Survey of Safety Risks and Attacks in Embodied AI
A recent study published on arXiv (2605.02900) provides a comprehensive review of safety research related to embodied artificial intelligence (AI), which combines perception, cognition, planning, and interaction for agents functioning in open-world settings. In contrast to digital AI, embodied agents encounter unpredictable sensing, limited knowledge, and evolving human-robot interactions, where failures may lead to physical dangers. The research analyzes vulnerabilities and protective measures throughout the entire embodied process—from perception and cognition to planning, action, interaction, and agentic systems. It presents a multi-tiered taxonomy to consolidate disparate research and link safety insights specific to embodied systems with the wider AI safety discourse. The survey addresses fields like transportation, healthcare, and robotics, highlighting the technical and social importance of safety as these systems become more autonomous.
Key facts
- Survey published on arXiv with ID 2605.02900
- Focuses on embodied AI safety, covering perception, cognition, planning, action, interaction, and agentic systems
- Embodied AI integrates perception, cognition, planning, and interaction
- Domains include transportation, healthcare, industrial robotics, and assistive robotics
- Failures in embodied AI can lead to physical harm
- Introduces a multi-level taxonomy to unify fragmented safety research
- Examines attacks and defenses across the full embodied pipeline
- Connects embodied-specific safety findings with broader AI safety literature
Entities
Institutions
- arXiv