[ { "id": "eai-legacy-0001", "t": "3D-VLA: A 3D Vision-Language-Action Generative World Model", "y": 2024, "u": "https://arxiv.org/abs/2403.09631", "m": [ "Learned dynamics", "VLA policy", "Language planner" ], "k": "zhen20243dvla3dvisionlanguageactiongenerative", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0001", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "3D perception and diffusion-generated goal observations support embodied planning; the abstract states no hazard criterion or safety enforcement.", "full_text_question": "Does the generated state preserve safety-relevant variables?", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "26ac266835e64c9d028f4eeb4fe8450b735f753f3d7f3354f7a7cbcaeca06128", "source_url": "https://arxiv.org/abs/2403.09631", "source_legacy_rows": [ 1 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0002", "t": "A call for embodied AI", "y": 2024, "u": "https://arxiv.org/abs/2402.03824", "m": [ "Language planner" ], "k": "paolo2024embodiedai", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0002", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "An active-inference account of embodied cognition motivates perception-action integration without an evaluated safeguard.", "full_text_question": "Use for conceptual background only.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "0be97b0aaebb82a6ec99f481d8470cb5af3944d1607bedd56b469c450d3c4c5b", "source_url": "https://arxiv.org/abs/2402.03824", "source_legacy_rows": [ 2 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0003", "t": "A Comprehensive Survey on Embodied Intelligence: Advancements, Challenges, and Future Perspectives", "y": 2024, "u": "https://www.sciopen.com/article/10.26599/AIR.2024.9150042", "m": [], "k": "sun2024survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0003", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "A broad perception-cognition-behavior survey discusses embodiment and ethical challenges rather than a predictive safety mechanism.", "full_text_question": "Check overlap with narrower safety reviews before adding citations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "2b3b2bcfb1cafca5840c4d38b6bf01aaf7d291817329db0d5ca1706dbc4400b7", "source_url": "https://www.sciopen.com/article/10.26599/AIR.2024.9150042", "source_legacy_rows": [ 3 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0004", "t": "A Framework for Benchmarking and Aligning Task-Planning Safety in LLM-Based Embodied Agents", "y": 2025, "u": "https://arxiv.org/abs/2504.14650v1", "m": [ "Language planner" ], "k": "huang2025framework", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0004", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Safe-BeAl combines a task-planning hazard benchmark with safety alignment and reports improved safety while preserving task completion.", "full_text_question": "Check hazard predicates, evaluation denominators and whether alignment prevents violations during execution.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "da7c99f45b132b6274a3cd08ad3469afe1a6abc4184b4be193a351be5232c490", "source_url": "https://arxiv.org/abs/2504.14650v1", "source_legacy_rows": [ 4 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0005", "t": "A Framework for Inherently Safer AGI through Language-Mediated Active Inference", "y": 2025, "u": "https://arxiv.org/abs/2508.05766", "m": [ "Language planner" ], "k": "wen2025frameworkinherentlysaferagi", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0005", "study_type": "position", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Language-mediated active inference proposes structural AGI safety, but validation is a future ARC research agenda rather than embodied control evidence.", "full_text_question": "Do not treat proposed guarantees as demonstrated robot safety.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "97a51798b4e5dccd55c9eb3d9fab3bc05f54863b32857f06628e040e8576a685", "source_url": "https://arxiv.org/abs/2508.05766", "source_legacy_rows": [ 5 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0006", "t": "A hierarchical architecture for embodied AI: planner-skills separation, distributed safety contracts, and real-time enforcement", "y": 2025, "u": "https://www.techrxiv.org/doi/full/10.36227/techrxiv.176344191.17191825", "m": [ "VLA policy", "Language planner" ], "k": "dandjinou2025hierarchical", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0006", "study_type": "architecture", "relevance": "direct", "full_text_priority": "high", "assessment": "Planner-skill separation, local safety contracts and a runtime supervisor explicitly connect plans to constrained actuation.", "full_text_question": "Verify the claimed non-violability of envelopes, actual experiments and response-time assumptions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "91f209e6ae0411c299caaccef4948938e3c1f3355673037285883cef7c1581d1", "source_url": "https://www.techrxiv.org/doi/full/10.36227/techrxiv.176344191.17191825", "source_legacy_rows": [ 6 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0007", "t": "A Human-Centered Safe Robot Reinforcement Learning Framework with Interactive Behaviors", "y": 2023, "u": "https://arxiv.org/abs/2302.13137", "m": [], "k": "gu2023humancenteredsaferobotreinforcement", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0007", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "A human-centered safe-RL framework proposes interactive exploration, value alignment and collaboration; the abstract presents an agenda.", "full_text_question": "Distinguish proposed interaction mechanisms from evaluated safety results.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "5c8fa809cda8690e4a02f88c06108c1a2695948eabd893a6edc70cc45ae163c4", "source_url": "https://arxiv.org/abs/2302.13137", "source_legacy_rows": [ 7 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0008", "t": "A Mapping of Assurance Techniques for Learning Enabled Autonomous Systems to the Systems Engineering Lifecycle", "y": 2022, "u": "https://ieeexplore.ieee.org/document/9763657", "m": [], "k": "ellis2022assurance", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0008", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Maps assurance and verification techniques to development, acquisition and sustainment within a systems-engineering lifecycle.", "full_text_question": "Extract assurance obligations relevant to deployed learned controllers.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ec1c52dd8a0af03653eccaad5cdfc978ca8fd236d9cf514a2a4caed330981c04", "source_url": "https://ieeexplore.ieee.org/document/9763657", "source_legacy_rows": [ 8 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0009", "t": "A Multiplicative Value Function for Safe and Efficient Reinforcement Learning", "y": 2023, "u": "https://arxiv.org/abs/2303.04118", "m": [], "k": "bührer2023multiplicativevaluefunctionsafe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0009", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A safety critic predicts constraint-violation probability and multiplicatively discounts a reward critic, with navigation and sim-to-real evaluation.", "full_text_question": "Check probability calibration, violation metrics and conditions on the claimed safety benefit.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "97a650388dc249e0fdf41e2282ec7e943648b78dbd732d0755bfb485a66be451", "source_url": "https://arxiv.org/abs/2303.04118", "source_legacy_rows": [ 9 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0010", "t": "A review of embodied intelligence systems: a three-layer framework integrating multimodal perception, world modeling, and structured strategies", "y": 2025, "u": "https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2025.1668910/full?utm_source=chatgpt.com", "m": [ "Learned dynamics" ], "k": "zhang2025review", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0010", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "A perception-modeling-decision framework integrates multimodal alignment and world-state prediction without an explicit safety property in the abstract.", "full_text_question": "Separate proposed architecture from evidence on hazard-aware prediction.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1ce6e41adae04c1be032133c96d3863b85fa8ad19af80735ad9409da4b8b3c4f", "source_url": "https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2025.1668910/full?utm_source=chatgpt.com", "source_legacy_rows": [ 10 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0011", "t": "A Review of Safe Reinforcement Learning: Methods, Theory and Applications", "y": 2024, "u": "https://arxiv.org/abs/2205.10330", "m": [], "k": "gu2024reviewsaferl", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0011", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "Reviews safe-RL methods, theory, sample complexity, applications and benchmark implementations.", "full_text_question": "Map its safety formulations and theoretical assumptions to the review taxonomy.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "d8cc3d105e2f1d98c570d6dca94948edb4c5698cefac7101a6829a8d830961ff", "source_url": "https://arxiv.org/abs/2205.10330", "source_legacy_rows": [ 11 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0012", "t": "A survey of constraint formulations in safe reinforcement learning", "y": 2024, "u": "https://arxiv.org/abs/2402.02025", "m": [], "k": "wachi2024survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0012", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "Organizes safe-RL constraint formulations and their mathematical relationships, directly informing distinctions among safety objectives.", "full_text_question": "Verify equivalences and non-equivalences among statewise, chance and cumulative-cost constraints.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "b092227b871f6295ec3e0822d8a58a787174942ca44e48f3fc32b84cdab68202", "source_url": "https://arxiv.org/abs/2402.02025", "source_legacy_rows": [ 12 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0013", "t": "A Survey of Embodied Learning for Object-Centric Robotic Manipulation", "y": 2025, "u": "https://arxiv.org/abs/2408.11537", "m": [], "k": "zheng2025survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0013", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Surveys object-centric perception, policy learning and manipulation tasks; no explicit safety evaluation is stated.", "full_text_question": "Use as manipulation representation background.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "b3cd4e88924303a9c16d0e044af591b20e45c6f373a24807fe4ff3006563fb54", "source_url": "https://arxiv.org/abs/2408.11537", "source_legacy_rows": [ 13 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0014", "t": "A Survey of Safe Reinforcement Learning and Constrained MDPs: A Technical Survey on Single-Agent and Multi-Agent Safety", "y": 2025, "u": "https://arxiv.org/abs/2505.17342", "m": [], "k": "kushwaha2025survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0014", "study_type": "review", "relevance": "direct", "full_text_priority": "medium", "assessment": "Surveys CMDP foundations and single- and multi-agent safe RL, including constrained optimization and exploration.", "full_text_question": "Check the precise guarantee types and multi-agent assumptions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1156e4f0170919f9e6e3df64b6ac0dbf2f4cb75d17945bd7082530a8cfb0a2bb", "source_url": "https://arxiv.org/abs/2505.17342", "source_legacy_rows": [ 14 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0015", "t": "A Survey of Safe Reinforcement Learning Methods in Robotics", "y": 2025, "u": "https://www.itm-conferences.org/articles/itmconf/pdf/2025/09/itmconf_cseit2025_01014.pdf", "m": [], "k": "xie2025survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0015", "study_type": "review", "relevance": "direct", "full_text_priority": "medium", "assessment": "Compares control, formal and constrained-optimization approaches using existing safe-RL benchmark results.", "full_text_question": "Check whether the claimed superiority of constrained optimization is limited to particular tasks and metrics.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "6ae7c2d80d9667c8ab87cc4353d3fee8d12cd57a0f75ad6dcab36d6a3fcb2779", "source_url": "https://www.itm-conferences.org/articles/itmconf/pdf/2025/09/itmconf_cseit2025_01014.pdf", "source_legacy_rows": [ 15 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0016", "t": "A Survey of World Models for Autonomous Driving", "y": 2025, "u": "https://arxiv.org/abs/2501.11260", "m": [ "Learned dynamics" ], "k": "feng2025surveyworldmodelsautonomous", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0016", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys driving world models for future prediction, planning and rare-event simulation, without specifying an assurance mechanism.", "full_text_question": "Identify explicit collision-risk evaluations rather than general robustness claims.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "721d792f96dd34d1c4e00af279384e59f3521ccf73f1ad9eb2159a393c95db09", "source_url": "https://arxiv.org/abs/2501.11260", "source_legacy_rows": [ 16 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0017", "t": "A Survey on Robotics with Foundation Models: toward Embodied AI", "y": 2024, "u": "https://arxiv.org/abs/2402.02385", "m": [], "k": "xu2024surveyroboticsfoundationmodels", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0017", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Surveys foundation-model manipulation, planning, control and benchmarks; the abstract emphasizes capability and generalization.", "full_text_question": "Use for architectural context rather than safety effectiveness.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e30256b49ef1ff5eb459ff54c88154b633ea0b01704bd8b12e3d6e64cc8e5ec1", "source_url": "https://arxiv.org/abs/2402.02385", "source_legacy_rows": [ 17 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0018", "t": "A Survey on Vision-Language-Action Models for Autonomous Driving", "y": 2025, "u": "https://arxiv.org/abs/2506.24044", "m": [ "VLA policy", "Language planner" ], "k": "jiang2025survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0018", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys driving VLAs and benchmarks that jointly measure safety, accuracy and explanation quality.", "full_text_question": "Extract safety metrics and distinguish explanations from formal verification.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "92faabf40ca267eb0929da2bb186485baad19018f0de6c5fb34d8cb45cb89e6f", "source_url": "https://arxiv.org/abs/2506.24044", "source_legacy_rows": [ 18, 19 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0020", "t": "A Survey on Vision-Language-Action Models: An Action Tokenization Perspective", "y": 2025, "u": "https://arxiv.org/abs/2507.01925", "m": [ "VLA policy" ], "k": "long2025survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0020", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Classifies VLA action tokens from language and affordances to trajectories and raw controls, informing representation placement.", "full_text_question": "Which action representations expose variables needed by a runtime safety check?", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "bf7e4d8d113d185a0a11fa7311ea2f570632435c4ece70b6dff49ff73f01a4a0", "source_url": "https://arxiv.org/abs/2507.01925", "source_legacy_rows": [ 20 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0021", "t": "A Survey: Learning Embodied Intelligence from Physical Simulators and World Models", "y": 2025, "u": "https://arxiv.org/abs/2507.00917v3", "m": [ "Learned dynamics" ], "k": "long2025survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0021", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys complementary physical simulators and learned world models for embodied learning and sim-to-real transfer.", "full_text_question": "Distinguish safe training environments from guarantees on deployed behavior.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "18f45c66cca25927b4521fbd47f6a5f45fc217af1f38ed18045d6cf7cab0acfb", "source_url": "https://arxiv.org/abs/2507.00917v3", "source_legacy_rows": [ 21 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0022", "t": "A Systematic Literature Review on AI Safety: Identifying Trends, Challenges and Future Directions", "y": 2024, "u": "https://ieeexplore.ieee.org/document/10630784", "m": [], "k": "10630784", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0022", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Broad trustworthy-AI synthesis covers verification, robustness and failure modes without an embodied-controller focus.", "full_text_question": "Use only for general assurance terminology.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "83ad40e74938f29b4a93883a664a4b45b2b1fee628ecbd58ca5fbcab3ca0828d", "source_url": "https://ieeexplore.ieee.org/document/10630784", "source_legacy_rows": [ 22 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0023", "t": "A3VLM: Actionable Articulation-Aware Vision Language Model", "y": 2024, "u": "https://arxiv.org/abs/2406.07549", "m": [ "Language planner" ], "k": "huang2024a3vlm", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0023", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "A3VLM predicts object articulation and affordances transferable to robot primitives, with simulated and physical evaluation.", "full_text_question": "Does the representation encode hazardous interactions beyond task affordance?", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f9f4321d329db4a3d7f6225af12a64f45435e7ba2e9ba055c54bb9af2f3c77a2", "source_url": "https://arxiv.org/abs/2406.07549", "source_legacy_rows": [ 23 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0024", "t": "ActSafe: Active Exploration with Safety Constraints for Reinforcement Learning", "y": 2024, "u": "https://arxiv.org/abs/2410.09486", "m": [ "Learned dynamics" ], "k": "rev_actsafe", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0025", "t": "AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions", "y": 2025, "u": "https://arxiv.org/abs/2506.14697v3", "m": [ "Language planner" ], "k": "ying2025agentsafe", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0025", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "SAFE evaluates hazardous instructions across perception, planning and execution and reports failures in converting hazard recognition to safe action.", "full_text_question": "Verify benchmark naming, hazard labels and stage-specific failure attribution.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "856c9ccc7732a4887d4a3cc5f157f68a206a5741f4925ea0c7243a43ac06b035", "source_url": "https://arxiv.org/abs/2506.14697v3", "source_legacy_rows": [ 25 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0026", "t": "AI Risk Management Should Incorporate Both Safety and Security", "y": 2024, "u": "https://arxiv.org/abs/2405.19524", "m": [ "Language planner" ], "k": "qi2024airiskmanagement", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0026", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "Unifies AI safety and security concepts for risk management without evaluating embodied control.", "full_text_question": "Use for terminology, not physical-safety evidence.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7d61d5d7b7ef2deb0eef0d6c365a0e01801cdf0fac821683d67c82b68d4b42a0", "source_url": "https://arxiv.org/abs/2405.19524", "source_legacy_rows": [ 26 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0027", "t": "AI Safety for Everyone", "y": 2025, "u": "https://www.nature.com/articles/s42256-025-01020-y", "m": [], "k": "gyevnar2025ai", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0027", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Argues for a broad conception of AI safety using a literature review of present-day technical risks.", "full_text_question": "Use for scope framing only.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e35cafe879e816b96a6c9d89f92a14b9fc08e29d9de248be176f73ac8e7d017e", "source_url": "https://www.nature.com/articles/s42256-025-01020-y", "source_legacy_rows": [ 27 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0028", "t": "Align, Reason and Learn: Enhancing Medical Vision-and-Language Pre-training with Knowledge", "y": 2022, "u": "https://arxiv.org/abs/2209.07118v1", "m": [], "k": "chen2022align", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0028", "study_type": "method", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Knowledge-enhanced medical image-text pretraining is evaluated on downstream representation tasks rather than an embodied controller.", "full_text_question": "Exclude from robotic safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "25de5793663b0fe39d8dbc166581852e36d819a07d078ea53b497b5649579180", "source_url": "https://arxiv.org/abs/2209.07118v1", "source_legacy_rows": [ 28 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0029", "t": "Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI", "y": 2025, "u": "https://arxiv.org/abs/2407.06886v8", "m": [ "Learned dynamics" ], "k": "liu2025aligningcyberspacephysical", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0029", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Broad embodied-AI survey covers robots, simulators, perception, interaction and sim-to-real adaptation.", "full_text_question": "Check overlap with existing background reviews.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f73b76988ea51e2bbeb888d5e896f5b8c1b866c0a77a19b7aedbbc4f2d4b722e", "source_url": "https://arxiv.org/abs/2407.06886v8", "source_legacy_rows": [ 29 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0030", "t": "An Embodied Generalist Agent in 3D World", "y": 2024, "u": "https://arxiv.org/abs/2311.12871v3", "m": [ "VLA policy", "Language planner" ], "k": "huang2024embodiedgeneralistagent3d", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0030", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "LEO combines 3D grounding, reasoning, navigation and manipulation through instruction tuning without a stated safety criterion.", "full_text_question": "Separate task competence from hazard-sensitive behavior.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f5ec44dfc1f951f2a2980de6b8015b0c686e8fd747f3fc7c7ef20b5b474a901c", "source_url": "https://arxiv.org/abs/2311.12871v3", "source_legacy_rows": [ 30 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0031", "t": "An Online Robot Collision Detection and Identification Scheme by Supervised Learning and Bayesian Decision Theory", "y": 2021, "u": "https://ieeexplore.ieee.org/document/9109713", "m": [], "k": "zhang2021bayesian", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0031", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Supervised collision diagnosis with Bayesian confidence enables rapid reactions on a physical manipulator after interaction begins.", "full_text_question": "Check detection latency, false alarms and residual harm; detection is reactive rather than pre-contact avoidance.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "9902c59cd00e4a30fc65ea112942beb7a8ef8bc2d8f261f2bd47dafe0bd7dec2", "source_url": "https://ieeexplore.ieee.org/document/9109713", "source_legacy_rows": [ 31 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0032", "t": "Artificial Intelligence: Approaches to Safety", "y": 2025, "u": "https://compass.onlinelibrary.wiley.com/doi/10.1111/phc3.70039", "m": [], "k": "dalessandro2025aisafety", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0032", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Philosophical AI-safety review discusses oversight, interpretability and corrigibility without embodied-control evaluation.", "full_text_question": "Use only where philosophical safety concepts are needed.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "6d988b3c3a533b976ac1edfec342fa4d30d884dd8a9927260110f1510a3f0340", "source_url": "https://compass.onlinelibrary.wiley.com/doi/10.1111/phc3.70039", "source_legacy_rows": [ 32 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0033", "t": "Assistax: A Hardware-Accelerated Reinforcement Learning Benchmark for Assistive Robotics", "y": 2025, "u": "https://arxiv.org/abs/2507.21638v1", "m": [], "k": "hinckeldey2025assist", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0033", "study_type": "benchmark", "relevance": "context", "full_text_priority": "medium", "assessment": "Assistax accelerates assistive multi-agent simulation and tests coordination with diverse human partners.", "full_text_question": "Are physical harm or contact-force constraints measured separately from coordination reward?", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "92ff7ff3d89b927226a329047f249a1bce1db05fc4870e7ce0b56a6293784c36", "source_url": "https://arxiv.org/abs/2507.21638v1", "source_legacy_rows": [ 33 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0034", "t": "BadRobot: Jailbreaking Embodied LLMs in the Physical World", "y": 2025, "u": "https://arxiv.org/abs/2407.20242v4", "m": [ "Language planner" ], "k": "zhang2025badrobot", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0034", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "BadRobot attacks embodied LLMs through instruction manipulation and language-action mismatch, with malicious physical-action benchmarks.", "full_text_question": "Verify physical execution evidence and failure rates for each attack mechanism.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "70e5da63971c1fe33a27d687a7c3502c0389d81ab020c93185b3c891f5dc101f", "source_url": "https://arxiv.org/abs/2407.20242v4", "source_legacy_rows": [ 34 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0035", "t": "Bayesian optimization with safety constraints: safe and automatic parameter tuning in robotics", "y": 2023, "u": "https://link.springer.com/article/10.1007/s10994-021-06019-1", "m": [], "k": "berkenkamp2023bayesian", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0035", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Safe Bayesian optimization separates safety constraints from performance and reports probabilistic safe tuning on a quadrotor.", "full_text_question": "Check initial safe-set requirements, Gaussian-process regularity and confidence assumptions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e2ed45ef5849c1e04769d783f3a0b4a91511e7bc55c209aaa7d22e3d6b75f742", "source_url": "https://link.springer.com/article/10.1007/s10994-021-06019-1", "source_legacy_rows": [ 35 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0036", "t": "Benchmark Evaluations, Applications, and Challenges of Large Vision Language Models: A Survey", "y": 2025, "u": "https://arxiv.org/abs/2501.02189", "m": [ "Language planner" ], "k": "li2025surveystateartlarge", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0036", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "A general VLM survey includes robotics and safety among multiple applications; it is not exclusively non-embodied.", "full_text_question": "Use for broad VLM background rather than safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ff79fcd2ed1cf9692966176367e7fac6ba1cf2fa3f1614d97e14a168624fe242", "source_url": "https://arxiv.org/abs/2501.02189", "source_legacy_rows": [ 36 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0037", "t": "Benchmarking Reinforcement Learning Algorithms on Real-World Robots", "y": 2018, "u": "https://arxiv.org/abs/1809.07731v1", "m": [], "k": "mahmood2018benchmarking", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0037", "study_type": "benchmark", "relevance": "context", "full_text_priority": "medium", "assessment": "Physical-robot RL benchmarks expose hyperparameter sensitivity and cross-task differences without explicit safety metrics in the abstract.", "full_text_question": "Identify hardware safeguards and distinguish learning performance from safe exploration.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "086148f47ae066a2fbaa226a67b983c0361350c773a56ff5a3d8969cc60d605a", "source_url": "https://arxiv.org/abs/1809.07731v1", "source_legacy_rows": [ 37 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0038", "t": "Benchmarking Vision, Language, & Action Models on Robotic Learning Tasks", "y": 2024, "u": "https://arxiv.org/abs/2411.05821", "m": [ "VLA policy" ], "k": "guruprasad2024benchmarking", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0038", "study_type": "benchmark", "relevance": "context", "full_text_priority": "medium", "assessment": "Evaluates VLA manipulation across datasets and embodiments, emphasizing task performance and action-space sensitivity.", "full_text_question": "Check whether safety failures are separately annotated.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "9bf42743675338e245855b9ce5e0deccf4eded5840a06aa5fd6fb08d16bfeb3c", "source_url": "https://arxiv.org/abs/2411.05821", "source_legacy_rows": [ 38 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0039", "t": "Between world models and model worlds: on generality, agency, and worlding in machine learning", "y": 2025, "u": "https://link.springer.com/article/10.1007/s00146-024-02086-9", "m": [ "Learned dynamics", "Language planner" ], "k": "mitrokhov2025between", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0039", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "A cultural critique of model worlds questions claims of generality and grounding rather than proposing robotic assurance.", "full_text_question": "Use only for epistemic framing if needed.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "af5dd017f5c303f315144e78a33066c87dfee9534556b4f111a1bca90fba7877", "source_url": "https://link.springer.com/article/10.1007/s00146-024-02086-9", "source_legacy_rows": [ 39 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0040", "t": "Bridging the Embodiment Gap: Embodied AI for Enhanced Human-Machine Collaboration and Learning in Dynamic Environments", "y": 2024, "u": "https://www.preprints.org/frontend/manuscript/79a28712bada86655f522af0f1ced4b8/download_pub", "m": [], "k": "idowu2024bridging", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0040", "study_type": "position", "relevance": "unclear", "full_text_priority": "medium", "assessment": "Claims embodied collaboration benefits and experimental case studies but does not identify a concrete method, task or safety measure.", "full_text_question": "Verify publication source, actual experiments and operational safety definitions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "07368c759748b44d28b7fd6835b1a7c76c0bb87934cc6947552be0d29a44e8e8", "source_url": "https://www.preprints.org/frontend/manuscript/79a28712bada86655f522af0f1ced4b8/download_pub", "source_legacy_rows": [ 40 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0041", "t": "ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model", "y": 2025, "u": "https://aclanthology.org/2025.emnlp-main.273/", "m": [ "VLA policy", "Language planner" ], "k": "zhou2025chatvla", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0041", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "ChatVLA addresses forgetting and task interference with phased training and expert modules; results concern understanding and manipulation success.", "full_text_question": "No safety intervention is established by the abstract.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "6b775c14a1689c0689cc264eb868cdf1a0474165499e293682b5f309dab05f43", "source_url": "https://aclanthology.org/2025.emnlp-main.273/", "source_legacy_rows": [ 41 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0042", "t": "CoA-VLA: Improving Vision-Language-Action Models via Visual-Text Chain-of-Affordance", "y": 2025, "u": "https://openaccess.thecvf.com/content/ICCV2025/html/Li_CoA-VLA_Improving_Vision-Language-Action_Models_via_Visual-Text_Chain-of-Affordance_ICCV_2025_paper.html", "m": [ "VLA policy" ], "k": "li2025affordance", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0042", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "CoA-VLA reasons over object, grasp, spatial and collision-free movement affordances and reports obstacle-avoidance generalization.", "full_text_question": "Check collision metrics and whether predicted affordances constrain execution.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "a8905b603c5e997d8794a07355119e84ba314aabc5d9ef83932c6109e14f89db", "source_url": "https://openaccess.thecvf.com/content/ICCV2025/html/Li_CoA-VLA_Improving_Vision-Language-Action_Models_via_Visual-Text_Chain-of-Affordance_ICCV_2025_paper.html", "source_legacy_rows": [ 42 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0043", "t": "CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation", "y": 2024, "u": "https://arxiv.org/abs/2411.19650", "m": [ "VLA policy" ], "k": "li2024cogact", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0043", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "CogACT combines VLM cognition with diffusion action modeling and reports cross-embodiment task-success gains.", "full_text_question": "Distinguish generalization from explicit safety evaluation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3831dcd76fa76baf0238b81850beb9e5d75244f5e70241370febe680d8352357", "source_url": "https://arxiv.org/abs/2411.19650", "source_legacy_rows": [ 43 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0044", "t": "Communicating Plans, Not Percepts: Scalable Multi-Agent Coordination with Embodied World Models", "y": 2025, "u": "https://arxiv.org/abs/2508.02912v4", "m": [ "Learned dynamics" ], "k": "hill2025plan", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0044", "study_type": "method", "relevance": "context", "full_text_priority": "medium", "assessment": "World-model-generated plans support multi-agent communication and coordination in grid-world task allocation.", "full_text_question": "Check whether collision or other safety constraints are modeled beyond coordination reward.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1b3a6ea4f639f2cdcb290b4ae957e49195fb199a5cfa7dabdc54980ea4a870b6", "source_url": "https://arxiv.org/abs/2508.02912v4", "source_legacy_rows": [ 44 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0045", "t": "Constrained Variational Policy Optimization for Safe Reinforcement Learning", "y": 2022, "u": "https://arxiv.org/abs/2201.11927", "m": [], "k": "liu2022constrained", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0045", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "CVPO uses constrained variational inference and trust-region policy updates for improved constraint satisfaction in robotic tasks.", "full_text_question": "Separate optimization guarantees from guarantees on the executed policy and training violations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3a5e99a5c1946420f07e714e899cf4df43cef03957a5acb6c2184e6ac721bc03", "source_url": "https://arxiv.org/abs/2201.11927", "source_legacy_rows": [ 45 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0046", "t": "Continuous-Discrete Reinforcement Learning for Hybrid Control in Robotics", "y": 2020, "u": "https://arxiv.org/abs/2001.00449v1", "m": [], "k": "neunert2020continuous", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0046", "study_type": "method", "relevance": "context", "full_text_priority": "medium", "assessment": "Hybrid RL jointly optimizes discrete and continuous robot actions and proposes meta-actions to reduce wear.", "full_text_question": "Determine whether wear reduction is a measured safety objective or a motivation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "81498e68bbb438d8074ece8bad55fbd7e19107721d2f7241d91927deb955fe5a", "source_url": "https://arxiv.org/abs/2001.00449v1", "source_legacy_rows": [ 46 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0047", "t": "cVLA: Towards Efficient Camera-Space VLAs", "y": 2025, "u": "https://arxiv.org/abs/2507.02190", "m": [ "VLA policy", "Language planner" ], "k": "argus2025cvla", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0047", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "Camera-space VLA predicts end-effector waypoints and demonstrates sim-to-real transfer without an explicit hazard criterion.", "full_text_question": "Check trajectory feasibility and collision validation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "42a82656f5f2dcf688fd17bd035ab551cb0e5138e39a94f765b8280d008963ec", "source_url": "https://arxiv.org/abs/2507.02190", "source_legacy_rows": [ 47 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0048", "t": "Deception Game: Closing the Safety-Learning Loop in Interactive Robot Autonomy", "y": 2023, "u": "https://arxiv.org/abs/2309.01267v2", "m": [], "k": "hu2023deceptiongameclosingsafetylearning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0048", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Deception Game jointly models physical dynamics and online learning to synthesize adaptive safe control through adversarial RL.", "full_text_question": "Inspect the uncertainty evolution, adversary assumptions and safety conditions under adaptation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "31159c4ceaef174c010a156aee17ff0dc066e16dfd459ae812925cf7eb68b46e", "source_url": "https://arxiv.org/abs/2309.01267v2", "source_legacy_rows": [ 48 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0049", "t": "Deep Reinforcement Learning for Decentralized Multi-Robot Control: A DQN Approach to Robustness and Information Integration", "y": 2024, "u": "https://arxiv.org/abs/2408.11339v1", "m": [], "k": "wu2024decentralized", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0049", "study_type": "method", "relevance": "context", "full_text_priority": "low", "assessment": "Decentralized DQN coordination reports simulated efficiency and fault-tolerance benefits but no defined safety constraint.", "full_text_question": "Do not equate robustness with physical-safety assurance.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "72fa1e423def61edc6397ad2bc7e8745e8975378e1fa4e7d417b1fb5ec27cd3d", "source_url": "https://arxiv.org/abs/2408.11339v1", "source_legacy_rows": [ 49 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0050", "t": "Designing a Multi-Objective Reward Function for Creating Teams of Robotic Bodyguards Using Deep Reinforcement Learning", "y": 2019, "u": "https://arxiv.org/abs/1901.09837v1", "m": [], "k": "sheikh2019designing", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0050", "study_type": "method", "relevance": "direct", "full_text_priority": "medium", "assessment": "Reward design for teams of robotic bodyguards targets physical protection through competing objectives.", "full_text_question": "Check harm definitions, simulation setting and whether protection is enforced or merely rewarded.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "592f97eaf773f47c3e8e57c72de642f30deaa8d1d63931ac7322672c283db0d1", "source_url": "https://arxiv.org/abs/1901.09837v1", "source_legacy_rows": [ 50 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0051", "t": "Designing Control Barrier Function Via Probabilistic Enumeration for Safe Reinforcement Learning Navigation", "y": 2025, "u": "https://ieeexplore.ieee.org/document/11117187", "m": [], "k": "marzari2025cbf", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0051", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Probabilistic enumeration identifies unsafe regions for a CBF correction layer, evaluated in simulation and robot navigation.", "full_text_question": "Verify coverage of unsafe regions, barrier validity and real-time correction feasibility.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "549c435b7ab29b9050d993666e463c860dd3469a32a757b426a5f34396c1054b", "source_url": "https://ieeexplore.ieee.org/document/11117187", "source_legacy_rows": [ 51 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0052", "t": "Digital twins to embodied artificial intelligence: review and perspective", "y": 2025, "u": "https://www.oaepublish.com/articles/ir.2025.11", "m": [], "k": "li2025digital", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0052", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Reviews digital twins as training, monitoring and adaptation infrastructure for embodied AI.", "full_text_question": "Distinguish simulated predeployment testing from runtime safety assurance.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1cb40ef7920b70e72d63e43e710b9926da011088c06f566e417da671f1cc6842", "source_url": "https://www.oaepublish.com/articles/ir.2025.11", "source_legacy_rows": [ 52 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0053", "t": "Do What? Teaching Vision-Language-Action Models to Reject the Impossible", "y": 2025, "u": "https://arxiv.org/abs/2508.16292", "m": [ "VLA policy" ], "k": "hsieh2025whatteachingvisionlanguageactionmodels", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0053", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "IVA detects false-premise instructions, requests correction and grounds alternatives before acting.", "full_text_question": "Check whether infeasibility detection reduces physical hazards rather than only instruction errors.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e95957eee399addb1e601b0c89e61f0df9adebec6e28f377e5feb8bcb8d959c1", "source_url": "https://arxiv.org/abs/2508.16292", "source_legacy_rows": [ 53 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0054", "t": "Don't do it: Safer Reinforcement Learning With Rule-based Guidance", "y": 2022, "u": "https://arxiv.org/abs/2212.13819", "m": [], "k": "nikonova2022dontitsaferreinforcement", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0054", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Relational safety rules override unsafe exploration actions through a modified epsilon-greedy policy.", "full_text_question": "Verify rule completeness, observation assumptions and training-versus-deployment guarantees.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "02fa6eced6a0680184a5fb86d445beceee7078889e414ef5e4fd351507f0b50f", "source_url": "https://arxiv.org/abs/2212.13819", "source_legacy_rows": [ 54 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0055", "t": "Don't Let Your Robot be Harmful: Responsible Robotic Manipulation", "y": 2025, "u": "https://arxiv.org/abs/2411.18289", "m": [ "Learned dynamics" ], "k": "ni2025dontletrobotharmful", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0055", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Safety-as-policy combines generated risk scenarios with consequence reflection and reports synthetic and physical manipulation evaluation.", "full_text_question": "Inspect world-model fidelity, SafeBox hazard labels and execution-time intervention.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e67dfe62783a5a9852229172a8da465269f1fa74c61e43434f4e96016ee50102", "source_url": "https://arxiv.org/abs/2411.18289", "source_legacy_rows": [ 55 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0056", "t": "EAIRiskBench: Towards Evaluating Physical Risk Awareness for Task Planning of Foundation Model-based Embodied AI Agents", "y": 2024, "u": "https://arxiv.org/abs/2408.04449", "m": [ "Language planner" ], "k": "zhu2024earbenchevaluatingphysicalrisk", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0056", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "EAIRiskBench generates and judges physical-risk planning scenarios and evaluates prompting mitigations.", "full_text_question": "Audit model-generated scenarios, model-based judges and the denominator behind the reported task-risk rate.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "9b1de733364f6953611c091ca467c47c44d378d7f90251bc6a77709e44835ed2", "source_url": "https://arxiv.org/abs/2408.04449", "source_legacy_rows": [ 56 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0057", "t": "Edge General Intelligence Through World Models and Agentic AI: Fundamentals, Solutions, and Challenges", "y": 2025, "u": "https://arxiv.org/abs/2508.09561", "m": [ "Analytical dynamics", "Learned dynamics" ], "k": "zhao2025edgegeneralintelligenceworld", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0057", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys edge world models for vehicular, UAV and network applications, treating safety guarantees as an open challenge.", "full_text_question": "Separate embodied-control evidence from network optimization examples.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f6b9693afdf06062d896b6e41f0e15b4a171710eeea009482c013c422302602b", "source_url": "https://arxiv.org/abs/2508.09561", "source_legacy_rows": [ 57 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0058", "t": "Embodied AI Through Cloud-Fog Computing: A Framework for Everywhere Intelligence", "y": 2024, "u": "https://ieeexplore.ieee.org/document/10595837", "m": [], "k": "10595837", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0058", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "Cloud-fog infrastructure is proposed to meet embodied-AI computing demands without an evaluated safety mechanism.", "full_text_question": "Check latency implications only if using this infrastructure discussion.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "4eeef89fdf735a130c53943f90086efd17e0db6779bc614ed1ee1768376d4928", "source_url": "https://ieeexplore.ieee.org/document/10595837", "source_legacy_rows": [ 58 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0059", "t": "Embodied AI with Two Arms: Zero-shot Learning, Safety and Modularity", "y": 2024, "u": "https://arxiv.org/abs/2404.03570", "m": [ "Language planner" ], "k": "varley2024embodiedaiarmszeroshot", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0059", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A modular dual-arm system combines language planning with trajectory optimization and compliant control for human proximity.", "full_text_question": "Verify semantic checks, collision constraints, compliance limits and physical test coverage.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "dad14bd3177c1441b2f4d2170b95f417cf1734d4896576d9832d26cf50617ff7", "source_url": "https://arxiv.org/abs/2404.03570", "source_legacy_rows": [ 59 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0060", "t": "Embodied ai: Emerging risks and opportunities for policy action", "y": 2025, "u": "https://arxiv.org/abs/2509.00117", "m": [], "k": "perlo2025embodied", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0060", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "Embodied-AI policy analysis categorizes physical and societal risks and recommends governance measures.", "full_text_question": "Check possible preprint/proceedings overlap with eai-legacy-0073 before treating them as distinct studies.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f55519d7c504bd92b71f038703d966a930bb17e9c0c09e396b0ad4fbc6509ce9", "source_url": "https://arxiv.org/abs/2509.00117", "source_legacy_rows": [ 60 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0061", "t": "Embodied AI: From LLMs to World Models", "y": 2025, "u": "https://arxiv.org/abs/2509.20021v1", "m": [ "Learned dynamics", "Language planner" ], "k": "feng2025embodiedaillmsworld", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0061", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys LLM- and world-model-driven embodiment and motivates their joint architecture.", "full_text_question": "Identify explicit assurance evidence rather than assumed physical-law compliance.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3551818b47acad76bf0b3c26f22a5923eab90643497d22ad4693eeaaef2ec855", "source_url": "https://arxiv.org/abs/2509.20021v1", "source_legacy_rows": [ 61 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0062", "t": "Embodied Arena: A Comprehensive, Unified, and Evolving Evaluation Platform for Embodied AI", "y": 2025, "u": "https://arxiv.org/abs/2509.15273v2", "m": [ "Language planner" ], "k": "ni2025embodiedarenacomprehensiveunified", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0062", "study_type": "benchmark", "relevance": "context", "full_text_priority": "medium", "assessment": "Embodied Arena unifies perception, reasoning and task-execution benchmarks across models.", "full_text_question": "Check whether hazard-specific dimensions are present in its capability taxonomy.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "749d46378b8692724cc36c655561795106d73ad0bdefecb505dd0bb02408e0fc", "source_url": "https://arxiv.org/abs/2509.15273v2", "source_legacy_rows": [ 62 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0063", "t": "Embodied artificial intelligence security and governance", "y": 2024, "u": "https://bulletinofcas.researchcommons.org/journal/vol40/iss3/4/", "m": [], "k": "xu2024embodied", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0063", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "Proposes security and governance principles across information, physical and social domains of embodied AI.", "full_text_question": "Distinguish governance prescriptions from technically enforced safeguards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "4035d02d1dbb58e5345fc03362c77a297f6b6257ec2fef06a8a70af5b1fb8af0", "source_url": "https://bulletinofcas.researchcommons.org/journal/vol40/iss3/4/", "source_legacy_rows": [ 63 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0064", "t": "Embodied Evolution in Collective Robotics: A Review", "y": 2017, "u": "https://arxiv.org/abs/1709.08992v2", "m": [], "k": "", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0064", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Reviews online distributed evolutionary learning in robot collectives, without an explicit safety objective.", "full_text_question": "Use for historical multi-robot learning context only.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "da7a4ea345a56dc322c6680601b59232b076b818a7c197a8f621ea879638eca6", "source_url": "https://arxiv.org/abs/1709.08992v2", "source_legacy_rows": [ 64 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0065", "t": "Embodied Intelligence", "y": 2025, "u": "https://link.springer.com/content/pdf/10.1007/978-981-97-8440-0_8-1.pdf", "m": [ "Language planner" ], "k": "huang2025embodied", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0065", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "A chapter discusses shared control and override authority in human-AI collaboration across safety-relevant domains.", "full_text_question": "Trace any claimed safety benefit to concrete control or evaluation evidence.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "5d363b32b1e52fe6febe0862c501d8a2fd101d4ff687058d5702f399e6298a8f", "source_url": "https://link.springer.com/content/pdf/10.1007/978-981-97-8440-0_8-1.pdf", "source_legacy_rows": [ 65 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0066", "t": "Embodied Intelligence Through World Models", "y": 2024, "u": "https://utoronto.scholaris.ca/server/api/core/bitstreams/8a278445-a1ce-4a53-9275-4f1a0c6739b9/content", "m": [ "Learned dynamics" ], "k": "hafner2024embodied", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0066", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "A world-model thesis develops prediction-based planning and sample-efficient physical robot learning.", "full_text_question": "Check safety during online learning separately from efficiency and task reward.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "9bea2272186fab709a0d8fc68633dcc47e2ba0ece35ffea924fce480106281f2", "source_url": "https://utoronto.scholaris.ca/server/api/core/bitstreams/8a278445-a1ce-4a53-9275-4f1a0c6739b9/content", "source_legacy_rows": [ 66 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0067", "t": "Embodied Neuromorphic Artificial Intelligence for Robotics: Perspectives, Challenges, and Research Development Stack", "y": 2024, "u": "https://arxiv.org/abs/2404.03325v2", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0067", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "Neuromorphic robotics perspective identifies safety, reliability and security as development challenges.", "full_text_question": "No tested predictive safeguard is specified in the abstract.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "44f67c032e9f027ebd84f65b9f7a11aeb64faf13dd1eea906496b098f0637520", "source_url": "https://arxiv.org/abs/2404.03325v2", "source_legacy_rows": [ 67 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0068", "t": "Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation", "y": 2025, "u": "https://arxiv.org/abs/2508.13998v1", "m": [ "Language planner" ], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0068", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "Embodied-R1 uses pointing and reinforced fine-tuning for spatial reasoning and cross-embodiment manipulation.", "full_text_question": "Check whether disturbance robustness includes harm-specific evaluation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "abb40dbab0f8e55c2272c7b4a5e5f93a5e568c5aec8d8d74ccb195d2deae7b36", "source_url": "https://arxiv.org/abs/2508.13998v1", "source_legacy_rows": [ 68 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0069", "t": "Embodied-RAG: General Non-parametric Embodied Memory for Retrieval and Generation", "y": 2024, "u": "https://arxiv.org/abs/2409.18313v5", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0069", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "Embodied-RAG retrieves hierarchical spatial-semantic memory for navigation and explanation.", "full_text_question": "No explicit safety property or intervention is described.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "8733ed9e3237a0425e4f7218b7ba3d6c6dcf797c642ec594f5b761751eb4f803", "source_url": "https://arxiv.org/abs/2409.18313v5", "source_legacy_rows": [ 69 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0070", "t": "EmbodiedCity: A Benchmark Platform for Embodied Agent in Real-world City Environment", "y": 2024, "u": "https://arxiv.org/abs/2410.09604v1", "m": [ "Language planner" ], "k": "gao2024embodiedcitybenchmarkplatformembodied", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0070", "study_type": "benchmark", "relevance": "context", "full_text_priority": "medium", "assessment": "EmbodiedCity provides a city-scale simulator with pedestrians, vehicles and embodied evaluation tasks.", "full_text_question": "Check collision definitions, dynamics validity and whether reported results are simulation-only.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e4315c00f3e6517e0895392b01ce10c29ddf4abff7591fbac97b34d8b9d6a98b", "source_url": "https://arxiv.org/abs/2410.09604v1", "source_legacy_rows": [ 70 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0071", "t": "Embracing the Future: Navigating the Challenges and Solutions in Embodied Artificial Intelligence", "y": 2025, "u": "https://link.springer.com/chapter/10.1007/978-3-031-68256-8_13", "m": [], "k": "khan2025embracing", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0071", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "A broad chapter surveys embodied-AI applications, security and long-term safety challenges.", "full_text_question": "Use for background rather than demonstrated assurance.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "487ba4d5bc75948e91338761c5113d2b2ccfd0f51d1d59c70b15b45b133278b6", "source_url": "https://link.springer.com/chapter/10.1007/978-3-031-68256-8_13", "source_legacy_rows": [ 71 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0072", "t": "Emerging Practices in Frontier AI Safety Frameworks", "y": 2025, "u": "https://arxiv.org/pdf/2503.04746", "m": [], "k": "buhl2025emerging", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0072", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "Frontier-AI safety frameworks organize risk assessment, mitigation and governance without a robotic-controller focus.", "full_text_question": "Use for governance context only.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "bc3525f5c4c5a3bcc2a94262cd50109539ba19e22973c86588392676b3c37d9f", "source_url": "https://arxiv.org/pdf/2503.04746", "source_legacy_rows": [ 72 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0073", "t": "Emerging Risks from Embodied AI Require Urgent Policy Action", "y": 2025, "u": "https://openreview.net/pdf?id=fXiPp3qvrW", "m": [], "k": "perloemerging", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0073", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "An embodied-AI policy position paper identifies physical and societal risks and regulatory gaps.", "full_text_question": "Check possible version overlap with eai-legacy-0060; policy recommendations are not control guarantees.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "0b7eb0251f191afa1f4d339ac1cde751f68886957196eca2e467efca24dc9168", "source_url": "https://openreview.net/pdf?id=fXiPp3qvrW", "source_legacy_rows": [ 73 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0074", "t": "Emma-X: An Embodied Multimodal Action Model with Grounded Chain of Thought and Look-ahead Spatial Reasoning", "y": 2024, "u": "https://arxiv.org/abs/2412.11974", "m": [ "VLA policy", "Language planner" ], "k": "sun2024emmax", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0074", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "Emma-X uses grounded reasoning and look-ahead spatial guidance for manipulation, with trajectory segmentation to reduce hallucination.", "full_text_question": "Check whether spatial reasoning is assessed against hazards or only task success.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "df0d0cf4097e45307584581ebd2f04a285193be69c34e5fc8e26413ab3623cf5", "source_url": "https://arxiv.org/abs/2412.11974", "source_legacy_rows": [ 74 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0075", "t": "EMMOE: A Comprehensive Benchmark for Embodied Mobile Manipulation in Open Environments", "y": 2025, "u": "https://arxiv.org/abs/2503.08604", "m": [ "Language planner" ], "k": "li2025emmoe", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0075", "study_type": "benchmark", "relevance": "context", "full_text_priority": "medium", "assessment": "EMMOE evaluates long-horizon mobile manipulation with replanning and error detection; supplied text contains unresolved model/dataset macros.", "full_text_question": "Recover the canonical abstract and inspect whether failures are safety violations or task errors.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ad964645466244fa5393e7a615122cf3031f5ce7e9031f876c4bb072ef4d5521", "source_url": "https://arxiv.org/abs/2503.08604", "source_legacy_rows": [ 75 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0076", "t": "Emotion in Reinforcement Learning Agents and Robots: A Survey", "y": 2017, "u": "https://arxiv.org/abs/1705.05172v1", "m": [], "k": "moerland2017emotion", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0076", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Surveys computational emotions in RL and robots for learning efficiency and social signaling.", "full_text_question": "No predictive physical-safety mechanism is identified in the abstract.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "29476faf67378dd4bc055ded2e1af35c892e74fa8bba4d5ff68bd3f6a9b7d9c8", "source_url": "https://arxiv.org/abs/1705.05172v1", "source_legacy_rows": [ 76 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0077", "t": "EVA: An Embodied World Model for Future Video Anticipation", "y": 2025, "u": "https://arxiv.org/abs/2410.15461v2", "m": [ "Learned dynamics" ], "k": "chi2025evaembodiedworldmodel", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0077", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "EVA combines reasoning and video generation for embodied future anticipation and OOD evaluation.", "full_text_question": "Compare hazard prediction with video fidelity; no safety condition is stated.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "005667a109231bfd31a8c0063f860847acaa5eb07b6810ffa3d9ccd1479c30f5", "source_url": "https://arxiv.org/abs/2410.15461v2", "source_legacy_rows": [ 77 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0078", "t": "Evaluating uncertainty and quality of visual language action-enabled robots", "y": 2025, "u": "https://arxiv.org/abs/2507.17049", "m": [ "VLA policy" ], "k": "valle2025evaluating", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0078", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "Proposes uncertainty and execution-quality metrics correlated with expert ratings of VLA manipulation.", "full_text_question": "Check sampling bias from successful runs and whether uncertainty predicts harmful outcomes under shift.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "0bba69bc146a280839af119f425a7f7bf6ee0db9704049c65b830772c44c185b", "source_url": "https://arxiv.org/abs/2507.17049", "source_legacy_rows": [ 78 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0079", "t": "Evo-0: Vision-Language-Action Model with Implicit Spatial Understanding", "y": 2025, "u": "https://arxiv.org/abs/2507.00416", "m": [ "VLA policy", "Language planner" ], "k": "lin2025evo", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0079", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "Evo-0 injects implicit 3D features to improve spatial VLA tasks without a stated safety objective.", "full_text_question": "Do not equate spatial-task success with collision avoidance.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "58ad14c9524520728e278382311e7c9839c90664f29dc5c18626c331e0f36e77", "source_url": "https://arxiv.org/abs/2507.00416", "source_legacy_rows": [ 79 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0080", "t": "Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics", "y": 2025, "u": "https://arxiv.org/abs/2411.13587", "m": [ "VLA policy" ], "k": "wang2025adversarial", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0080", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Adversarial patches target VLA action positions and trajectories, exposing execution vulnerabilities.", "full_text_question": "Distinguish task-success degradation from measured harmful motion and verify physical attack evidence.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "63b2bedce6b5650e1d06c826a1c114eff1a76753feef6b0d67f3fa7df3e2099b", "source_url": "https://arxiv.org/abs/2411.13587", "source_legacy_rows": [ 80 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0081", "t": "Feasible Actor-Critic: Constrained Reinforcement Learning for Ensuring Statewise Safety", "y": 2021, "u": "https://arxiv.org/abs/2105.10682", "m": [], "k": "ma2021feasibleactorcriticconstrainedreinforcement", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0081", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "FAC models statewise feasibility through learned Lagrange multipliers and evaluates locomotion and safe exploration.", "full_text_question": "Check assumptions behind feasible-state guarantees and behavior from infeasible states.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "d104394257ffe896f9f043dd657b6072cbe73a09403221000410eb52c9905c5d", "source_url": "https://arxiv.org/abs/2105.10682", "source_legacy_rows": [ 81 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0082", "t": "Few-shot reasoning-based safe reinforcement learning framework for autonomous robot navigation", "y": 2024, "u": "https://www.inderscienceonline.com/doi/abs/10.1504/IJAAC.2024.135093", "m": [], "k": "wang2024few", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0082", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Few-shot action-safety reasoning and return-to-safe-state behavior reduce navigation training failures.", "full_text_question": "Verify recoverability assumptions and uncertainty for unexplored actions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "81ca0487f71297cc89f84a41a6b0b5f13cdc726dd38c73cc5b06cc41ebc6c4b8", "source_url": "https://www.inderscienceonline.com/doi/abs/10.1504/IJAAC.2024.135093", "source_legacy_rows": [ 82 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0083", "t": "FOSP: Fine-tuning Offline Safe Policy through World Models", "y": 2025, "u": "https://arxiv.org/abs/2407.04942", "m": [ "Learned dynamics" ], "k": "cao2025fospfinetuningofflinesafe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0083", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "FOSP combines offline learning, world-model expansion and reachability guidance for safer online adaptation, including a physical robot.", "full_text_question": "Check reachability assumptions and safety during fine-tuning under unseen constraints.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "15efcd0caeb312e35dbef1010935316c173494929ad728b557fa25807052a44b", "source_url": "https://arxiv.org/abs/2407.04942", "source_legacy_rows": [ 83 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0084", "t": "Foundation Model Driven Robotics: A Comprehensive Review", "y": 2025, "u": "https://arxiv.org/abs/2507.10087", "m": [ "Language planner" ], "k": "khan2025foundationmodeldrivenrobotics", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0084", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Reviews foundation-model robotics at system level, identifying grounding, real-time operation and safety bottlenecks.", "full_text_question": "Extract concrete assurance requirements rather than architectural feasibility claims.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "9ad9563c0b5c90023c5dd063a9aabeb1bdf612c59221fa92fb5b15c02b6cab8a", "source_url": "https://arxiv.org/abs/2507.10087", "source_legacy_rows": [ 84 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0085", "t": "Foundation Models in Robotics: Applications, Challenges, and the Future", "y": 2023, "u": "https://arxiv.org/abs/2312.07843", "m": [ "Language planner" ], "k": "firoozi2023foundationmodelsroboticsapplications", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0085", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys foundation models across perception and control, highlighting uncertainty, safety guarantees and execution latency as open issues.", "full_text_question": "Use to position the assurance-focused synthesis against general robotics surveys.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "152dc29f3aa12b01ed411b0954e243737d12273fb7070099c65289199c7240e0", "source_url": "https://arxiv.org/abs/2312.07843", "source_legacy_rows": [ 85 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0086", "t": "From Screens to Scenes: A Survey of Embodied AI in Healthcare", "y": 2025, "u": "https://arxiv.org/abs/2501.07468", "m": [ "Learned dynamics", "Language planner" ], "k": "liu2025medical", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0086", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Healthcare embodiment survey covers clinical and assistive applications and identifies safety and benchmark gaps.", "full_text_question": "Identify domain-specific harms and distinguish physical intervention from informational healthcare tasks.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "c1340dafb6f8c3514f40adb4238b0c7c25f6bca7ecddf6ec1dbbc71f9f19fff4", "source_url": "https://arxiv.org/abs/2501.07468", "source_legacy_rows": [ 86, 87 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0088", "t": "Generative AI Agents in Autonomous Machines: A Safety Perspective", "y": 2024, "u": "https://dl.acm.org/doi/10.1145/3676536.3698390", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0088", "study_type": "position", "relevance": "direct", "full_text_priority": "medium", "assessment": "A safety perspective on generative autonomous machines recommends comprehensive safety scorecards.", "full_text_question": "Assess scorecard contents and separate reporting frameworks from validated safeguards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "814ed21d8f13e3f501037691a8472400b7745b7f35b7602f78d9448352df5c4a", "source_url": "https://dl.acm.org/doi/10.1145/3676536.3698390", "source_legacy_rows": [ 88 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0089", "t": "Generative World Models of Tasks: LLM-Driven Hierarchical Scaffolding for Embodied Agents", "y": 2025, "u": "https://arxiv.org/abs/2509.04731v3", "m": [ "Learned dynamics", "Language planner" ], "k": "hill2025generativeworldmodelstasks", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0089", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "Hierarchical task environments use LLM-generated task semantics to scaffold multi-agent soccer learning.", "full_text_question": "No explicit physical-safety criterion is stated.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "4dd8d560a24403b39b3b739e3bbfbc415d8f9a7c3286ba59de48ec3e4af1ff37", "source_url": "https://arxiv.org/abs/2509.04731v3", "source_legacy_rows": [ 89 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0090", "t": "GenSafe: A Generalizable Safety Enhancer for Safe Reinforcement Learning Algorithms Based on Reduced Order Markov Decision Process Model", "y": 2025, "u": "https://arxiv.org/abs/2406.03912", "m": [], "k": "zhou2025gensafe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0090", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "GenSafe refines actions using a reduced-order proxy for CMDP costs, especially during early learning.", "full_text_question": "Check abstraction error and whether constraint satisfaction is empirical or guaranteed.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "b5aab9ea0459af8eb898d47e77b0d2dd81e1ed7ebcdbb1d7dad22435bdfa70bd", "source_url": "https://arxiv.org/abs/2406.03912", "source_legacy_rows": [ 90 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0091", "t": "GeoVLA: Empowering 3D Representations in Vision-Language-Action Models", "y": 2025, "u": "https://arxiv.org/abs/2508.09071", "m": [ "VLA policy", "Language planner" ], "k": "sun2025geovlaempowering3drepresentations", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0091", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "GeoVLA fuses point-cloud and vision-language embeddings for spatially robust manipulation.", "full_text_question": "Check explicit collision outcomes before interpreting robustness as safety.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "275f993ef7175c256da16b846535dba3bb08377b7f33724af4ba4db6fe062618", "source_url": "https://arxiv.org/abs/2508.09071", "source_legacy_rows": [ 91 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0092", "t": "GigaWorld-0: World Models as Data Engine to Empower Embodied AI", "y": 2025, "u": "https://arxiv.org/abs/2511.19861v2", "m": [ "Learned dynamics", "VLA policy" ], "k": "angen2025gigaworld", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0092", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "GigaWorld-0 generates video and 3D interaction data for VLA training and reports physical-robot generalization.", "full_text_question": "Verify physics consistency and hazard coverage independently of generated-image quality.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7d2c6f6e8e008e285deaff67d63d03a27a5942b553227e570ce4dc34407c7380", "source_url": "https://arxiv.org/abs/2511.19861v2", "source_legacy_rows": [ 92 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0093", "t": "Grounding Actions in Camera Space: Observation-Centric Vision-Language-Action Policy", "y": 2025, "u": "https://arxiv.org/abs/2508.13103", "m": [ "VLA policy" ], "k": "zhang2025groundingactionscameraspace", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0093", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "OC-VLA aligns camera observations and action coordinates using extrinsic calibration for cross-view generalization.", "full_text_question": "Assess sensitivity to calibration error and downstream unsafe actions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3048656d81bf83bb3a3a9f5f23a4a54b7477ee85378ae86fcf49936ad4cfa6ca", "source_url": "https://arxiv.org/abs/2508.13103", "source_legacy_rows": [ 93 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0094", "t": "Handling Long-Term Safety and Uncertainty in Safe Reinforcement Learning", "y": 2024, "u": "https://arxiv.org/abs/2409.12045v2", "m": [], "k": "günster2024handlinglongtermsafetyuncertainty", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0094", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Extends ATACOM with learned constraints to address long-term safety and uncertainty.", "full_text_question": "Inspect how uncertainty affects the learned constraint manifold and whether guarantees survive approximation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "01a5af1dd473f546d829f507bdb8c274c4953e0b38afa017dec0346df251f175", "source_url": "https://arxiv.org/abs/2409.12045v2", "source_legacy_rows": [ 94 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0095", "t": "Hierarchical Motion Planning Under Probabilistic Temporal Tasks and Safe-Return Constraints", "y": 2023, "u": "https://ieeexplore.ieee.org/document/10044248", "m": [], "k": "10044248", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0095", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Hierarchical MDP/LTL planning jointly synthesizes outbound and probabilistic safe-return policies with hardware validation.", "full_text_question": "Check return-policy feasibility, probabilistic labels and loss of guarantees under abstraction.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "eb3ac144e89ca1477776dadd47329aef4dafb582ee61cf0182ce102debdbbf5b", "source_url": "https://ieeexplore.ieee.org/document/10044248", "source_legacy_rows": [ 95 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0096", "t": "How Safe Will I Be Given What I Saw? Calibrated Prediction of Safety Chances for Image-Controlled Autonomy", "y": 2025, "u": "https://arxiv.org/abs/2508.09346", "m": [ "Learned dynamics" ], "k": "rev_calibration", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0097", "t": "Human Centric General Physical Intelligence for Agile Manufacturing Automation", "y": 2025, "u": "https://arxiv.org/abs/2508.11960", "m": [ "VLA policy" ], "k": "kanta2025general", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0097", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Manufacturing-focused VLA review includes uncertainty, safety and industrial deployment readiness.", "full_text_question": "Verify the described ablation study and industrial safety metrics.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7c9929b29a1c1fd6a30c865be7064f4d9c02ee9bbcc02024f42eba6bb9941da5", "source_url": "https://arxiv.org/abs/2508.11960", "source_legacy_rows": [ 97 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0098", "t": "Implicit Safe Set Algorithm for Provably Safe Reinforcement Learning", "y": 2025, "u": "https://arxiv.org/abs/2405.02754", "m": [ "Analytical dynamics" ], "k": "zhao2025implicitsafesetalgorithm", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0098", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Implicit safe-set synthesis queries black-box dynamics and claims finite-time recovery and invariance, with Safety Gym evaluation.", "full_text_question": "Clarify that model-free here still requires dynamics queries and check theorem assumptions and simulator error.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "4936b581066af684935fb35d129b98a119a2959c7d13b18b32c11c51bb6ec428", "source_url": "https://arxiv.org/abs/2405.02754", "source_legacy_rows": [ 98 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0099", "t": "In-Context Learning Enables Robot Action Prediction in LLMs", "y": 2025, "u": "https://arxiv.org/abs/2410.12782", "m": [ "Language planner" ], "k": "yin2025incontextlearningenablesrobot", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0099", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "RoboPrompt predicts robot actions from text-only in-context demonstrations without training.", "full_text_question": "Check action validity and runtime safeguards if used as a nominal-policy example.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "a4f6c93e75cb4d2ac7a741adbd59a31148875ca44173d0a43387158710c65325", "source_url": "https://arxiv.org/abs/2410.12782", "source_legacy_rows": [ 99 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0100", "t": "In-House Evaluation Is Not Enough: Towards Robust Third-Party Flaw Disclosure for General-Purpose AI", "y": 2025, "u": "https://arxiv.org/pdf/2503.16861", "m": [], "k": "longpre2025house", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0100", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "General-purpose AI flaw-disclosure proposals concern reporting and governance rather than robotic control.", "full_text_question": "Use only as evaluation-governance context.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "170ae806aad6f4dc0d2f4fa596c9f6065cad883ae0ab6b1ccfe021894e03d0f4", "source_url": "https://arxiv.org/pdf/2503.16861", "source_legacy_rows": [ 100 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0101", "t": "Introduction to AI Safety, Ethics, and Society", "y": 2024, "u": "https://arxiv.org/abs/2411.01042", "m": [], "k": "hendrycks2024introductionaisafetyethics", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0101", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "An introductory AI-risk book consolidates safety, ethics and societal concepts without a specific embodied method.", "full_text_question": "Use for definitions only.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "35c653a7b16a821348455395d27db432f34b934f781d0ea094222c855c27da95", "source_url": "https://arxiv.org/abs/2411.01042", "source_legacy_rows": [ 101 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0102", "t": "Irl-vla: Training an vision-language-action policy via reward world model", "y": 2025, "u": "https://arxiv.org/abs/2508.06571", "m": [ "Learned dynamics", "VLA policy" ], "k": "jiang2025irlvla", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0102", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "IRL-VLA trains driving policies against a reward world model balancing incidents, comfort and efficiency.", "full_text_question": "Determine whether safety is a reward trade-off or a separately enforced constraint.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "66956ec48b4be627173ddee5c2dd74353c65d47bb42961ca3de6f01da6791d61", "source_url": "https://arxiv.org/abs/2508.06571", "source_legacy_rows": [ 102 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0103", "t": "Is Sora a World Simulator? A Comprehensive Survey on General World Models and Beyond", "y": 2025, "u": "https://arxiv.org/abs/2405.03520", "m": [ "Learned dynamics" ], "k": "zhu2025soraworldsimulatorcomprehensive", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0103", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys general video, driving and agent world models and their limitations.", "full_text_question": "Separate visual simulation capability from predictive validity for safety.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "d2b91f535308d386bf818bc2e9aa105989b4c9ea4071e1596e71269fa9eacf00", "source_url": "https://arxiv.org/abs/2405.03520", "source_legacy_rows": [ 103 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0104", "t": "IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks", "y": 2025, "u": "https://arxiv.org/abs/2506.16402", "m": [], "k": "lu2025isbench", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0104", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "IS-Bench measures whether household agents mitigate dynamically emerging risks in the required procedural order.", "full_text_question": "Inspect temporal safety predicates, intermediate-step validation and safety-task trade-offs.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f12fc11a1e836892ea7e7aeb09a1f7ac74f5430266983159f710c2cba5d4c9a2", "source_url": "https://arxiv.org/abs/2506.16402", "source_legacy_rows": [ 104 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0105", "t": "Iterative Batch Reinforcement Learning via Safe Diversified Model-based Policy Search", "y": 2024, "u": "https://arxiv.org/abs/2411.09722v1", "m": [], "k": "najib2024iterativebatchreinforcementlearning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0105", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Ensemble-based iterative batch policy search adds safety and diversity objectives for informative deployment within data support.", "full_text_question": "Check the safety criterion, support estimator and guarantees during new data collection.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "772ada8c8657435993ad9108da11c43f24a3ff036f195d2f84eb46b4172d6b23", "source_url": "https://arxiv.org/abs/2411.09722v1", "source_legacy_rows": [ 105 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0106", "t": "Joint Action Language Modelling for Transparent Policy Execution", "y": 2025, "u": "https://arxiv.org/abs/2504.10055", "m": [], "k": "wulff2025joint", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0106", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "Joint language-action modeling produces action descriptions alongside control tokens in Language-Table.", "full_text_question": "Test whether verbal explanations faithfully describe executed behavior; transparency alone is not a safeguard.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "46d049545592dcbf42148c940875047af7e6243a80e6aded9b5961c0deeb550e", "source_url": "https://arxiv.org/abs/2504.10055", "source_legacy_rows": [ 106 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0107", "t": "Language Acquisition is Embodied, Interactive, Emotive: a Research Proposal", "y": 2021, "u": "https://arxiv.org/abs/2105.04633v1", "m": [ "Language planner" ], "k": "", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0107", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "A language-acquisition proposal explicitly includes a robot-dialogue system, but no safety objective.", "full_text_question": "Treat as embodied-language background rather than labeling the work wholly non-robotic.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "0d404dd4217252c218160adf136d7d4300f2e72d0d69070b3ca73ddf8c7140d6", "source_url": "https://arxiv.org/abs/2105.04633v1", "source_legacy_rows": [ 107 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0108", "t": "Language-conditioned Learning for Robotic Manipulation: A Survey", "y": 2025, "u": "https://arxiv.org/abs/2312.10807", "m": [ "Language planner" ], "k": "yao2025survey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0108", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys language-conditioned manipulation and identifies safety as a future direction.", "full_text_question": "Extract relevant planner-control interfaces and avoid treating open questions as evaluated methods.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e4ce3591b27575ac8e7094460706982d24a4e73e0e527465e142ad031e1d10f8", "source_url": "https://arxiv.org/abs/2312.10807", "source_legacy_rows": [ 108 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0109", "t": "Large language and vision-language models for robot: safety challenges, mitigation strategies and future directions", "y": 2025, "u": "https://doi.org/10.1108/IR-02-2025-0074", "m": [ "Language planner" ], "k": "hu2025large", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0109", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "Directly surveys safety challenges and mitigations for LLM/VLM robots, including monitoring and adversarial robustness.", "full_text_question": "Compare its scope and existing mitigations against the manuscript's claimed contribution.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "8c9cc9421c616db12681613738db593f02fb2f1d7e2860fc85fd87f2a6f2da45", "source_url": "https://doi.org/10.1108/IR-02-2025-0074", "source_legacy_rows": [ 109 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0110", "t": "Large model empowered embodied ai: A survey on decision-making and embodied learning", "y": 2025, "u": "https://arxiv.org/abs/2508.10399", "m": [ "Learned dynamics", "VLA policy" ], "k": "liang2025largemodelempoweredembodied", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0110", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys hierarchical and end-to-end embodied decision-making, learning and world-model integration.", "full_text_question": "Distinguish prediction placement from enforcement of safety conditions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e9047508131e753714ba7487aa8faf010fd9acf2b656d581f5de6eafb25a0323", "source_url": "https://arxiv.org/abs/2508.10399", "source_legacy_rows": [ 110 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0111", "t": "Large VLM-based Vision-Language-Action Models for Robotic Manipulation: A Survey", "y": 2025, "u": "https://scispace.com/papers/large-vlm-based-vision-language-action-models-for-robotic-ndleipk1f39p", "m": [ "Learned dynamics", "VLA policy", "Language planner" ], "k": "", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0111", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Provides a VLA manipulation taxonomy separating monolithic and hierarchical architectures and intermediate representations.", "full_text_question": "Check which interfaces support auditable safeguards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1bfb3c001903f36b8345db2c136717954040bc318683286a11c381a370d40567", "source_url": "https://scispace.com/papers/large-vlm-based-vision-language-action-models-for-robotic-ndleipk1f39p", "source_legacy_rows": [ 111 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0112", "t": "Latent Action Pretraining from Videos", "y": 2025, "u": "https://arxiv.org/abs/2410.11758", "m": [ "VLA policy" ], "k": "ye2025wmvideos", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0112", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "LAPA learns latent actions from unlabeled videos and maps them to robot actions using limited fine-tuning.", "full_text_question": "Assess whether latent actions retain hazard-relevant information and physical feasibility.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "fb30888a615c83726531a1dc4eba4061a5f66ee37366702dab7c97d6479885a2", "source_url": "https://arxiv.org/abs/2410.11758", "source_legacy_rows": [ 112 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0113", "t": "Learning 3D Persistent Embodied World Models", "y": 2025, "u": "https://arxiv.org/abs/2505.05495v1", "m": [ "Learned dynamics" ], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0113", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "A persistent 3D map conditions video world modeling to improve long-horizon consistency under partial observation.", "full_text_question": "Check uncertainty in generated unseen regions and relevance to hazard prediction.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "954b06a3b0fc2f0d2d3c9ab2c30f35cbbe88bc4f786de746f57584f72b8ee659", "source_url": "https://arxiv.org/abs/2505.05495v1", "source_legacy_rows": [ 113 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0114", "t": "Learning adaptive safety for multi-agent systems", "y": 2024, "u": "https://ieeexplore.ieee.org/abstract/document/10611037/", "m": [], "k": "berducci2024learning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0114", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "ASRL learns policy and CBF coefficients to adapt to other agents while limiting empirical costs.", "full_text_question": "Check barrier feasibility and the distinction between cost limits and statewise invariance.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "5e443cc3f1aa01c5dc77a6805efe6de9965a3f061c51afe7ae8ceefaac55d944", "source_url": "https://ieeexplore.ieee.org/abstract/document/10611037/", "source_legacy_rows": [ 114 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0115", "t": "Learning Collision-free and Torque-limited Robot Trajectories based on Alternative Safe Behaviors", "y": 2022, "u": "https://ieeexplore.ieee.org/document/10000077", "m": [], "k": "kiemel2022safe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0115", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Robot motions execute only when simulated braking continuations satisfy collision, torque and kinematic constraints.", "full_text_question": "Inspect model fidelity, stored backup validity and real-time hardware performance.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3525f60ca40f3b89762a61510a7c10e11b221dd0cb25c0792f83769b9129482e", "source_url": "https://ieeexplore.ieee.org/document/10000077", "source_legacy_rows": [ 115 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0116", "t": "Learning to be Safe: Deep RL with a Safety Critic", "y": 2020, "u": "https://arxiv.org/abs/2010.14603", "m": [], "k": "srinivasan2020learningsafedeeprl", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0116", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A transferred safety critic reduces incidents while learning new navigation, locomotion and manipulation tasks.", "full_text_question": "Check transfer assumptions, critic calibration and whether risk estimates can fail under novel hazards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f9902c20a8026d8e2ef8ce31e08696d7138cfeaa503fd029a04a7b2aa09c01e7", "source_url": "https://arxiv.org/abs/2010.14603", "source_legacy_rows": [ 116 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0117", "t": "Learning to walk autonomously via reset-free quality-diversity", "y": 2022, "u": "https://dl.acm.org/doi/10.1145/3512290.3528715", "m": [], "k": "lim2022reset", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0117", "study_type": "method", "relevance": "direct", "full_text_priority": "medium", "assessment": "Reset-free quality-diversity learning selects imagined behaviors for automatic recovery in obstacle-constrained locomotion zones.", "full_text_question": "Verify safety constraints and behavior-selection checks beyond successful resets.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1aa667c2926d378d12094e65bd99924cb281d4f28ee3497bddb7790b478de14f", "source_url": "https://dl.acm.org/doi/10.1145/3512290.3528715", "source_legacy_rows": [ 117 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0118", "t": "Learning with safety constraints: Sample complexity of reinforcement learning for constrained mdps", "y": 2021, "u": "https://ojs.aaai.org/index.php/AAAI/article/view/16937", "m": [], "k": "hasanzadezonuzy2021learning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0118", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Analyzes PAC sample complexity for unknown CMDPs with objective and constraint accuracy requirements.", "full_text_question": "Distinguish final-policy feasibility guarantees from safety during sample collection.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7ae9845d06fc4268428c94ea5b080a8772e8eefd4bccd5406586b8f97b40cad6", "source_url": "https://ojs.aaai.org/index.php/AAAI/article/view/16937", "source_legacy_rows": [ 118 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0119", "t": "Manipulation Facing Threats: Evaluating Physical Vulnerabilities in End-to-End Vision Language Action Models", "y": 2025, "u": "https://arxiv.org/abs/2409.13174", "m": [ "VLA policy", "Language planner" ], "k": "cheng2025manipulationfacingthreatsevaluating", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0119", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "PVEP evaluates VLA vulnerabilities to OOD scenes, visual prompts and adversarial patches.", "full_text_question": "Check whether physical harm is measured separately from task-performance changes.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "23e8bd8ece47bf27d1ff79a6047ea893fd57ffcc1035cc1767524ea10b63bd70", "source_url": "https://arxiv.org/abs/2409.13174", "source_legacy_rows": [ 119 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0120", "t": "Mechanistic interpretability for steering vision-language-action models", "y": 2025, "u": "https://arxiv.org/abs/2509.00328", "m": [ "VLA policy", "Language planner" ], "k": "haon2025mechanistic", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0120", "study_type": "method", "relevance": "context", "full_text_priority": "medium", "assessment": "Activation steering changes VLA speed and direction in simulation and on a robot using interpretable internal directions.", "full_text_question": "Does behavioral steering reliably enforce a specified safety limit?", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "81eb9f679b763b79f6397660d241a14fc7efc9254329f751fba1f78e95781c4f", "source_url": "https://arxiv.org/abs/2509.00328", "source_legacy_rows": [ 120 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0121", "t": "Meta-Learning Priors for Safe Bayesian Optimization", "y": 2023, "u": "https://arxiv.org/abs/2210.00762v3", "m": [ "Learned dynamics" ], "k": "rothfuss2023metalearningpriorssafebayesian", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0121", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Meta-learned priors and empirical uncertainty criteria accelerate safe Bayesian optimization on a motion system.", "full_text_question": "Check prior calibration under transfer and the conditions needed for safety-compliant exploration.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3e070005650891b7a808677ed96265a2e240d33433cb3ac291ea5f8868e3148c", "source_url": "https://arxiv.org/abs/2210.00762v3", "source_legacy_rows": [ 121 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0122", "t": "Minimalist Vision with Freeform Pixels", "y": 2024, "u": "https://arxiv.org/abs/2501.00142v1", "m": [], "k": "klotz2024pixels", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0122", "study_type": "method", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Task-designed low-pixel cameras address sensing efficiency and privacy, without an embodied-controller safety evaluation.", "full_text_question": "Exclude from safety-effectiveness comparisons; sensing context only if necessary.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "cd4c6050a5244f78be35f369452735a3d6ee8e1511377e6e5eb05e0e58fa1a79", "source_url": "https://arxiv.org/abs/2501.00142v1", "source_legacy_rows": [ 122 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0123", "t": "MoDem-V2: Visuo-Motor World Models for Real-World Robot Manipulation", "y": 2024, "u": "https://arxiv.org/abs/2309.14236", "m": [ "Learned dynamics" ], "k": "lancaster2024modemv2visuomotorworldmodels", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0123", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "MoDem-V2 uses demonstrations, exploration centering, agency handover and ensembles for contact-rich physical robot learning.", "full_text_question": "Extract actual safety-fault measurements and the assumptions behind safe exploration.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f239b84fb10161efa3c62716cdb53bb2f44d42e48338e613c91030292b787a4b", "source_url": "https://arxiv.org/abs/2309.14236", "source_legacy_rows": [ 123 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0124", "t": "MoveRL: To a Safer Robotic Reinforcement Learning Environment", "y": 2021, "u": "https://link.springer.com/chapter/10.1007/978-3-030-93842-0_14", "m": [], "k": "liu2021moverl", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0124", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "MoveRL inserts collision-aware MoveIt planning between RL actions and robot execution, with moving-obstacle simulation tests.", "full_text_question": "Check planner latency, collision-check resolution and the gap between simulated and physical validation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ab469b0fd7b9fc9dc6747e9ba553d69c7de520104592d77dd1214cad1a20e747", "source_url": "https://link.springer.com/chapter/10.1007/978-3-030-93842-0_14", "source_legacy_rows": [ 124 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0125", "t": "Multi-agent embodied ai: Advances and future directions", "y": 2025, "u": "https://arxiv.org/abs/2505.05108", "m": [ "Language planner" ], "k": "feng2025multi", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0125", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys multi-agent embodiment, collaboration and adaptation in open environments.", "full_text_question": "Identify explicit inter-agent safety constraints rather than general coordination challenges.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "d984c9bbc82ce405206771430abf9e08677b9496701ab00dae58edc56349ed83", "source_url": "https://arxiv.org/abs/2505.05108", "source_legacy_rows": [ 125 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0126", "t": "Multi-Modal Masked Autoencoders for Medical Vision-and-Language Pre-Training", "y": 2022, "u": "https://arxiv.org/abs/2209.07098v1", "m": [], "k": "", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0126", "study_type": "method", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Medical image-text masked autoencoding evaluates representation tasks without robot control.", "full_text_question": "Exclude from embodied safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "a08ca330a0ff757035f1da025cae14755d59a8d9625c1aabcc0c1c09b6e75441", "source_url": "https://arxiv.org/abs/2209.07098v1", "source_legacy_rows": [ 126 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0127", "t": "Multi-objective Model-based Policy Search for Data-efficient Learning with Sparse Rewards", "y": 2018, "u": "https://arxiv.org/abs/1806.09351v3", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0127", "study_type": "method", "relevance": "context", "full_text_priority": "medium", "assessment": "Multi-DEX trades off novelty, reward and model accuracy for sparse-reward robot exploration.", "full_text_question": "Model-confidence preference is not an explicit harm constraint; check whether safety is measured.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "646924bc9099415ba8c1eeeb8b936ef7f04935da01e34a0d4e5f04c87a0823c1", "source_url": "https://arxiv.org/abs/1806.09351v3", "source_legacy_rows": [ 127 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0128", "t": "Multi-View Dreaming: Multi-View World Model with Contrastive Learning", "y": 2023, "u": "https://arxiv.org/abs/2203.11024", "m": [ "Learned dynamics" ], "k": "kinose2023multi", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0128", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "Multi-view Dreaming fuses latent distributions to mitigate occlusion in robot control.", "full_text_question": "Check whether improved observability translates to calibrated hazard uncertainty.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "6cb683edc464b2fa90421621fb0453664936e1415e64d066735a8d612fb15555", "source_url": "https://arxiv.org/abs/2203.11024", "source_legacy_rows": [ 128 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0129", "t": "Navigating the sociotechnical labyrinth: Dynamic certification for responsible embodied AI", "y": 2025, "u": "https://scispace.com/papers/navigating-the-sociotechnical-labyrinth-dynamic-77del0zsp2wt", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0129", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "Dynamic certification is proposed as an adaptive governance framework for evolving embodied systems.", "full_text_question": "Extract concrete assurance-update obligations without presenting certification as a technical guarantee.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7940e4850d3ee7aa2099a13b71a985f79c578a147a3fbdc546876f93841e2768", "source_url": "https://scispace.com/papers/navigating-the-sociotechnical-labyrinth-dynamic-77del0zsp2wt", "source_legacy_rows": [ 129 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0130", "t": "NeuroAI for AI Safety", "y": 2025, "u": "https://arxiv.org/abs/2411.18526", "m": [], "k": "mineault2025neuroaiaisafety", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0130", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "Neuroscience-inspired safety roadmap includes robust sensory-motor systems but no evaluated robot safeguard.", "full_text_question": "Use as a research perspective rather than exclude it solely as digital-only.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1c837e384d85cd7d798e77b2e96c62b588fe5306e4257b25c44073bb439d263a", "source_url": "https://arxiv.org/abs/2411.18526", "source_legacy_rows": [ 130 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0131", "t": "Null Space Based Efficient Reinforcement Learning with Hierarchical Safety Constraints", "y": 2021, "u": "https://ieeexplore.ieee.org/document/9568848", "m": [], "k": "yang2021safety", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0131", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Null-space RL prioritizes joint-limit and obstacle constraints over learned tasks in redundant robot simulations.", "full_text_question": "Check rank changes, competing constraints and assumptions behind guaranteed safe exploration.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "079c1880e254062c41c5eca6fa31a2e94d0a5d681c01862f28bdd23b5e5eeb1d", "source_url": "https://ieeexplore.ieee.org/document/9568848", "source_legacy_rows": [ 131 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0132", "t": "Object Detection with Multimodal Large Vision-Language Models: An In-depth Review", "y": 2025, "u": "https://arxiv.org/abs/2508.19294v2", "m": [ "Language planner" ], "k": "", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0132", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "LVLM object-detection survey mentions robotic applications but evaluates perception capability rather than control safety.", "full_text_question": "Use only for perception background.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1cca0e6508856302b6b4b552b2fc78c48eeec6b1230729f4bad9fa1224120a87", "source_url": "https://arxiv.org/abs/2508.19294v2", "source_legacy_rows": [ 132 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0133", "t": "On the Safety Concerns of Deploying LLMs/VLMs in Robotics: Highlighting the Risks and Vulnerabilities", "y": 2025, "u": "https://arxiv.org/abs/2402.10340", "m": [ "Language planner" ], "k": "wu2025vulnerabilityllmvlmcontrolledrobotics", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0133", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Formalizes attacks on LLM/VLM robotics and reports performance degradation from prompt and perceptual changes.", "full_text_question": "Identify actual hazardous outcomes separately from general effectiveness loss.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "0294a9bd673c0ad69347533372288d3ae3fbd7e74e9c736b5fbb2d59a19457bb", "source_url": "https://arxiv.org/abs/2402.10340", "source_legacy_rows": [ 133, 244 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0134", "t": "On-Policy Robot Imitation Learning from a Converging Supervisor", "y": 2019, "u": "https://arxiv.org/abs/1907.03423v7", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0134", "study_type": "method", "relevance": "context", "full_text_priority": "medium", "assessment": "Imitation learning from an improving supervisor provides regret guarantees and faster policy evaluation.", "full_text_question": "Do not interpret regret guarantees as constraints on unsafe behavior.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7fd8f8f7a21730c0b894b5faf9069e25314817d6af550db10eac5ec770596c09", "source_url": "https://arxiv.org/abs/1907.03423v7", "source_legacy_rows": [ 134 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0135", "t": "One-Shot Reinforcement Learning for Robot Navigation with Interactive Replay", "y": 2017, "u": "https://arxiv.org/abs/1711.10137v2", "m": [ "Learned dynamics" ], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0135", "study_type": "method", "relevance": "context", "full_text_priority": "medium", "assessment": "Interactive replay from one environment traversal enables visual navigation transfer with limited real-world data.", "full_text_question": "Check collision evaluation and coverage of unobserved hazards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "292631d6ee6cfceaab2ad9553e4c017851ad3ea70a8648ff724fbe9a9d9d6615", "source_url": "https://arxiv.org/abs/1711.10137v2", "source_legacy_rows": [ 135 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0136", "t": "OphGLM: Training an Ophthalmology Large Language-and-Vision Assistant based on Instructions and Dialogue", "y": 2023, "u": "https://arxiv.org/abs/2306.12174v2", "m": [ "Language planner" ], "k": "", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0136", "study_type": "method", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "OphGLM provides ophthalmic diagnosis and dialogue from images, not embodied actuation.", "full_text_question": "Exclude from robotic safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "b87f5351aa5cd48064a5d2b62f381ff95bd29286bd447acfe9270689a126f327", "source_url": "https://arxiv.org/abs/2306.12174v2", "source_legacy_rows": [ 136 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0137", "t": "Opinion: A Unified World Model is the cornerstone for integrating perception, reasoning, and decision-making in embodied AI", "y": 2025, "u": "https://openreview.net/forum?id=WjmfSw2eg9", "m": [ "Learned dynamics", "Language planner" ], "k": "xu2025opinion", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0137", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "A unified-world-model opinion proposes memory, differentiable imagination and uncertainty-aware evaluation.", "full_text_question": "Separate research agenda from demonstrated predictive safety mechanisms.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "564a3d0af3c44cd9974af08349236c57d7897526e409d32c81700ce3bab4f389", "source_url": "https://openreview.net/forum?id=WjmfSw2eg9", "source_legacy_rows": [ 137 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0138", "t": "PhysicalAgent: Towards General Cognitive Robotics with Foundation World Models", "y": 2025, "u": "https://arxiv.org/abs/2509.13903", "m": [ "Learned dynamics" ], "k": "lykov2025physicalagent", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0138", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "PhysicalAgent generates video trajectories and replans after failures, improving cumulative manipulation success.", "full_text_question": "Determine whether initial failures cause harm; recovery success does not establish safe execution.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ec221824366d7ecd86c0e380fa776ff4692fb33c838a3dd60b32f6acb75e3830", "source_url": "https://arxiv.org/abs/2509.13903", "source_legacy_rows": [ 138 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0139", "t": "Planning for potential: efficient safe reinforcement learning", "y": 2022, "u": "https://link.springer.com/article/10.1007/s10994-022-06143-6", "m": [], "k": "den2022planning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0139", "study_type": "method", "relevance": "direct", "full_text_priority": "medium", "assessment": "Symbolic safety automata guide RL independently of reward, with grid-world and recommender evaluation.", "full_text_question": "Assess transfer of abstract constraint satisfaction to robot dynamics; no physical validation is stated.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e0bdfc573e93f9945f38e6bb5548abc1d69763d19a0445f408b1fc3eb48c1007", "source_url": "https://link.springer.com/article/10.1007/s10994-022-06143-6", "source_legacy_rows": [ 139 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0140", "t": "Provable Safe Reinforcement Learning with Binary Feedback", "y": 2022, "u": "https://arxiv.org/abs/2210.14492", "m": [], "k": "bennett2022provablesafereinforcementlearning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0140", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "SABRE uses binary safety-oracle queries and active learning to claim safe exploration and a near-optimal safe policy.", "full_text_question": "Check oracle correctness, initial information and assumptions needed for zero unsafe actions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "805a28215c617d94edb409f2b282596636c1404bd2f180492141bb1c03d01c78", "source_url": "https://arxiv.org/abs/2210.14492", "source_legacy_rows": [ 140 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0141", "t": "Provably Safe Deep Reinforcement Learning for Robotic Manipulation in Human Environments", "y": 2022, "u": "https://arxiv.org/abs/2205.06311", "m": [], "k": "thumm2022provablysafedeepreinforcement", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0141", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Reachability shielding stops manipulators before human contact, evaluated using motion-capture-driven simulation.", "full_text_question": "Inspect human-motion bounds, braking feasibility and the precise standard-related claim.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "c2da1c2b759674ffc4fc1f6681e3c407665709b78608c4f46e0b080b779b0ae5", "source_url": "https://arxiv.org/abs/2205.06311", "source_legacy_rows": [ 141 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0142", "t": "Provably Safe Reinforcement Learning: Conceptual Analysis, Survey, and Benchmarking", "y": 2022, "u": "https://arxiv.org/abs/2205.06750", "m": [], "k": "rev_comparison", "role": "secondary", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0143", "t": "Pure vision language action (vla) models: A comprehensive survey", "y": 2025, "u": "https://arxiv.org/abs/2509.19012", "m": [ "VLA policy", "Language planner" ], "k": "zhang2025purevisionlanguageaction", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0143", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "A broad VLA taxonomy covers autoregressive, diffusion, reinforcement and hybrid policies.", "full_text_question": "Check overlap with existing background reviews and relevant safety-specific content.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "a301e3184e25a398098d2cc8cb1130432f0216edecb9bfc615bf1b4e5c49669c", "source_url": "https://arxiv.org/abs/2509.19012", "source_legacy_rows": [ 143 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0144", "t": "RaceVLA: VLA-based Racing Drone Navigation with Human-like Behaviour", "y": 2025, "u": "https://arxiv.org/abs/2503.02572", "m": [ "VLA policy" ], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0144", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "RaceVLA adapts language-conditioned control to drone racing and reports mixed generalization gains.", "full_text_question": "Check collision rates and flight safety independently of speed and task metrics.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "812701197a41f0af48b509b582bf3151e57e51d6196068374a5c871424d1c62a", "source_url": "https://arxiv.org/abs/2503.02572", "source_legacy_rows": [ 144 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0145", "t": "Rationalvla: A rational vision-language-action model with dual system", "y": 2025, "u": "https://arxiv.org/abs/2506.10826", "m": [ "VLA policy", "Language planner" ], "k": "song2025rationalvla", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0145", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "RationalVLA combines a defective-instruction benchmark with dual-system rejection and physical manipulation trials.", "full_text_question": "Inspect safety-related instruction labels, false rejections and downstream execution hazards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ccc05cc29ad302bcd431f9214e2c31fd1d32cd074037991f8b90ad797be28fe6", "source_url": "https://arxiv.org/abs/2506.10826", "source_legacy_rows": [ 145 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0146", "t": "Recipe for Vision-Language-Action Models in Robotic Manipulation: A Survey", "y": 2025, "u": "https://www.techrxiv.org/doi/full/10.36227/techrxiv.175624610.06665789", "m": [ "VLA policy", "Language planner" ], "k": "motoda2025recipe", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0146", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "VLA survey emphasizes training, data and generalization while identifying long-horizon recovery and deployment limits.", "full_text_question": "Use to contextualize failure recovery rather than demonstrate safe recovery.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "cdfb400805f608974e61f951e49fd77224920e42f4e7a570fbfbb03c89403d2e", "source_url": "https://www.techrxiv.org/doi/full/10.36227/techrxiv.175624610.06665789", "source_legacy_rows": [ 146, 230 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0147", "t": "Redefining Safety in Light of Human-Robot Interaction: A Critical Review of Current Standards and Regulations", "y": 2021, "u": "https://www.frontiersin.org/journals/chemical-engineering/articles/10.3389/fceng.2021.666237/full", "m": [], "k": "martinetti2021redefining", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0147", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Critically broadens robot safety to interaction, psychosocial, cyber, temporal and societal dimensions.", "full_text_question": "Identify which dimensions the technical synthesis operationalizes and which remain outside it.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "96cd41c734bd732565e7b06a209cac3233e7303c7edfb42b94b14ebaffab86af", "source_url": "https://www.frontiersin.org/journals/chemical-engineering/articles/10.3389/fceng.2021.666237/full", "source_legacy_rows": [ 147 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0148", "t": "Reducing Safety Interventions in Provably Safe Reinforcement Learning", "y": 2023, "u": "https://arxiv.org/abs/2303.03339", "m": [], "k": "thumm2023reducingsafetyinterventionsprovably", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0148", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Action replacement and projection reduce failsafe interventions while reporting zero violations in tested tasks.", "full_text_question": "Check whether intervention reduction preserves the underlying shield assumptions and behavior smoothness.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "9799219c5d6b055a2bab00d8a9bcc45a5a3319af3af77418948377ef8437f898", "source_url": "https://arxiv.org/abs/2303.03339", "source_legacy_rows": [ 148 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0149", "t": "Reinforcement Learning Controllers for Soft Robots using Learned Environments", "y": 2024, "u": "https://arxiv.org/abs/2410.18519v2", "m": [ "Analytical dynamics" ], "k": "berdica2024soft", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0149", "study_type": "method", "relevance": "direct", "full_text_priority": "medium", "assessment": "Soft-robot model learning uses a mean-reverting exploration protocol intended to collect physically safe data.", "full_text_question": "Verify safe actuation bounds and whether learned-model rollout errors can induce unsafe control.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "730e3d7a5c947fbcc603b4dfff987ade789a5c16ad12ea59fa68f14d0e0b8884", "source_url": "https://arxiv.org/abs/2410.18519v2", "source_legacy_rows": [ 149 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0150", "t": "Reinforcement Learning in a Safety-Embedded MDP with Trajectory Optimization", "y": 2024, "u": "https://arxiv.org/abs/2310.06903", "m": [], "k": "yang2024reinforcementlearningsafetyembeddedmdp", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0150", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Trajectory optimization transforms RL outputs into constrained trajectories, including physical box-pushing around obstacles.", "full_text_question": "Check optimizer feasibility, model assumptions and the meaning of near-zero versus guaranteed violations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1264c2c38182543ceb896b60fc50802903ccdc430d78cf5aa1d13e8e47e750de", "source_url": "https://arxiv.org/abs/2310.06903", "source_legacy_rows": [ 150 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0151", "t": "Repairing Neural Networks for Safety in Robotic Systems using Predictive Models", "y": 2024, "u": "https://arxiv.org/abs/2411.04408", "m": [], "k": "majd2024repairingneuralnetworkssafety", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0151", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Predictive-model-constrained neural policy repair is evaluated in navigation and physical prostheses.", "full_text_question": "Inspect the repair domain, prediction error and whether constraints hold outside demonstrated states.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "c51aff5ab8d4c8cf48fb34d5b0beffaabeb298e8a56740b44dd70fe1b709399a", "source_url": "https://arxiv.org/abs/2411.04408", "source_legacy_rows": [ 151 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0152", "t": "Research on Task Planning Methods for Embodied Intelligence Based on Large Language Models", "y": 2025, "u": "https://ieeexplore.ieee.org/document/11168047", "m": [ "Language planner" ], "k": "niu2025embodied", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0152", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys LLM task planning and multi-agent structures, naming safety verification as an unresolved challenge.", "full_text_question": "Identify concrete verification approaches rather than broad planning claims.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7efe0e146c7ef81599d2c6e9221cf85488c3754b9295937e49028707b20d15de", "source_url": "https://ieeexplore.ieee.org/document/11168047", "source_legacy_rows": [ 152 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0153", "t": "Residual Skill Policies: Learning an Adaptable Skill-based Action Space for Reinforcement Learning for Robotics", "y": 2022, "u": "https://arxiv.org/abs/2211.02231v1", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0153", "study_type": "method", "relevance": "context", "full_text_priority": "low", "assessment": "Residual skill policies accelerate exploration and adaptation across manipulation tasks.", "full_text_question": "No explicit safety constraint is stated for residual actions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "85100f324ee7e97fe4de2c3c5ec2955a39d4f622db0c78344071db59f5be8668", "source_url": "https://arxiv.org/abs/2211.02231v1", "source_legacy_rows": [ 153 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0154", "t": "RoboBallet: Planning for Multi-Robot Reaching with Graph Neural Networks and Reinforcement Learning", "y": 2025, "u": "https://arxiv.org/abs/2509.05397v1", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0154", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "RoboBallet jointly allocates tasks and plans collision-free multi-arm motion through graph-based RL.", "full_text_question": "Check collision certification, real-time replanning claims and whether evaluation includes hardware.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "b95222771ebf5fa678cd549811be33e85a11687ba6525bd780232a21f63ce75b", "source_url": "https://arxiv.org/abs/2509.05397v1", "source_legacy_rows": [ 154 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0155", "t": "RoboNurse-VLA: Robotic Scrub Nurse System based on Vision-Language-Action Model", "y": 2024, "u": "https://arxiv.org/abs/2409.19590", "m": [ "VLA policy", "Language planner" ], "k": "li2024robonursevlaroboticscrubnurse", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0155", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "RoboNurse-VLA performs voice-conditioned surgical-instrument grasping and handover with task-success evaluation.", "full_text_question": "Identify patient/operator harm metrics, contact limits and actual clinical versus laboratory settings.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "c394b4adab305e04d9d72344f4d12f8c83d1a96521f308eb0ddc84e2feb33596", "source_url": "https://arxiv.org/abs/2409.19590", "source_legacy_rows": [ 155 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0156", "t": "RoboScape: Physics-informed Embodied World Model", "y": 2025, "u": "https://arxiv.org/abs/2506.23135", "m": [ "Learned dynamics" ], "k": "shang2025roboscape", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0156", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "RoboScape jointly learns video, depth and keypoint dynamics for physically plausible robot world models.", "full_text_question": "Verify contact-physics fidelity and hazard-prediction accuracy beyond image realism.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1d4098147891ddd2812ad3110a8c9f5bd1e15aa145a9e33db8e20cd02ab3ef87", "source_url": "https://arxiv.org/abs/2506.23135", "source_legacy_rows": [ 156 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0157", "t": "Robot Learning: A Tutorial", "y": 2025, "u": "https://arxiv.org/abs/2510.12403v1", "m": [], "k": "capuano2025robotlearningtutorial", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0157", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "A practical robot-learning tutorial covers RL, cloning and generalist policies.", "full_text_question": "Use as introductory background only.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e00954f9e5d50cb1b9e5bed0b54ddd5c9bdb0f88ff05e4db222e58f702bf9187", "source_url": "https://arxiv.org/abs/2510.12403v1", "source_legacy_rows": [ 157 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0158", "t": "Robot reinforcement learning on the constraint manifold", "y": 2022, "u": "https://proceedings.mlr.press/v164/liu22c.html", "m": [], "k": "liu2022manifold", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0158", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Constraint-manifold robot learning uses known models and constraints to maintain safe exploration in simulation.", "full_text_question": "Inspect manifold construction, feasibility and handling of approximate dynamics.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "51d33865996da415eafbabfb18a92754d161fb0f8f0cab465b5faaf08b5b1889", "source_url": "https://proceedings.mlr.press/v164/liu22c.html", "source_legacy_rows": [ 158 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0159", "t": "Robotic Control via Embodied Chain-of-Thought Reasoning", "y": 2025, "u": "https://arxiv.org/abs/2407.08693", "m": [ "VLA policy", "Language planner" ], "k": "zawalski2025roboticcontrolembodiedchainofthought", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0159", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "ECoT grounds multi-step VLA reasoning in plans, motions and sensory features, improving task generalization and correction.", "full_text_question": "Check whether interpretable reasoning predicts or prevents hazardous actions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "c7e34eec437d42a3d8cbbdcc6e5ebaa7101755f1d31cfee3b0193744b140a817", "source_url": "https://arxiv.org/abs/2407.08693", "source_legacy_rows": [ 159 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0160", "t": "Robotic Manipulation Datasets for Offline Compositional Reinforcement Learning", "y": 2023, "u": "https://arxiv.org/abs/2307.07091v2", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0160", "study_type": "benchmark", "relevance": "context", "full_text_priority": "low", "assessment": "Compositional offline manipulation datasets test task transfer and skill recombination.", "full_text_question": "No harm-specific annotations or safety constraints are stated.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "db897e2068cb9b10d83281669cf037e949e78e2a94e430f539767c2955e6e2a5", "source_url": "https://arxiv.org/abs/2307.07091v2", "source_legacy_rows": [ 160 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0161", "t": "Robotic world models—conceptualization, review, and engineering best practices", "y": 2023, "u": "https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2023.1253049/full", "m": [ "Learned dynamics" ], "k": "sakagami2023robotic", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0161", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Reviews four decades of robotic world models and proposes consistent terminology and engineering design principles.", "full_text_question": "Use to avoid conflating learned dynamics with all robotic world representations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "856e2948db0262c0819bcbd38d60389365be932ea1f881977548fb6a7a4f1010", "source_url": "https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2023.1253049/full", "source_legacy_rows": [ 161 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0162", "t": "ROS-LLM: A ROS framework for embodied AI with task feedback and structured reasoning", "y": 2024, "u": "https://arxiv.org/abs/2406.19741", "m": [ "Language planner" ], "k": "mower2024rosllmrosframeworkembodied", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0162", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "ROS-LLM executes structured behaviors with human/environment feedback and reflection.", "full_text_question": "Check permission boundaries, action checks and hazards during automated execution.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "6086ea56ad8d65c540681e69831c6cc45a76cefd3cc5a90c9bcf44286082550c", "source_url": "https://arxiv.org/abs/2406.19741", "source_legacy_rows": [ 162 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0163", "t": "RynnEC: Bringing MLLMs into Embodied World", "y": 2025, "u": "https://arxiv.org/abs/2508.14160v2", "m": [ "Language planner" ], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0163", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "RynnEC learns region-level video perception and spatial reasoning with a cognition benchmark.", "full_text_question": "No control-loop safety outcome is specified.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "8a97556cd74d72a474eacfeefc661aba4a1dc68dde53df3acbb9ebf85d200dac", "source_url": "https://arxiv.org/abs/2508.14160v2", "source_legacy_rows": [ 163 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0164", "t": "Safe & Accurate at Speed with Tendons: A Robot Arm for Exploring Dynamic Motion", "y": 2024, "u": "https://arxiv.org/abs/2307.02654", "m": [], "k": "guist2024safeaccuratespeed", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0164", "study_type": "method", "relevance": "context", "full_text_priority": "medium", "assessment": "A tendon-driven compliant arm reduces peak collision forces through mechanical design and supports learned dynamic motion.", "full_text_question": "Use as complementary physical risk mitigation, not a predictive-controller guarantee.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "2eff7af38db06d7db17932b188a4ad3fa018d15b7dd490654a208b16bf2fba83", "source_url": "https://arxiv.org/abs/2307.02654", "source_legacy_rows": [ 164 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0165", "t": "Safe CoR: A Dual-Expert Approach to Integrating Imitation Learning and Safe Reinforcement Learning Using Constraint Rewards", "y": 2024, "u": "https://scispace.com/papers/safe-cor-a-dual-expert-approach-to-integrating-imitation-3v0yrw0jju", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0165", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Safe CoR combines performance and safety demonstrations through constraint rewards and reports physical Jackal violation reductions.", "full_text_question": "Check residual violations and whether constraint rewards enforce hard limits.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "c77d8e8af615a7a2253473cddc0067c447dbbd2fcc90367b6576028eaac18a02", "source_url": "https://scispace.com/papers/safe-cor-a-dual-expert-approach-to-integrating-imitation-3v0yrw0jju", "source_legacy_rows": [ 165 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0166", "t": "Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning", "y": 2021, "u": "https://arxiv.org/abs/2108.06266", "m": [], "k": "rev_brunke", "role": "secondary", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0167", "t": "Safe Reinforcement Learning for a Robot Being Pursued but with Objectives Covering More Than Capture-avoidance", "y": 2022, "u": "https://scispace.com/papers/safe-reinforcement-learning-for-a-robot-being-pursued-but-1wt8gs4f", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0167", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A pursuit-evasion safe-RL framework claims guarantees while maintaining additional robot objectives, with simulations and experiments.", "full_text_question": "Identify pursuer assumptions, capture-avoidance conditions and hardware scope.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "c218162b3d5ce06e438b256405a3e7b04de49c3b3bcbe878a6867387449226e7", "source_url": "https://scispace.com/papers/safe-reinforcement-learning-for-a-robot-being-pursued-but-1wt8gs4f", "source_legacy_rows": [ 167 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0168", "t": "Safe Reinforcement Learning for Learning from Human Demonstrations", "y": 2023, "u": "https://scispace.com/papers/safe-reinforcement-learning-for-learning-from-human-x6ixihx202", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0168", "study_type": "method", "relevance": "direct", "full_text_priority": "medium", "assessment": "Combines learning from human behavior with safe exploration on a planar robot.", "full_text_question": "Inspect the unspecified safety mechanism, constraint definitions and simulation-versus-hardware evidence.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "20a39f17fb0359d28b6c9bc64d910a4d89611adc1c52ea474e1ae4fb60bbf920", "source_url": "https://scispace.com/papers/safe-reinforcement-learning-for-learning-from-human-x6ixihx202", "source_legacy_rows": [ 168 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0169", "t": "Safe reinforcement learning for probabilistic reachability and safety specifications: A Lyapunov-based approach", "y": 2020, "u": "https://scispace.com/papers/safe-reinforcement-learning-for-probabilistic-reachability-58gktwzta3", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0169", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Lyapunov-based reachability learning expands a probabilistic safe set and shields exploration.", "full_text_question": "Check initial safe-policy assumptions and whether guarantees survive deep actor-critic relaxation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "52c86fb778562f06b73b7d98c23e5ffd9928c192dd27e818580a74eec4ed5333", "source_url": "https://scispace.com/papers/safe-reinforcement-learning-for-probabilistic-reachability-58gktwzta3", "source_legacy_rows": [ 169 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0170", "t": "Safe Reinforcement Learning for Vision-Based Robotic Manipulation in Human-Centered Environments", "y": 2025, "u": "https://link.springer.com/article/10.1007/s41315-025-00507-6", "m": [], "k": "khan2025safe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0170", "study_type": "method", "relevance": "direct", "full_text_priority": "medium", "assessment": "Constrained PPO reduces average safety cost in limited visual manipulation simulations while trading off reward.", "full_text_question": "Do not interpret nonzero average costs as statewise safety or physical validation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "4ea3be43895afa5e36cd661d9b8892b23e0e365542bb7fde12ba22fa0e9f47a2", "source_url": "https://link.springer.com/article/10.1007/s41315-025-00507-6", "source_legacy_rows": [ 170 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0171", "t": "Safe Reinforcement Learning in a Simulated Robotic Arm", "y": 2023, "u": "https://scispace.com/papers/safe-reinforcement-learning-in-a-simulated-robotic-arm-58qnpvf7mp", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0171", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "medium", "assessment": "Extends Safety Gym with a simulated Panda arm and pilot constrained-PPO comparisons.", "full_text_question": "Check constraint definitions, training costs and benchmark reproducibility.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "453aecdf90ce9e9b19059c5e34ed343aba2dfe7d334faa61266f56d516b8b603", "source_url": "https://scispace.com/papers/safe-reinforcement-learning-in-a-simulated-robotic-arm-58qnpvf7mp", "source_legacy_rows": [ 171, 172 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0173", "t": "Safe Reinforcement Learning of Dynamic High-Dimensional Robotic Tasks: Navigation, Manipulation, Interaction", "y": 2023, "u": "https://scispace.com/papers/safe-reinforcement-learning-of-dynamic-high-dimensional-2qwzdz7m", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0173", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Constraint-manifold exploration handles learned collision constraints and reports TIAGo++ physical deployment.", "full_text_question": "Verify uncertainty in learned constraints and distinctions from related ATACOM publications.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "779020118855dff05c829d6435b93baeddf5c1e2aa6e236ca77c6c34494fa325", "source_url": "https://scispace.com/papers/safe-reinforcement-learning-of-dynamic-high-dimensional-2qwzdz7m", "source_legacy_rows": [ 173, 174, 175 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0176", "t": "Safe Reinforcement Learning on the Constraint Manifold: Theory and Applications", "y": 2025, "u": "https://scispace.com/papers/safe-reinforcement-learning-on-the-constraint-manifold-1l8w56ygxq", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0176", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Differential-geometric safe action spaces support constrained robot learning, demonstrated in physical air hockey.", "full_text_question": "Check invariance conditions, discretization and relation to earlier constraint-manifold papers.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "539189bffb7122ca1d56f447f914d31c6b7bcdc32e1c80943d1bf4e1b6b0e664", "source_url": "https://scispace.com/papers/safe-reinforcement-learning-on-the-constraint-manifold-1l8w56ygxq", "source_legacy_rows": [ 176 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0177", "t": "Safe reinforcement learning using robust control barrier functions", "y": 2025, "u": "https://ieeexplore.ieee.org/abstract/document/9928337/", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0177", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A differentiable robust-CBF layer couples action projection with model-based learning and modular reward policies.", "full_text_question": "Inspect disturbance bounds, feasibility and guarantee transfer across constraints.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3eeb6d236a9eb0d6b5588239f053fcc42cfb7d1a684a6f90e1e55bef4a9d4099", "source_url": "https://ieeexplore.ieee.org/abstract/document/9928337/", "source_legacy_rows": [ 177 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0178", "t": "Safe reinforcement learning via hierarchical adaptive chance-constraint safeguards", "y": 2024, "u": "https://ieeexplore.ieee.org/abstract/document/10801338/", "m": [], "k": "chen2024safe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0178", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Adaptive chance safeguards use recovery-rate surrogates and claim forward invariance with near-zero empirical violations.", "full_text_question": "Check conditions connecting the surrogate chance constraint to invariance and stochastic deployment.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "8fa92aa4a3f13ee39c2e429645daea0e30792a3f1094a76c990bcdaa8c3dfb03", "source_url": "https://ieeexplore.ieee.org/abstract/document/10801338/", "source_legacy_rows": [ 178 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0179", "t": "Safe Reinforcement Learning with Contrastive Risk Prediction", "y": 2022, "u": "https://arxiv.org/abs/2209.09648", "m": [], "k": "zhang2022safereinforcementlearningcontrastive", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0179", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Contrastive risk prediction estimates unsafe-state probabilities to guide exploration and shape rewards in simulation.", "full_text_question": "Assess calibration, prediction horizon and residual violations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "df357163277fcaa52c6d8936c53de7483763940bc40284b7e4492598b4a590d7", "source_url": "https://arxiv.org/abs/2209.09648", "source_legacy_rows": [ 179 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0180", "t": "Safe reinforcement learning with free-form natural language constraints and pre-trained language models", "y": 2024, "u": "https://arxiv.org/abs/2401.07553", "m": [ "Language planner" ], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0180", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Pretrained language models infer costs from free-form constraints for robot and grid-world safe policy learning.", "full_text_question": "Validate learned cost semantics independently; absence of ground-truth evaluation can hide violations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "52e1f419a20cb867e6b8ce3260ef0bae846400f843e5a20d712ab1a357a1396d", "source_url": "https://arxiv.org/abs/2401.07553", "source_legacy_rows": [ 180 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0181", "t": "Safe reinforcement learning with natural language constraints", "y": 2020, "u": "https://proceedings.neurips.cc/paper/2021/hash/72f67e70f6b7cdc4cc893edaddf0c4c6-Abstract.html", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0181", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A language-constraint interpreter encodes forbidden states for constrained-policy navigation.", "full_text_question": "Check temporal/spatial interpretation errors and remaining constraint violations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ea74f40d901f0b99ae0c6e43b60c498d4ffd667d2f2088e547b02817521ddd44", "source_url": "https://proceedings.neurips.cc/paper/2021/hash/72f67e70f6b7cdc4cc893edaddf0c4c6-Abstract.html", "source_legacy_rows": [ 181 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0182", "t": "Safe Robot Navigation Using Constrained Hierarchical Reinforcement Learning", "y": 2022, "u": "https://scispace.com/papers/safe-robot-navigation-using-constrained-hierarchical-m0gexw1j", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0182", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A safety layer modifies low-level hierarchical navigation actions and improves simulated safety-performance trade-offs.", "full_text_question": "Inspect the action correction, feasible region and controller hierarchy assumptions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "2b2fd54e1a8cccf37e1190c2e81d58e0f2619ba651516fe16e432e5ac078f0c9", "source_url": "https://scispace.com/papers/safe-robot-navigation-using-constrained-hierarchical-m0gexw1j", "source_legacy_rows": [ 182, 183 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0184", "t": "SAFE: Multitask Failure Detection for Vision-Language-Action Models", "y": 2025, "u": "https://arxiv.org/abs/2506.09937", "m": [ "VLA policy" ], "k": "gu2025safemultitaskfailuredetection", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0184", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "SAFE predicts VLA task failure from internal features and uses conformal prediction to balance detection accuracy and timing.", "full_text_question": "Separate task failure from harm and verify calibration assumptions and available intervention time.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e4310e11cd5836e37c3109dbab5c886ebec9849fb40c51cd75aa3bc6934caabb", "source_url": "https://arxiv.org/abs/2506.09937", "source_legacy_rows": [ 184 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0185", "t": "SafeDreamer: Safe Reinforcement Learning with World Models", "y": 2023, "u": "https://arxiv.org/abs/2307.07176", "m": [ "Learned dynamics" ], "k": "rev_safedreamer", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0186", "t": "SafeEmbodAI: a Safety Framework for Mobile Robots in Embodied AI Systems", "y": 2024, "u": "https://arxiv.org/abs/2409.01630", "m": [ "Language planner" ], "k": "zhang2024safeembodaisafetyframeworkmobile", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0186", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "SafeEmbodAI combines secure prompting, state management and output validation for mobile-robot navigation under injected commands.", "full_text_question": "Check collision metrics and whether improvement in mission score implies reduced harm.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "88b2e41f05634d4ad9e73b1e3f3e5e3173fc8bf5e2f59f192e9f3e1a210b449e", "source_url": "https://arxiv.org/abs/2409.01630", "source_legacy_rows": [ 186 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0187", "t": "Safeguarding AI Agents: Developing and Analyzing Safety Architectures", "y": 2025, "u": "https://arxiv.org/abs/2409.03793", "m": [ "Language planner" ], "k": "domkundwar2025safeguardingaiagentsdeveloping", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0187", "study_type": "method", "relevance": "unclear", "full_text_priority": "medium", "assessment": "Compares LLM filters, safety agents and hierarchical checks on agentic use cases without specifying physical robots.", "full_text_question": "Determine whether evaluated tasks are embodied before including effectiveness claims.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1c1465fa2d8d8f4a611b718240a61a1460668731ef64cc9c3b8fe273e8cb9dd1", "source_url": "https://arxiv.org/abs/2409.03793", "source_legacy_rows": [ 187 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0188", "t": "SafeMind: Benchmarking and Mitigating Safety Risks in Embodied LLM Agents", "y": 2025, "u": "https://arxiv.org/abs/2509.25885v1", "m": [ "Language planner" ], "k": "chen2025safemindbenchmarkingmitigatingsafety", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0188", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "SafeMind combines factual, causal and temporal safety benchmarks with cascaded planner-executor checks.", "full_text_question": "Inspect constraint grounding, benchmark judges and intermediate-action safety.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "2a091ec70d3569bf5da69134c05274dadf0932731b9da311b6cdbefb03eb46ce", "source_url": "https://arxiv.org/abs/2509.25885v1", "source_legacy_rows": [ 188 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0189", "t": "Safety Assurance of Artificial Intelligence-Based Systems: A Systematic Literature Review on the State of the Art and Guidelines for Future Work", "y": 2022, "u": "https://scispace.com/papers/safety-assurance-of-artificial-intelligence-based-systems-a-1x700s8s", "m": [], "k": "", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0189", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "Systematic assurance review organizes testing, envelopes, fail-safe design, explainability and lifecycle processes.", "full_text_question": "Use to strengthen prior-work positioning and trace technical assurance evidence.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e03854d0b496d2856b5816816c31b1bee7dd0c0d1a24ef29076bcbaeacfe2af9", "source_url": "https://scispace.com/papers/safety-assurance-of-artificial-intelligence-based-systems-a-1x700s8s", "source_legacy_rows": [ 189 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0190", "t": "Safety case template for frontier AI: A cyber inability argument", "y": 2024, "u": "https://scispace.com/papers/safety-case-template-for-frontier-ai-a-cyber-inability-7j8v8t76uubv", "m": [], "k": "", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0190", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "A cyber-capability safety-case template illustrates explicit claims-arguments-evidence reasoning.", "full_text_question": "Use assurance structure as context; cyber inability is not robotic physical safety.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "cfd30bd2d349d62bc743fdf545936c4c48f5a7fdeca7a3d1fe76343dc09f7084", "source_url": "https://scispace.com/papers/safety-case-template-for-frontier-ai-a-cyber-inability-7j8v8t76uubv", "source_legacy_rows": [ 190 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0191", "t": "Safety Control of Service Robots with LLMs and Embodied Knowledge Graphs", "y": 2024, "u": "https://scispace.com/papers/safety-control-of-service-robots-with-llms-and-embodied-2m2qp0ycj2", "m": [ "Language planner" ], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0191", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Service-robot prompts are checked against embodied knowledge graphs and reportedly improve safety compliance in real tasks.", "full_text_question": "Verify task details, knowledge completeness, standards claims and action-validation timing.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "004d4aaf05067e2acf10ce5b4631907714d33077b8316bb4a0e781a0339b3570", "source_url": "https://scispace.com/papers/safety-control-of-service-robots-with-llms-and-embodied-2m2qp0ycj2", "source_legacy_rows": [ 191 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0192", "t": "Safety Correction from Baseline: Towards the Risk-aware Policy in Robotics via Dual-agent Reinforcement Learning", "y": 2022, "u": "https://scispace.com/papers/safety-correction-from-baseline-towards-the-risk-aware-3s2cen9u", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0192", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Dual-agent learning corrects reward-maximizing baselines into safer locomotion and manipulation policies.", "full_text_question": "Check how safety constraints are enforced and whether unsafe baseline data collection is counted.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3d957c49493355f23d0758591f81f9883662cf937becde7d707a36a90b195d2f", "source_url": "https://scispace.com/papers/safety-correction-from-baseline-towards-the-risk-aware-3s2cen9u", "source_legacy_rows": [ 192, 193 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0194", "t": "Safety Guided Policy Optimization", "y": 2022, "u": "https://scispace.com/papers/safety-guided-policy-optimization-2cpezlpz", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0194", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Safety Guided Policy Optimization predicts near-future constraints, corrects actions and trains policies to imitate corrections.", "full_text_question": "Inspect prediction horizon, error bounds and physical validation of corrected actions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "c665393ae5b3507864f0306a56f3c5e72920a11ef98ac69519a0f041a42c2c89", "source_url": "https://scispace.com/papers/safety-guided-policy-optimization-2cpezlpz", "source_legacy_rows": [ 194, 195 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0196", "t": "Safety gymnasium: A unified safe reinforcement learning benchmark", "y": 2023, "u": "https://proceedings.neurips.cc/paper_files/paper/2023/hash/3c557a3d6a48cc99444f85e924c66753-Abstract-Datasets_and_Benchmarks.html", "m": [], "k": "ji2023safety", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0196", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "Safety-Gymnasium provides visual/vector and single/multi-agent safety tasks with standardized algorithm implementations.", "full_text_question": "Extract exact cost definitions, budgets and evaluation protocols.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e628f82217f1acb51929b7825d33042f9426e6712b41ff4a6c16838c15b687b9", "source_url": "https://proceedings.neurips.cc/paper_files/paper/2023/hash/3c557a3d6a48cc99444f85e924c66753-Abstract-Datasets_and_Benchmarks.html", "source_legacy_rows": [ 196 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0197", "t": "Safety of Embodied Navigation: A Survey", "y": 2025, "u": "https://arxiv.org/abs/2508.05855", "m": [ "Language planner" ], "k": "rev_navsurvey", "role": "secondary", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0198", "t": "Safety-Driven Deep Reinforcement Learning Framework for Cobots: A Sim2Real Approach", "y": 2024, "u": "https://scispace.com/papers/safety-driven-deep-reinforcement-learning-framework-for-2wmiaud86c", "m": [], "k": "", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0198", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Safety-constrained sim-to-real cobot learning reports velocity limits, obstacle avoidance and testbed evaluation.", "full_text_question": "Scrutinize standard-conformance and integrity-level claims separately from observed zero violations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ff615e804083bec0a75dbaa1d93ccbc66d7a71abcec3ae05cc9b94932bb349b6", "source_url": "https://scispace.com/papers/safety-driven-deep-reinforcement-learning-framework-for-2wmiaud86c", "source_legacy_rows": [ 198 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0199", "t": "SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning", "y": 2025, "u": "https://arxiv.org/abs/2503.03480", "m": [ "VLA policy" ], "k": "rev_safevla", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0200", "t": "Scalable, Training-Free Visual Language Robotics: a modular multi-model framework for consumer-grade GPUs", "y": 2025, "u": "https://ieeexplore.ieee.org/document/10870971", "m": [ "VLA policy", "Language planner" ], "k": "samson2025scalable", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0200", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "SVLR composes lightweight models for training-free pick-and-place and explicitly acknowledges limited evaluation.", "full_text_question": "No safety mechanism or broad generalization is established by initial demonstrations.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e922c28bb598feda760317e11f30a041d298541f859094d7570db9b682206672", "source_url": "https://ieeexplore.ieee.org/document/10870971", "source_legacy_rows": [ 200 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0201", "t": "Securing Embodied AI: Addressing Cybersecurity Challenges in Physical Systems", "y": 2025, "u": "https://link.springer.com/chapter/10.1007/978-3-031-68256-8_21", "m": [], "k": "khalifa2025securing", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0201", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "A chapter surveys cyber threats, data integrity and physical attacks affecting embodied systems.", "full_text_question": "Trace concrete mitigations to technical evidence rather than broad security assurances.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "000230c9365b334e9b9a98c64685aa1312c2d9ff791277c0e3b1a76175faff85", "source_url": "https://link.springer.com/chapter/10.1007/978-3-031-68256-8_21", "source_legacy_rows": [ 201 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0202", "t": "Security Considerations in AI-Robotics: A Survey of Current Methods, Challenges, and Opportunities", "y": 2024, "u": "https://arxiv.org/abs/2310.08565", "m": [], "k": "neupane2024securityconsiderationsairoboticssurvey", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0202", "study_type": "review", "relevance": "direct", "full_text_priority": "medium", "assessment": "Surveys robotic attack surfaces, defenses and human-interaction security across perception, planning and control.", "full_text_question": "Separate physical-safety mechanisms from ethical/legal discussions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f9414f80340c3502cef77a32dbd6c1c36a0e34d460273042ee6d36704b1d5cb3", "source_url": "https://arxiv.org/abs/2310.08565", "source_legacy_rows": [ 202 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0203", "t": "Selective Visual Representations Improve Convergence and Generalization for Embodied AI", "y": 2024, "u": "https://arxiv.org/abs/2311.04193v2", "m": [], "k": "eftekhar2024selectivevisualrepresentationsimprove", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0203", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "A task-conditioned visual bottleneck improves navigation and displacement learning by discarding irrelevant information.", "full_text_question": "Check whether discarded features can still determine harm, a direct representation-risk concern.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "53a70f4f21d30260d3a515670c09568b494fa79458cbb1a144f1f860190feede", "source_url": "https://arxiv.org/abs/2311.04193v2", "source_legacy_rows": [ 203 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0204", "t": "Semantics-aware Adaptive Knowledge Distillation for Sensor-to-Vision Action Recognition", "y": 2020, "u": "https://arxiv.org/abs/2009.00210v5", "m": [], "k": "", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0204", "study_type": "method", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Sensor-to-video distillation improves human action recognition on datasets without an embodied safety controller.", "full_text_question": "Exclude from safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "879b228a786cea0e805e934fb2448957a9a708f9e19f39e20daec784411ac882", "source_url": "https://arxiv.org/abs/2009.00210v5", "source_legacy_rows": [ 204 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0205", "t": "Smooth Exploration for Robotic Reinforcement Learning", "y": 2020, "u": "https://arxiv.org/abs/2005.05719v2", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0205", "study_type": "method", "relevance": "direct", "full_text_priority": "medium", "assessment": "Generalized state-dependent exploration reduces jerky motions and supports direct learning on three physical robots.", "full_text_question": "Measure mechanical damage or safety violations separately from smoothness and reward.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "57e0441d57f0ee7b199a405edb62f92fbcfe84e31f848d46a94edd4f8e10b9ee", "source_url": "https://arxiv.org/abs/2005.05719v2", "source_legacy_rows": [ 205 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0206", "t": "State-of-the-art in Robot Learning for Multi-Robot Collaboration: A Comprehensive Survey", "y": 2024, "u": "https://arxiv.org/abs/2408.11822v1", "m": [], "k": "", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0206", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Surveys learning for multi-robot cooperation and applications.", "full_text_question": "Check for explicit inter-robot safety evidence before using it in the technical comparison.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e83933b307527eb99a6a67a48f557ccf3584e27bb7a30caf8deaca372a2bb2c9", "source_url": "https://arxiv.org/abs/2408.11822v1", "source_legacy_rows": [ 206 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0207", "t": "Subtle risks, critical failures: A framework for diagnosing physical safety of llms for embodied decision making", "y": 2025, "u": "https://aclanthology.org/2025.emnlp-main.1305/", "m": [ "Language planner" ], "k": "son2025subtle", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0207", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "SAFEL diagnoses command refusal and plan safety through PDDL-grounded scenarios and decomposed reasoning stages.", "full_text_question": "Validate generated scenarios, symbolic transition fidelity and correspondence to physical execution.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "26fa348965982353f3c933431b02e5be79575995906f0bffa7a42d8210061f30", "source_url": "https://aclanthology.org/2025.emnlp-main.1305/", "source_legacy_rows": [ 207 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0208", "t": "Survey of Vision-Language-Action Models for Embodied Manipulation", "y": 2025, "u": "https://arxiv.org/abs/2508.15201", "m": [ "VLA policy" ], "k": "li2025surveyvisionlanguageactionmodelsembodied", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0208", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Surveys VLA manipulation architectures, datasets, training and evaluation.", "full_text_question": "Use for background taxonomy and distinguish deployment challenges from evaluated safeguards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "acfa27edde232c4980399b15abb452b8fc347546cf655edc5703dc6c30d95127", "source_url": "https://arxiv.org/abs/2508.15201", "source_legacy_rows": [ 208 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0209", "t": "The Essential Role of Causality in Foundation World Models for Embodied AI", "y": 2024, "u": "https://arxiv.org/abs/2402.06665", "m": [ "Learned dynamics" ], "k": "gupta2024essentialrolecausalityfoundation", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0209", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "Argues causal world models are necessary for predicting physical interaction outcomes.", "full_text_question": "Identify testable causal criteria without treating the viewpoint as a validated safety result.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3f7cba5d526a5800720a8bbabec758b1b6103453a94dc1c1cbe3dbf422fa454e", "source_url": "https://arxiv.org/abs/2402.06665", "source_legacy_rows": [ 209 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0210", "t": "The Ethics of Advanced AI Assistants", "y": 2024, "u": "https://arxiv.org/abs/2404.16244", "m": [], "k": "gabriel2024ethicsadvanced", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0210", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "Advanced-assistant ethics analysis addresses alignment, relationships and societal effects without a specific robotic safeguard.", "full_text_question": "Use only for broader risk framing.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "493eaf3651118d434d3d59216a8f994b8de4608af9023c4faeb73ed70b5a522e", "source_url": "https://arxiv.org/abs/2404.16244", "source_legacy_rows": [ 210 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0211", "t": "The Evolution of First Person Vision Methods: A Survey", "y": 2014, "u": "https://arxiv.org/abs/1409.1484v3", "m": [], "k": "", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0211", "study_type": "review", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Reviews first-person video-analysis methods for perception and activity recognition rather than embodied safety control.", "full_text_question": "Exclude from safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "fbc9655d16b21415bb93371b31f4e039f8dfa28702a1c60408e004cab84f3369", "source_url": "https://arxiv.org/abs/1409.1484v3", "source_legacy_rows": [ 211 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0212", "t": "The Physical Basis of Prediction: World Model Formation in Neural Organoids via an LLM-Generated Curriculum", "y": 2025, "u": "https://arxiv.org/abs/2509.04633v3", "m": [ "Learned dynamics", "Language planner" ], "k": "", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0212", "study_type": "position", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Proposes virtual-task curricula and evaluation for neural organoids, not a robotic safety method.", "full_text_question": "Do not infer demonstrated learning or safety from a proposed biological-agent framework.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f918bdd32974002bbad2065a339db771e720a8f6f96b5274903e90227238b73e", "source_url": "https://arxiv.org/abs/2509.04633v3", "source_legacy_rows": [ 212 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0213", "t": "The Road to Know-Where: An Object-and-Room Informed Sequential BERT for Indoor Vision-Language Navigation", "y": 2021, "u": "https://arxiv.org/abs/2104.04167v2", "m": [], "k": "", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0213", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "Object-and-room-informed BERT improves instruction-following navigation through spatial-semantic encoding.", "full_text_question": "No hazard-specific evaluation is described.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "bec931fc7874586b124d9a0620ad4c951a11e677ee031cdd61106349082e71f9", "source_url": "https://arxiv.org/abs/2104.04167v2", "source_legacy_rows": [ 213 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0214", "t": "The Safety Challenge of World Models for Embodied AI Agents: A Review", "y": 2025, "u": "https://arxiv.org/abs/2510.05865", "m": [ "Learned dynamics" ], "k": "rev_wmsurvey", "role": "secondary", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0215", "t": "ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning", "y": 2025, "u": "https://arxiv.org/pdf/2507.16815", "m": [ "VLA policy", "Language planner" ], "k": "huang2025thinkactvisionlanguageactionreasoningreinforced", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0215", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "ThinkAct uses reinforced visual latent plans for long-horizon robot action and self-correction.", "full_text_question": "Check whether correction handles hazards or only task failures.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "60c2a44d4ad17a7012e6d4ed6ca9202e91e6ebc99a4b283d39e230a5f2a8242c", "source_url": "https://arxiv.org/pdf/2507.16815", "source_legacy_rows": [ 215 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0216", "t": "Towards a Robotic Intrusion Prevention System: Combining Security and Safety in Cognitive Social Robots", "y": 2025, "u": "https://arxiv.org/abs/2407.06669", "m": [], "k": "martin2025safety", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0216", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "RIPS detects communication-level intrusions and switches robot functionality modes to mitigate cyber-physical threats.", "full_text_question": "Inspect mode-transition safety, detection latency and physical threat experiments.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1dcbf044edb9087640a441cef8d5a8403a40b9d8f04ec1342469e495f32141da", "source_url": "https://arxiv.org/abs/2407.06669", "source_legacy_rows": [ 216 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0217", "t": "Towards Embodied Agentic AI: Review and Classification of LLM- and VLM-Driven Robot Autonomy and Interaction", "y": 2025, "u": "https://arxiv.org/abs/2508.05294", "m": [ "VLA policy", "Language planner" ], "k": "salimpour2025embodiedagenticaireview", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0217", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Classifies model-driven robot agents, tooling and ROS integrations, including non-peer-reviewed community projects.", "full_text_question": "Track evidence provenance and distinguish engineering examples from validated safety methods.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "d68a1ae79ce2f72e4c4e9f67322f1aeaf35d45302d33e1a6453ff49dd223f04e", "source_url": "https://arxiv.org/abs/2508.05294", "source_legacy_rows": [ 217 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0218", "t": "Towards provable probabilistic safety for scalable embodied AI systems", "y": 2025, "u": "https://arxiv.org/abs/2506.05171v2", "m": [], "k": "he2025provableprobabilisticsafetyscalable", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0218", "study_type": "position", "relevance": "direct", "full_text_priority": "high", "assessment": "A probabilistic-safety perspective proposes scalable system-level risk boundaries and a research roadmap.", "full_text_question": "Check operational definitions and do not treat proposed probabilistic assurance as an implemented result.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "2d3c5660d4cd6771c47681c9305667cc45239c89eaba9be7650af5aadfef6439", "source_url": "https://arxiv.org/abs/2506.05171v2", "source_legacy_rows": [ 218 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0219", "t": "Towards Robust and Secure Embodied AI: A Survey on Vulnerabilities and Attacks", "y": 2025, "u": "https://arxiv.org/abs/2502.13175", "m": [ "Language planner" ], "k": "rev_securitysurvey", "role": "secondary", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0220", "t": "Towards Safe and Trustworthy Embodied AI: Foundations, Status, and Prospects", "y": 2025, "u": "https://openreview.net/pdf?id=Eu6Yt21Alv", "m": [], "k": "tan2025safetrustworthyeai", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0220", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "A safe/trustworthy embodied-AI review organizes risks across operational stages and advocates closed-loop systems analysis.", "full_text_question": "Compare overlap with this manuscript's systems-level contribution and prior-review positioning.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "29bcad90cc02adc3d672dad414adcf654a3d69177db6ae875d88719f386e599d", "source_url": "https://openreview.net/pdf?id=Eu6Yt21Alv", "source_legacy_rows": [ 220 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0221", "t": "Towards safe reinforcement learning with a safety editor policy", "y": 2022, "u": "https://proceedings.neurips.cc/paper_files/paper/2022/hash/11afefdd848d1bc9ac9f1604d9f45817-Abstract-Conference.html", "m": [], "k": "yu2022safety", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0221", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "SEditor learns a separate action-editing policy with very low but nonzero violations in safety and racing tasks.", "full_text_question": "Check constraint budgets, residual risk and safety during learning the editor.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "cb6436d411baaaefccb784f4a45f0031d12a9109e3ed6a07e0646e2278dce865", "source_url": "https://proceedings.neurips.cc/paper_files/paper/2022/hash/11afefdd848d1bc9ac9f1604d9f45817-Abstract-Conference.html", "source_legacy_rows": [ 221 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0222", "t": "Towards Testing and Evaluating Vision-Language-Action Models for Robotic Manipulation: An Empirical Study", "y": 2025, "u": "https://arxiv.org/abs/2409.12894", "m": [ "VLA policy" ], "k": "wang2025vlatest", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0222", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "VLATest generates manipulation scenes to expose VLA sensitivity to viewpoints, lighting and novel objects.", "full_text_question": "Separate robustness failures from harm and verify coverage of safety-specific outcomes.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f11a7ff36e1263fd68a0cead2e6c059b39fd6efef73b5ab26739eb9d28efcf3a", "source_url": "https://arxiv.org/abs/2409.12894", "source_legacy_rows": [ 222, 231 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0223", "t": "Towards Vision-Language Mechanistic Interpretability: A Causal Tracing Tool for BLIP", "y": 2023, "u": "https://arxiv.org/abs/2308.14179v1", "m": [ "Language planner" ], "k": "palit2023visionlanguagemechanisticinterpretabilitycausal", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0223", "study_type": "method", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Causal tracing of BLIP is evaluated on visual question answering, not embodied policy execution.", "full_text_question": "Use only if a general interpretability precursor is needed.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "b09a03a31bc435039f22e0445eac0d20821c678a53b08216ddf00921599611d5", "source_url": "https://arxiv.org/abs/2308.14179v1", "source_legacy_rows": [ 223 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0224", "t": "TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies", "y": 2025, "u": "https://arxiv.org/abs/2412.10345", "m": [ "VLA policy" ], "k": "zheng2025tracevlavisualtraceprompting", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0224", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "TraceVLA encodes past state-action trajectories visually to improve spatial-temporal action prediction.", "full_text_question": "Check whether trace-based awareness is evaluated for collision or other hazards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e301bbcd9d6ab033ad5f181757352d768e3e19a32ee292e4f99b2541dc05652a", "source_url": "https://arxiv.org/abs/2412.10345", "source_legacy_rows": [ 224 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0225", "t": "TriVLA: A Triple-System-Based Unified Vision-Language-Action Model for General Robot Control", "y": 2025, "u": "https://arxiv.org/abs/2507.01424", "m": [ "VLA policy", "Language planner" ], "k": "liu2025trivlatriplesystembasedunifiedvisionlanguageaction", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0225", "study_type": "architecture", "relevance": "context", "full_text_priority": "medium", "assessment": "TriVLA integrates predicted dynamics with vision-language reasoning and real-time action generation.", "full_text_question": "Determine whether prediction is only policy conditioning or an explicit safety assessment.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "12c0fb06f3588615c3bcd05d0f16d1366de1c7257b784df095dad2874e1a4637", "source_url": "https://arxiv.org/abs/2507.01424", "source_legacy_rows": [ 225 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0226", "t": "Trustworthy, Responsible, and Safe AI: A Comprehensive Architectural Framework for AI Safety with Challenges and Mitigations", "y": 2025, "u": "https://arxiv.org/abs/2408.12935", "m": [ "Language planner" ], "k": "chen2025trustworthyresponsiblesafeai", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0226", "study_type": "review", "relevance": "context", "full_text_priority": "low", "assessment": "Broad AI-safety architecture survey separates trustworthy, responsible and safe AI perspectives.", "full_text_question": "Use for terminology rather than embodied-controller effectiveness.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "92bb21e0d593d83c5541d1a9667503faeebea2ce01098ed0585719bff1d60c9e", "source_url": "https://arxiv.org/abs/2408.12935", "source_legacy_rows": [ 226 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0227", "t": "Understanding World or Predicting Future? A Comprehensive Survey of World Models", "y": 2025, "u": "https://dl.acm.org/doi/10.1145/3746449", "m": [ "Learned dynamics", "Language planner" ], "k": "ding2025worldmodels", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0227", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Categorizes world models into present-world understanding and future-state prediction across applications.", "full_text_question": "Use to clarify representation versus prediction without implying safety guarantees.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "353d94c699939b5550f91cc5d69a0a1accc31372a59ee3c8456a2f22cafbd5ff", "source_url": "https://dl.acm.org/doi/10.1145/3746449", "source_legacy_rows": [ 227 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0228", "t": "Updating Robot Safety Representations Online from Natural Language Feedback", "y": 2024, "u": "https://arxiv.org/abs/2409.14580", "m": [ "Language planner" ], "k": "santos2024updatingrobotsafetyrepresentations", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0228", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Language feedback updates VLM-inferred constraints and warm-started Hamilton-Jacobi reachability control online.", "full_text_question": "Verify semantic grounding, update latency and safety during controller recomputation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "43db74ac58e66f774c200465b53c3e4a4471a37806f1273e50aeb5060b874bc0", "source_url": "https://arxiv.org/abs/2409.14580", "source_legacy_rows": [ 228 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0229", "t": "Video Event Reasoning and Prediction by Fusing World Knowledge from LLMs with Vision Foundation Models", "y": 2025, "u": "https://arxiv.org/abs/2507.05822", "m": [ "Language planner" ], "k": "dubois2025videoeventreasoningprediction", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0229", "study_type": "architecture", "relevance": "context", "full_text_priority": "low", "assessment": "Vision-language fusion supports video-event reasoning and prediction; robotics is a prospective application.", "full_text_question": "Confirm any actual robot evaluation before treating it as embodied evidence.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7181013e25b967d0f76a9fa8c6aab906309c53b628bdf3203e4c31e0613d8ecb", "source_url": "https://arxiv.org/abs/2507.05822", "source_legacy_rows": [ 229 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0232", "t": "Whole-Body Constrained Learning for Legged Locomotion via Hierarchical Optimization", "y": 2025, "u": "https://ieeexplore.ieee.org/document/11027695", "m": [], "k": "wang2025legged", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0232", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Hierarchical whole-body optimization adds hard and soft constraints to legged RL with outdoor hexapod deployment.", "full_text_question": "Separate hard safety limits from soft objectives and inspect contact/friction assumptions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "562dd7c13eec4be9e96f64105c5fee2afedf7d13f2a32e5e4d265f51dc442e66", "source_url": "https://ieeexplore.ieee.org/document/11027695", "source_legacy_rows": [ 232 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0233", "t": "World models and predictive coding for cognitive and developmental robotics: frontiers and challenges", "y": 2023, "u": "https://www.tandfonline.com/doi/full/10.1080/01691864.2023.2225232", "m": [ "Learned dynamics" ], "k": "taniguchi2023world", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0233", "study_type": "review", "relevance": "context", "full_text_priority": "medium", "assessment": "Connects world models, predictive coding and active inference in developmental robotics.", "full_text_question": "Use for conceptual history and avoid conflating prediction with a safety certificate.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "eccbd90f4ace6d8863c62552a122b7396d35e0dc22fad92247329e43bfa6265d", "source_url": "https://www.tandfonline.com/doi/full/10.1080/01691864.2023.2225232", "source_legacy_rows": [ 233 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0234", "t": "World Models: The Safety Perspective", "y": 2024, "u": "https://arxiv.org/pdf/2411.07690", "m": [ "Learned dynamics", "Language planner" ], "k": "zeng2024world", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0234", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "Reviews world models explicitly through safety and trustworthiness concerns.", "full_text_question": "Compare identified challenges and coverage with the manuscript's novelty claims.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "b1857353c2f05259b5ebad408204f07543317453ead7f79260deeb659b0a28cd", "source_url": "https://arxiv.org/pdf/2411.07690", "source_legacy_rows": [ 234 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0235", "t": "World-in-World: World Models in a Closed-Loop World", "y": 2025, "u": "https://arxiv.org/abs/2510.18135v1", "m": [ "Learned dynamics" ], "k": "zhang2025worldinworldworldmodelsclosedloop", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0235", "study_type": "benchmark", "relevance": "context", "full_text_priority": "high", "assessment": "World-in-World evaluates world models in closed-loop tasks and finds visual quality insufficient for task success.", "full_text_question": "Check whether any outcomes measure harm; task utility remains distinct from safety.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7c4da7eb073ae01960b06b51ead882afe92cf73759757a38fa6095d5c122b6ac", "source_url": "https://arxiv.org/abs/2510.18135v1", "source_legacy_rows": [ 235 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0236", "t": "X-Risk Analysis for AI Research", "y": 2022, "u": "https://arxiv.org/abs/2206.05862", "m": [], "k": "hendrycks2022xriskanalysisairesearch", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0236", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "A general AI existential-risk guide draws on hazard analysis and systems-safety concepts.", "full_text_question": "Use only for general assurance framing.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "097b4ca86b716c3fa3b7a05eff7f2115c74c217a5a00fed0ea3ada786f8f7697", "source_url": "https://arxiv.org/abs/2206.05862", "source_legacy_rows": [ 236 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0237", "t": "Zero-shot Safety Prediction for Autonomous Robots with Foundation World Models", "y": 2024, "u": "https://arxiv.org/abs/2404.00462", "m": [ "Learned dynamics", "Language planner" ], "k": "rev_zeroshot", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0238", "t": "VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model", "y": 2026, "u": "https://arxiv.org/abs/2602.10098", "m": [ "Learned dynamics", "VLA policy" ], "k": "rev_vlajepa", "role": "architecture_background", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0239", "t": "VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer", "y": 2025, "u": "https://arxiv.org/abs/2512.11891", "m": [ "VLA policy" ], "k": "rev_aegis", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0240", "t": "World Action Models are Zero-shot Policies", "y": 2026, "u": "https://arxiv.org/abs/2602.15922", "m": [ "VLA policy", "Joint world-action" ], "k": "rev_dreamzero", "role": "architecture_background", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0241", "t": "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks", "y": 2025, "u": "https://arxiv.org/abs/2510.04374", "m": [ "Learned dynamics" ], "k": "patwardhan2025gdpval", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0241", "study_type": "benchmark", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "GDPval measures economically valuable professional deliverables rather than embodied-controller safety.", "full_text_question": "Exclude from embodied safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "5a9c3a4ec7c1383929f5aff95716b9b34a4ac43c7ec0f021555ce92f9b1ecdb1", "source_url": "https://arxiv.org/abs/2510.04374", "source_legacy_rows": [ 241 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0242", "t": "Embodied AI Agents: Modeling the World", "y": 2025, "u": "https://arxiv.org/abs/2506.22355", "m": [ "Learned dynamics" ], "k": "fung2025embodiedaiagents", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0242", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "Proposes physical and user mental-world models for multimodal embodied agents and collaboration.", "full_text_question": "Distinguish conceptual modeling goals from evaluated predictive safeguards.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "21438c1dc77ff1f1bd9eecc8285fd6480db8420e4c4e47ed614e528a70c3a334", "source_url": "https://arxiv.org/abs/2506.22355", "source_legacy_rows": [ 242 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0243", "t": "AI2-THOR: An Interactive 3D Environment for Visual AI", "y": 2017, "u": "https://arxiv.org/abs/1712.05474", "m": [], "k": "kolve2017ai2thor", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0243", "study_type": "benchmark", "relevance": "context", "full_text_priority": "medium", "assessment": "AI2-THOR supplies interactive indoor simulation infrastructure for embodied learning and planning.", "full_text_question": "Check which later safety benchmarks add hazard semantics beyond simulator affordances.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e59e289f2e3ab6079ead3fac9760d339c9119034c47514750990ef1186d97252", "source_url": "https://arxiv.org/abs/1712.05474", "source_legacy_rows": [ 243 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0245", "t": "Don't Trade Off Safety: Diffusion Regularization for Constrained Offline RL", "y": 2025, "u": "https://arxiv.org/abs/2502.12391", "m": [], "k": "guo2025dont", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0245", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "DRCORL combines diffusion-based offline behavior modeling and safety adaptation to meet empirical cost limits.", "full_text_question": "Check dataset support, cost constraints and safety under deployment distribution shift.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1d2cbbf0dfde4a6944ded9b48c03c420f3b0e42aa0e6a503af6e5db18959bc1d", "source_url": "https://arxiv.org/abs/2502.12391", "source_legacy_rows": [ 245 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0246", "t": "HASARD: A Benchmark for Vision-Based Safe Reinforcement Learning in Embodied Agents", "y": 2025, "u": "https://arxiv.org/abs/2503.08241", "m": [], "k": "tomilin2025hasard", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0246", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "HASARD uses egocentric Doom tasks with reward-cost trade-offs and multiple difficulty levels for safe RL.", "full_text_question": "State the game-based abstraction and limits of transfer to physical robot safety.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "6033db8991221ba22944cf8bfcc2205c361a43583fbc2bd67536724b25de3c29", "source_url": "https://arxiv.org/abs/2503.08241", "source_legacy_rows": [ 246 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0247", "t": "SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents", "y": 2025, "u": "https://arxiv.org/abs/2510.12985", "m": [], "k": "authors2026sentinel", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0247", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "SENTINEL uses temporal logic to verify semantic interpretation, plans and trajectory trees in embodied simulators.", "full_text_question": "Check specification grounding, transition semantics and finite-trace versus deployed-system guarantees.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "8a300500d60849b065f118a78c43ced7ba5060a7a3ffdd21e39fbfbb2038dcc0", "source_url": "https://arxiv.org/abs/2510.12985", "source_legacy_rows": [ 247 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0248", "t": "ABNet: Attention BarrierNet for Safe and Scalable Robot Learning", "y": 2024, "u": "https://arxiv.org/abs/2406.13025", "m": [], "k": "xie2025abnet", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0248", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Attention BarrierNet combines multiple barrier heads while claiming formal safety and improved robustness across robot tasks.", "full_text_question": "Inspect composition conditions, perception assumptions and scalability of certification.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "35323e8062f12ecee5ca57b286ac4f4f609022a422fd84f0ae336cc26535c032", "source_url": "https://arxiv.org/abs/2406.13025", "source_legacy_rows": [ 248 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0249", "t": "Learning Control Barrier Functions and their application in Reinforcement Learning: A Survey", "y": 2024, "u": "https://arxiv.org/abs/2404.16879", "m": [], "k": "guerrier2024learning", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0249", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "Surveys learning CBFs and their integration into safe RL for robotics.", "full_text_question": "Map learned-certificate verification and training-data assumptions to assurance claims.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "81159388028311c30717cee96ccd9ae44a03a4f98300f59367dc554053b8f7ea", "source_url": "https://arxiv.org/abs/2404.16879", "source_legacy_rows": [ 249 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0250", "t": "Scalable Verification of Neural Control Barrier Functions Using Linear Bound Propagation", "y": 2025, "u": "https://arxiv.org/abs/2511.06341", "m": [ "Analytical dynamics" ], "k": "vertovec2025scalable", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0250", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Linear bound propagation verifies neural-CBF conditions and gradients with adaptive refinement for control-affine systems.", "full_text_question": "Inspect domain coverage, numerical soundness and conservatism of the certificate.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "527bcd243728a1aa02ff129805ee530c6385e8265fc98f23e3071ae40d7e8f06", "source_url": "https://arxiv.org/abs/2511.06341", "source_legacy_rows": [ 250 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0251", "t": "Safe Reinforcement Learning of Robot Trajectories in the Presence of Moving Obstacles", "y": 2024, "u": "https://arxiv.org/abs/2411.05784", "m": [], "k": "arxiv2024safe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0251", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A learned evasive backup supports collision-risk screening under moving obstacles using simulation or learned estimators.", "full_text_question": "Compare estimator error, backup feasibility and response deadlines, especially where no permanently safe state exists.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "ae7bd7f1fa19e98373a43c89f9b228185a71633d9d5a708c4fadbaea043e4b53", "source_url": "https://arxiv.org/abs/2411.05784", "source_legacy_rows": [ 251 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0252", "t": "Learning Fast, Tool aware Collision Avoidance for Collaborative Robots", "y": 2025, "u": "https://arxiv.org/abs/2508.20457", "m": [], "k": "authors2024learning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0252", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Tool-aware perception predicts collisions under occlusion and constrained RL generates fast avoidance on physical robots.", "full_text_question": "Check occlusion uncertainty, end-to-end latency and tool-geometry generalization.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1916fda62621aa1b35d50537a392a8a4adb40480bdad1325711269d9021fecc3", "source_url": "https://arxiv.org/abs/2508.20457", "source_legacy_rows": [ 252 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0253", "t": "Probabilistic Shielding for Safe Reinforcement Learning", "y": 2025, "u": "https://arxiv.org/abs/2503.07671", "m": [], "k": "rev_shield", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0254", "t": "Safe Planner: Empowering Safety Awareness in Large Pre-Trained Models for Robot Task Planning", "y": 2025, "u": "https://arxiv.org/abs/2411.06920", "m": [ "Language planner" ], "k": "li2025safe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0254", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Safe Planner guides a pretrained task planner with a simulation-trained safety predictor transferred to physical robots.", "full_text_question": "Verify safety-prediction calibration and how risk scores constrain execution.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "2254f6b7e01bf705423b4603dd31068c5bb8f86508520e4359606aeb873e867b", "source_url": "https://arxiv.org/abs/2411.06920", "source_legacy_rows": [ 254 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0255", "t": "SafeAgentBench: A Benchmark for Safe Task Planning of Embodied LLM Agents", "y": 2024, "u": "https://arxiv.org/abs/2412.13178", "m": [ "Language planner" ], "k": "rev_safeagentbench", "role": "benchmark", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0256", "t": "RoboSafe: Safeguarding Embodied Agents via Executable Safety Logic", "y": 2025, "u": "https://arxiv.org/abs/2512.21220", "m": [ "Language planner" ], "k": "paper2025robosafe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0256", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "RoboSafe uses backward temporal checks and forward risk prediction to generate executable safety predicates and trigger replanning.", "full_text_question": "Audit generated predicate correctness, memory updates and remaining hazardous actions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "99f2cfbc352ea34584478f7663b7d450afa41f70c88f0a1d4adb6fa3be16735b", "source_url": "https://arxiv.org/abs/2512.21220", "source_legacy_rows": [ 256 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0257", "t": "Gameplay Filters: Robust Zero-Shot Safety through Adversarial Imagination", "y": 2024, "u": "https://arxiv.org/abs/2405.00846", "m": [], "k": "rev_gameplay", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0258", "t": "FREA: Feasibility-Guided Generation of Safety-Critical Scenarios with Reasonable Adversariality", "y": 2024, "u": "https://arxiv.org/abs/2406.02983", "m": [], "k": "chen2024frea", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0258", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "FREA generates adversarial driving scenarios while constraining them to an estimated feasible region.", "full_text_question": "Check feasibility-estimation error and distinguish avoidable challenge from unavoidable collision.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "119392345dc53ab9224883300d61ba4014b4f79b410d7b399971d7609c3b1926", "source_url": "https://arxiv.org/abs/2406.02983", "source_legacy_rows": [ 258 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0259", "t": "Learning Performance-Oriented Control Barrier Functions Under Complex Safety Constraints and Limited Actuation", "y": 2024, "u": "https://arxiv.org/abs/2401.05629", "m": [], "k": "manda2024learning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0259", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Physics-informed neural-CBF learning approximates invariant sets under composed constraints and actuation limits.", "full_text_question": "Verify candidate certification rather than relying only on PDE training loss.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "2050447581bf7d30af5a3a44c33de12e18408e89dd7c24c808979b605522770b", "source_url": "https://arxiv.org/abs/2401.05629", "source_legacy_rows": [ 259 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0260", "t": "Semantically Safe Robot Manipulation: From Semantic Scene Understanding to Motion Safeguards", "y": 2024, "u": "https://arxiv.org/abs/2410.15185", "m": [ "Language planner" ], "k": "brunke2024semantically", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0260", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Semantic scene understanding and LLM-derived conditions are mapped to CBF safeguards in physical kitchen manipulation.", "full_text_question": "Check semantic-to-geometric translation, perception uncertainty and limits of the barrier certificate.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "5f0e6cafb459892fef3efaca121a4d8175edf191eb0803a197a7dc748a248bf7", "source_url": "https://arxiv.org/abs/2410.15185", "source_legacy_rows": [ 260, 262 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0261", "t": "Beyond Model Jailbreak: Systematic Dissection of the \"Ten DeadlySins\" in Embodied Intelligence", "y": 2025, "u": "https://arxiv.org/abs/2512.06387", "m": [ "Language planner" ], "k": "huang2025beyond", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0261", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "A physical robot security audit finds cross-layer vulnerabilities enabling unauthorized control beyond model jailbreaks.", "full_text_question": "Assess system-level safety implications and disclosure provenance without reproducing exploitation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "e994b7fa9598f1bb0f5de97b73016c20bedea45a037e276065f72a784eab8c93", "source_url": "https://arxiv.org/abs/2512.06387", "source_legacy_rows": [ 261 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0263", "t": "Adversarial Attacks on Robotic Vision Language Action Models", "y": 2025, "u": "https://arxiv.org/abs/2506.03350", "m": [ "VLA policy", "Language planner" ], "k": "jones2025adversarial", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0263", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Textual adversarial inputs manipulate VLA actions across the reachable action space and can persist over rollouts.", "full_text_question": "Inspect threat assumptions and distinguish action-control authority from measured physical harm.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "df911e6c369461c1d4f7fcb2498e1fd1c1ebbedecd4125653075239e26011911", "source_url": "https://arxiv.org/abs/2506.03350", "source_legacy_rows": [ 263 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0264", "t": "Model-agnostic Adversarial Attack and Defense for Vision-Language-Action Models", "y": 2025, "u": "https://arxiv.org/abs/2510.13237", "m": [ "VLA policy" ], "k": "xu2025modelagnostic", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0264", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Model-agnostic visual patches disrupt VLA embeddings and adversarial fine-tuning mitigates simulated task failures.", "full_text_question": "Check transfer across attacks and whether defenses reduce hazardous behavior rather than only failure rates.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "69035e71199091f5d0e103e817448058b8e605ed517fc0b1a09b8f5640daf7dc", "source_url": "https://arxiv.org/abs/2510.13237", "source_legacy_rows": [ 264 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0265", "t": "Visual Adversarial Attack on Vision-Language Models for Autonomous Driving", "y": 2024, "u": "https://arxiv.org/abs/2411.18275", "m": [ "Language planner" ], "k": "zhang2024visual", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0265", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "ADvLM attacks driving VLMs across semantically equivalent prompts and temporally linked views.", "full_text_question": "Verify closed-loop driving consequences and the scope of claimed real-world attack evidence.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "88cbd3d580933dae325127baf8b1c5fbdf00a0a3b0ccb64ad01f906818d8f0bc", "source_url": "https://arxiv.org/abs/2411.18275", "source_legacy_rows": [ 265 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0266", "t": "Transferable Adversarial Attacks on Black-Box Vision-Language Models", "y": 2025, "u": "https://arxiv.org/abs/2505.01050", "m": [ "Language planner" ], "k": "hu2025transferable", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0266", "study_type": "method", "relevance": "context", "full_text_priority": "medium", "assessment": "Transferable VLM perturbations alter visual interpretations across recognition, VQA and captioning tasks.", "full_text_question": "No robotic execution is tested in the abstract; use as an upstream perception threat.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "599b9342dcd4bcee34dd387be5b44976bc89e0e2ef00879b6f07dd1bed7e9388", "source_url": "https://arxiv.org/abs/2505.01050", "source_legacy_rows": [ 266 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0267", "t": "Safety Aware Task Planning via Large Language Models in Robotics", "y": 2025, "u": "https://arxiv.org/abs/2503.15707", "m": [ "Language planner" ], "k": "khan2025safety", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0267", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "SAFER combines planner feedback, LLM-based evaluation and CBF control with multi-robot hardware experiments.", "full_text_question": "Separate independent ground truth from LLM judging and identify which constraints the CBF actually guarantees.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "29eec543c340e6b79f314f06c80e908a03addc529ae0c168b68c3c1428827046", "source_url": "https://arxiv.org/abs/2503.15707", "source_legacy_rows": [ 267 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0268", "t": "Risk-Aware Decision Making in Restless Bandits: Theory and Algorithms for Planning and Learning", "y": 2024, "u": "https://arxiv.org/abs/2410.23029", "m": [], "k": "akbarzadeh2024riskaware", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0268", "study_type": "method", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Risk-aware restless bandits address resource allocation in replacement and patient scheduling rather than embodied control.", "full_text_question": "Exclude from robotic safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "75003d1e80ce9b14c51518fa899156ae86d13bca8a83738caf91b65d2e649aa2", "source_url": "https://arxiv.org/abs/2410.23029", "source_legacy_rows": [ 268 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0269", "t": "Learning Safe Autonomous Driving Policies Using Predictive Safety Representations", "y": 2025, "u": "https://arxiv.org/abs/2512.17586", "m": [], "k": "rev_driving", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0270", "t": "Formal Methods in Robot Policy Learning and Verification: A Survey on Current Techniques and Future Directions", "y": 2026, "u": "https://arxiv.org/abs/2602.06971", "m": [], "k": "manganaris2026formal", "role": "secondary", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0270", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "Surveys formal specification, policy learning and verification for robot safety and correctness.", "full_text_question": "Use to position the synthesis and compare scalability versus expressiveness of guarantees.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "8188edf352df7a6747e23d004baf7135b5d1d1c3eb5455e7400cd10790353146", "source_url": "https://arxiv.org/abs/2602.06971", "source_legacy_rows": [ 270 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0271", "t": "CN-CBF: Composite Neural Control Barrier Function for Safe Robot Navigation in Dynamic Environments", "y": 2026, "u": "https://arxiv.org/abs/2603.06921", "m": [], "k": "derajic2026cncbf", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0271", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "CN-CBF composes reachability-trained neural barriers for moving-obstacle avoidance on ground robots and quadrotors.", "full_text_question": "Check composition feasibility, learned-boundary certification and disturbance assumptions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "61752cee7dba5187cbfdd382810056007cde26148afe1690439f5a59355c9c87", "source_url": "https://arxiv.org/abs/2603.06921", "source_legacy_rows": [ 271 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0272", "t": "Control Barrier Functions with Audio Risk Awareness for Robot Safe Navigation on Construction Sites", "y": 2026, "u": "https://arxiv.org/abs/2602.12416", "m": [], "k": "mootz2026control", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0272", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "An audio risk detector adapts navigation CBF margins and reports zero violations in simulated construction scenes.", "full_text_question": "Check time-varying barrier conditions, missed audio hazards and deadlock versus safety trade-offs.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "386c7fed5e6da9d546fc12773ee47a5822803a87d32ce114918c4d6cf6a44ec6", "source_url": "https://arxiv.org/abs/2602.12416", "source_legacy_rows": [ 272 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0273", "t": "Learning to Nudge: A Scalable Barrier Function Framework for Safe Robot Interaction in Dense Clutter", "y": 2026, "u": "https://arxiv.org/abs/2601.02686", "m": [], "k": "jin2026learning", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0273", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Dense Contact Barrier Functions compose object-centric learned constraints for cluttered contact-rich interaction.", "full_text_question": "Verify composition guarantees, acceptable-contact definitions and simulation-only validation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "b9a3a200df30e06179d871801744e1317affd7771042e1bc7309e18852a9d125", "source_url": "https://arxiv.org/abs/2601.02686", "source_legacy_rows": [ 273 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0274", "t": "Contact-Aware Safety in Soft Robots Using High-Order Control Barrier and Lyapunov Functions", "y": 2025, "u": "https://arxiv.org/abs/2505.03841", "m": [ "Analytical dynamics" ], "k": "wong2025contactaware", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0274", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "High-order barrier and Lyapunov functions constrain whole-body soft-robot contact forces using a differentiable dynamics model.", "full_text_question": "Inspect model-error bounds, optimization feasibility and limits of planar simulation evidence.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "f39f02eaa9e804549c6e0a85d469bca6a2c5d56e1a33d38a72b3e0798cb10c4b", "source_url": "https://arxiv.org/abs/2505.03841", "source_legacy_rows": [ 274 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0275", "t": "Toward Maturity-Based Certification of Embodied AI: Quantifying Trustworthiness Through Measurement Mechanisms", "y": 2026, "u": "https://arxiv.org/abs/2601.03470", "m": [], "k": "darling2026toward", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0275", "study_type": "position", "relevance": "context", "full_text_priority": "medium", "assessment": "Maturity-based certification proposes quantitative trustworthiness measures illustrated by UAS detection uncertainty.", "full_text_question": "Distinguish a detection case study from certification of an entire embodied control loop.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "6e12f7a73bbdf7863cf3daec3058446a422bffe21e0dac6cb4277d8979e0752e", "source_url": "https://arxiv.org/abs/2601.03470", "source_legacy_rows": [ 275 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0276", "t": "Toward Autonomous Laboratory Safety Monitoring with Vision Language Models: Learning to See Hazards Through Scene Structure", "y": 2026, "u": "https://arxiv.org/abs/2602.00414", "m": [ "Language planner" ], "k": "chakraborty2026towardb", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0276", "study_type": "benchmark", "relevance": "context", "full_text_priority": "medium", "assessment": "Synthetic laboratory scenes and graphs evaluate VLM hazard detection and expose a pixel-to-structure gap.", "full_text_question": "Audit synthetic ground truth and note that autonomous robot intervention is not evaluated.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "5ca497611a2f579851de52f764642d25235ee9e1e7f165c639c0f2de00cdaf8c", "source_url": "https://arxiv.org/abs/2602.00414", "source_legacy_rows": [ 276 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0277", "t": "NavTrust: Benchmarking Trustworthiness for Embodied Navigation", "y": 2026, "u": "https://arxiv.org/abs/2603.19229", "m": [], "k": "jiang2026navtrust", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0277", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "NavTrust tests navigation under RGB, depth and instruction corruption and evaluates physical-robot mitigations.", "full_text_question": "Separate robustness gains from collision/harm reductions and inspect corruption realism.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "3f8b8f28a044681373cb2967707c4c5901cd2baf213ddee172f5030fe1b4d274", "source_url": "https://arxiv.org/abs/2603.19229", "source_legacy_rows": [ 277 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0278", "t": "HomeSafeBench: A Benchmark for Embodied Vision-Language Models in Free-Exploration Home Safety Inspection", "y": 2025, "u": "https://arxiv.org/abs/2509.23690", "m": [ "Language planner" ], "k": "gao2025homesafebench", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0278", "study_type": "benchmark", "relevance": "direct", "full_text_priority": "high", "assessment": "HomeSafeBench evaluates active visual exploration for detecting household hazards in simulated homes.", "full_text_question": "Inspect hazard ground truth, exploration coverage and precision/recall; detection does not imply mitigation.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "a9daae39c2a4415a9d951618bb2ed12936d42545e55a0adfbdeb1d3b2864bc15", "source_url": "https://arxiv.org/abs/2509.23690", "source_legacy_rows": [ 278 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0279", "t": "Safe Human Robot Navigation in Warehouse Scenario", "y": 2025, "u": "https://arxiv.org/abs/2503.21141", "m": [], "k": "farrell2025safe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0279", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Learning-based CBFs integrated with OpenRMF support warehouse robot avoidance of static and moving obstacles.", "full_text_question": "Verify multi-robot compatibility, human-motion assumptions and evaluation setting.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "80eee8f8a69b38f4c815729b707b975d673d897dd167611e6d75233e09fff6f0", "source_url": "https://arxiv.org/abs/2503.21141", "source_legacy_rows": [ 279 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0280", "t": "What Breaks Embodied AI Security:LLM Vulnerabilities, CPS Flaws,or Something Else?", "y": 2026, "u": "https://arxiv.org/abs/2602.17345", "m": [ "Language planner" ], "k": "ma2026what", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0280", "study_type": "review", "relevance": "direct", "full_text_priority": "high", "assessment": "A systems-security survey emphasizes semantic-physical mismatch, uncertainty and failure propagation across the control loop.", "full_text_question": "Compare overlap with the manuscript's systems argument and identify distinct contributions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "071d1f77c12198a6a2809c8c6412fdf4751961ea89d015a70a83948f995b757d", "source_url": "https://arxiv.org/abs/2602.17345", "source_legacy_rows": [ 280 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0281", "t": "A Hazard-Informed Data Pipeline for Robotics Physical Safety", "y": 2026, "u": "https://arxiv.org/abs/2603.06130", "m": [], "k": "odinokov2026hazardinformed", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0281", "study_type": "position", "relevance": "direct", "full_text_priority": "medium", "assessment": "A hazard-informed pipeline links asset analysis, synthetic data and learned safety envelopes.", "full_text_question": "Verify actual implementation and evidence supporting learned-envelope validity.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "0da6badadd7ad9914bb31cf5a90336c517871eec45f4ec38101fce290ffb55e1", "source_url": "https://arxiv.org/abs/2603.06130", "source_legacy_rows": [ 281 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0282", "t": "Adversarial attacks against Modern Vision-Language Models", "y": 2026, "u": "https://arxiv.org/abs/2603.16960", "m": [ "Language planner" ], "k": "latorre2026adversarial", "role": "out_of_scope", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0282", "study_type": "method", "relevance": "out_of_scope", "full_text_priority": "low", "assessment": "Adversarial VLM agents are tested in an e-commerce environment without physical embodiment.", "full_text_question": "Exclude from robotic safety-effectiveness comparisons.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "a458f35b8170e7ec9d83cb9ef43c6219c56ae518a5ff845b3bc9ba2353436c32", "source_url": "https://arxiv.org/abs/2603.16960", "source_legacy_rows": [ 282 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0283", "t": "EigenSafe: A Spectral Framework for Learning-Based Probabilistic Safety Assessment", "y": 2026, "u": "https://arxiv.org/abs/2509.17750", "m": [], "k": "jang2026eigensafe", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0283", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "EigenSafe learns an operator eigenpair tied to stochastic safety probability and uses it in safe policy learning and physical manipulation.", "full_text_question": "Inspect probability calibration, operator approximation and conditions for constraining policy updates.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "7378d1bff2a0ef9474b8919308ec16815bb10a6a26a9a5b0504dc7701066bcb0", "source_url": "https://arxiv.org/abs/2509.17750", "source_legacy_rows": [ 283 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0284", "t": "Cybersecurity AI: Hacking Consumer Robots in the AI Era", "y": 2026, "u": "https://arxiv.org/abs/2603.08665", "m": [], "k": "mayoralvilches2026cybersecurity", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0284", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Consumer-robot security case studies report AI-assisted discovery of vulnerabilities with safety-critical control implications.", "full_text_question": "Verify methodology and disclosure evidence without reproducing attacks or sensitive extracted data.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "1307bfb499d7dd8b0694164f872c0a0ac1b1a3cf32e0e88dae4b1367e799b13c", "source_url": "https://arxiv.org/abs/2603.08665", "source_legacy_rows": [ 284 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0285", "t": "Generating Robot Constitutions & Benchmarks for Semantic Safety", "y": 2026, "u": "https://arxiv.org/abs/2503.08663", "m": [ "Language planner" ], "k": "deepmind2026generating", "role": "evaluation_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0285", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "ASIMOV benchmarks semantic safety and derives robot constitutions that reject undesirable actions.", "full_text_question": "Audit generated scenarios, preference-based ground truth and residual unsafe actions.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "a417065a7d80797f285535c49601ee0c4a13d72209859324ea1a567ff105bb53", "source_url": "https://arxiv.org/abs/2503.08663", "source_legacy_rows": [ 285 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0286", "t": "Multimodal Adversarial Quality Policy for Safe Grasping", "y": 2026, "u": "https://arxiv.org/abs/2603.01479", "m": [], "k": "institutions2026multimodal", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0286", "study_type": "method", "relevance": "direct", "full_text_priority": "medium", "assessment": "Multimodal adversarial patches steer RGBD grasp-quality policies toward purportedly safer behavior on datasets and a cobot.", "full_text_question": "Clarify the safety objective and whether suppressing grasp predictions reliably prevents harm.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "aeb121e894332162f2bf7eba2e2eb12bbe2dbccb6798fbffe679f412c7e3aa99", "source_url": "https://arxiv.org/abs/2603.01479", "source_legacy_rows": [ 286 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0287", "t": "SAC-Loco: Safe and Adjustable Compliant Quadrupedal Locomotion", "y": 2025, "u": "https://arxiv.org/abs/2509.23223", "m": [], "k": "institutions2026sacloco", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0287", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "SAC-Loco uses a learned safety critic to switch between compliant locomotion and recovery under external forces.", "full_text_question": "Check critic reliability, switching stability and the boundary of recoverable disturbances.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "a24f2445af440182f8c753eba7c97f527259de24d2367ffc4a2e4093b4a19eec", "source_url": "https://arxiv.org/abs/2509.23223", "source_legacy_rows": [ 287 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0288", "t": "CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions", "y": 2025, "u": "https://arxiv.org/abs/2510.14959", "m": [], "k": "institutions2026cbfrl", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0288", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "CBF-RL filters training rollouts and claims safe deployment without a runtime filter, including humanoid experiments.", "full_text_question": "Determine precisely which guarantees transfer to the learned unfiltered policy under approximation and shift.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "0e49e3611091509fc8aca20ba65b084a3f2671900a3e9e1212a249e88dad58da", "source_url": "https://arxiv.org/abs/2510.14959", "source_legacy_rows": [ 288 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0289", "t": "Deep QP Safety Filter: Model-free Learning for Reachability-based Safety Filter", "y": 2026, "u": "https://arxiv.org/abs/2601.21297", "m": [], "k": "institutions2026deep", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0289", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Deep QP Safety Filter learns reachability values and derivatives for black-box systems with exact-setting convergence claims.", "full_text_question": "Separate exact contraction results from finite-data neural approximation and residual failures.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "bd10f63372938ef11dd9b92e432ec17d36271386352f1034f90be694da3a6eba", "source_url": "https://arxiv.org/abs/2601.21297", "source_legacy_rows": [ 289 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0290", "t": "Q-learning-based Model-free Safety Filter", "y": 2024, "u": "https://arxiv.org/abs/2411.19809", "m": [ "Analytical dynamics" ], "k": "institutions2026qlearningbased", "role": "safety_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0290", "study_type": "method", "relevance": "direct", "full_text_priority": "high", "assessment": "Q-learning estimates safety values and a theoretically supported filtering threshold, with soft-robot hardware evaluation.", "full_text_question": "Check threshold assumptions, learning error and unseen-state action admissibility.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "09f4545e090519d2d6ffd92c70f0e7091054886b4a5da2b9408b64090f484bc6", "source_url": "https://arxiv.org/abs/2411.19809", "source_legacy_rows": [ 290 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-legacy-0291", "t": "One Filter to Deploy Them All: Robust Safety for Quadrupedal Navigation in Unknown Environments", "y": 2024, "u": "https://arxiv.org/abs/2412.09989", "m": [], "k": "rev_ocr", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-legacy-0292", "t": "From Chat Control to Robot Control: Implications of the Chat Control Proposal for Human-Robot Interaction", "y": 2026, "u": "https://arxiv.org/abs/2601.02205", "m": [], "k": "mit2026chat", "role": "background_candidate", "verified": false, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "abstract_screened", "abstract_review": { "record_id": "eai-legacy-0292", "study_type": "position", "relevance": "context", "full_text_priority": "low", "assessment": "An HRI policy essay analyzes surveillance, privacy and security implications of proposed communication monitoring.", "full_text_question": "Use for governance discussion only; no technical predictive-safety mechanism is evaluated.", "reviewed_on": "2026-09-11", "reviewer": "OpenAI Codex; AI-assisted individual abstract reading", "review_basis": "Supplied title and complete abstract; not fresh publisher verification or full-text inspection", "author_validation": "pending", "abstract_sha256": "90dce9671f1658294d66b2815cf02a58b055aeffae886698447825ce479d7128", "source_url": "https://arxiv.org/abs/2601.02205", "source_legacy_rows": [ 292 ], "source_file": "data/review_embodied_ai - safety_embodied_ai.csv", "claims_status": "Authors reported abstract claims; technical validity not independently verified" } }, { "id": "eai-2c90f176f276", "t": "Control Barrier Function Based Quadratic Programs for Safety Critical Systems", "y": 2016, "u": "https://arxiv.org/abs/1609.06408", "m": [], "k": "rev_cbfqp", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-bd726552dd30", "t": "A predictive safety filter for learning-based control of constrained nonlinear dynamical systems", "y": 2018, "u": "https://arxiv.org/abs/1812.05506", "m": [], "k": "rev_psf", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-b82e3b6ce0bc", "t": "Predictive control barrier functions: Enhanced safety mechanisms for learning-based control", "y": 2021, "u": "https://arxiv.org/abs/2105.10241", "m": [], "k": "rev_pcbf", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-43425e92faf5", "t": "A predictive safety filter for learning-based racing control", "y": 2021, "u": "https://arxiv.org/abs/2102.11907", "m": [], "k": "rev_racing", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-e7fb51bde94b", "t": "ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models", "y": 2026, "u": "https://arxiv.org/abs/2606.27079", "m": [ "VLA policy" ], "k": "rev_foresight", "role": "benchmark", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-c03a46a60a0a", "t": "How to Learn from What a Human Would Avoid? Intervention-Aware World Models with Real-World RL for Dexterous Manipulation", "y": 2026, "u": "https://arxiv.org/abs/2609.06009", "m": [ "Learned dynamics" ], "k": "rev_intervention", "role": "safety_candidate", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-d29181274c9c", "t": "Security of World-Model-Based Embodied AI: A Lifecycle of Threats, Defenses, and Evaluation", "y": 2026, "u": "https://arxiv.org/abs/2607.28226", "m": [], "k": "rev_lifecyclesurvey", "role": "secondary", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null }, { "id": "eai-309fbf18adbc", "t": "Vision-Language-Action Safety: Threats, Challenges, Evaluations, and Mechanisms", "y": 2026, "u": "https://arxiv.org/abs/2604.23775", "m": [ "VLA policy" ], "k": "rev_vlasurvey", "role": "secondary", "verified": true, "tag_status": "title/abstract proposal; not validated full-text coding", "review_level": "full_text_inspected", "abstract_review": null } ]