A Survey · 2026

Predictive Safety in Embodied AI

From Control-Theoretic Filters to Foundation Models

  1. Kleyton da Costa1,2
  2. Adriano Koshiyama2
  3. Dimitrios Kanoulas1
  4. Philip Treleaven1
  1. 1University College London
  2. 2Holistic AI

Structured review · 10 September 2026

Predicting an unsafe future is useful only if the prediction supports a timely and effective intervention. We compare control-theoretic filters, constrained learning, world models, and vision-language-action policies through their safety properties, assumptions, and evaluation evidence. Predictive capability alone does not establish physical safety.

Candidates
288
Full-text inspected
26
Abstract screened
262
Primary methods
15
Updated
10 Sep 2026

All 288 candidates have an assessment: 26 full-text inspections and 262 additional AI-assisted readings of supplied abstracts. The search is purposive, not exhaustive. Abstract judgments, full-text evidence and historical keyword tags are kept separate.

The predictive-safety loop

Open figure PDF
Grounded state, anticipated consequences, and a feasible intervention must work together.

An auditable candidate collection

Open figure PDF
The audit distinguishes 26 full-text inspections from 262 additional abstract-only assessments.

A continuous line of research

Open figure PDF
Dates refer to first public sources; archival publication may be later.

Coverage of the full collection

Open figure PDF
Every candidate appears in both the study-role and publication-year totals. Full-text evidence roles and historical provisional classifications are distinguished.

Comparison tables

The manuscript provides native tables comparing related reviews, all 17 inspected methods and predictive architectures, benchmark capabilities, and research directions. Each technical comparison includes source citations. Read the tables in the paper.

Explore the candidate collection

Search titles or assessment text; use high, medium or low to find full-text priorities. Expand a paper to read its abstract assessment. These AI-assisted judgments await author validation and do not establish a safety guarantee.

    Citation and materials

    Read the paper · Download candidate metadata · Code, protocol, and evidence

    @misc{dacosta2026predictivesafety,
      title={Predictive Safety in Embodied AI: From Control-Theoretic Filters to Foundation Models},
      author={da Costa, Kleyton and Koshiyama, Adriano and Kanoulas, Dimitrios and Treleaven, Philip},
      year={2026}
    }