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

Embodied AI systems face unique safety vulnerabilities across their entire operational pipeline — from perception and cognition through planning and human interaction — because unlike digital AI, these robots operate in physical environments where failures can cause direct physical harm to people, making safety considerations essential as these systems enter transportation, healthcare, and assistive robotics domains.

Multimodal perception fusion represents a critical but overlooked weakness: The survey identifies the fragility of how embodied AI systems combine information from multiple sensors (vision, language, etc.) as a significant safety gap, suggesting that the integration points between different types of sensory data are particularly vulnerable to attacks and failures.

Planning systems become unstable when subjected to jailbreak attacks, meaning that adversaries can manipulate embodied AI agents into ignoring their safety constraints during decision-making, highlighting that current safeguards may not adequately protect the reasoning stage where robots decide what actions to take.

The survey synthesizes findings across 500+ papers on multiple attack types — including adversarial attacks, backdoor attacks, jailbreaks, and hardware-level attacks — demonstrating that threats to embodied AI safety are diverse and multifaceted, requiring defenses at multiple levels rather than single-point solutions.

Human-agent interaction trustworthiness remains unresolved in open-ended scenarios, meaning we still lack clear frameworks for ensuring that robots can safely and reliably interact with humans in unpredictable real-world situations, which is critical as these systems become more autonomous and prevalent in shared human spaces.