Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses

arXiv 26.05 2026 35 method

TLDR

A comprehensive survey of safety research in embodied AI, covering attacks and defenses across perception, cognition, planning, and interaction.

Reasoning

The paper provides a structured taxonomy and synthesizes over 500 papers, which is a strength for organizing fragmented research. However, as a survey, it lacks new empirical contributions or real-world experiments, limiting its direct impact.

Read-first score

Read-first score 35, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 0.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=github

Citation impact 15%

71.6

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.71574477

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Topical relevance 25%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 338.

Keyword Scores

world model

world simulator

generative world model

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Comprehensive structured review of safety research in embodied AI
Multi-level taxonomy unifying fragmented lines of work and connecting embodied-specific safety findings with broader advances in vision, language, and multimodal foundation models
Synthesis of insights from over 500 papers spanning adversarial, backdoor, jailbreak, and hardware-level attacks; attack detection, safe training and robust inference; and risk-aware human-agent interaction
Identification of overlooked challenges: fragility of multimodal perception fusion, instability of planning under jailbreak attacks, and trustworthiness of human-agent interaction in open-ended scenarios

Methodology

The survey reviews over 500 papers on attacks and defenses across the full embodied pipeline, from perception and cognition to planning, action and interaction, and agentic system. It introduces a multi-level taxonomy that unifies fragmented lines of work and connects embodied-specific safety findings with broader advances in vision, language, and multimodal foundation models.

Key Results

The survey identifies several overlooked challenges including the fragility of multimodal perception fusion, instability of planning under jailbreak attacks, and trustworthiness of human-agent interaction in open-ended scenarios, and provides a roadmap for building embodied agents that are safe, robust, and reliable in real-world deployment.

Resources

Paper Code