Silent Failures in Physical AI: A Literature Review of Runtime Action Authorization for Autonomous Systems

A literature review identifies a critical safety gap in Physical AI systems—autonomous robots, drones, and vehicles that make physically consequential decisions based on visual and language inputs. The research reveals that existing safety mechanisms from AI content moderation and robotics operate independently, leaving no unified runtime authorization system to prevent silent failures where confident but incorrect model outputs cause real-world harm before hardware safeguards activate.

Physical AI systems represent a convergence of foundation models, vision-language models, and world models deployed in hardware that directly affects the physical environment. Unlike content moderation failures, which are primarily informational harms, and traditional robot safety systems, which rely on hardware limits, Physical AI creates a novel failure mode: a system can appear confident and semantically correct while issuing physically dangerous commands due to sensor drift, distribution shift, hallucinated affordances, or invalid physical assumptions. This silent failure problem emerges because the model's black-box decision-making operates upstream of hardware controllers, meaning errors propagate into physical consequences before detection occurs.

The research synthesizes insights across eight technical domains—embodied foundation models, world models, simulation, safety benchmarks, safe control, runtime assurance, uncertainty estimation, and verification—yet identifies a critical gap: no existing approach provides a complete runtime authorization boundary between the AI model and physical execution. This fragmentation matters because foundation models are increasingly deployed in safety-critical applications like autonomous vehicles, industrial robots, and surgical systems. The gap represents a systemic risk in the AI industry's transition toward embodied systems.

The implications extend across development, deployment, and insurance landscapes. Engineers lack standardized guardrail evaluation frameworks, creating inconsistent safety standards across robotics companies. Investors in autonomous systems face unquantified liability exposure. The research catalyzes demand for runtime assurance architectures that operate independently of model confidence signals, incorporating redundant sensing, state validation, and physics-based constraint checking. This work signals that Physical AI safety requires new interdisciplinary approaches combining formal verification, uncertainty quantification, and hardware design.

• →Physical AI systems exhibit silent failures where confident models issue physically dangerous commands undetected by existing safety mechanisms.
• →Eight technical domains in robotics and AI have advanced safety independently without creating unified runtime authorization systems.
• →Silent failures stem from sensor drift, distribution shift, hallucinated affordances, and invalid physical assumptions in embodied foundation models.
• →No current framework standardizes evaluation of runtime guardrails for Physical AI applications across industry.
• →The gap demands new interdisciplinary approaches combining formal verification, uncertainty quantification, and physics-based constraint checking.