From bits to atoms: A survey of cross-layer safety in Embodied AI

Physical harm is a unique risk for AI systems that interact with the real world: Unlike traditional software, failures in embodied AI (robots, autonomous vehicles, etc.) can directly cause physical injury or damage because digital errors propagate from "bits to atoms." The survey identifies how vulnerabilities at any layer—from cloud data systems to physical actuators—can cascade across the system to produce real-world harm.

Safety requires coordinated defense across the entire system stack, not just the algorithm: The paper emphasizes that protecting embodied AI demands attention to four interconnected layers (physical interaction, system/middleware, algorithm/decision, and cloud-edge/data). A single weak point—whether a compromised sensor, unpatched software, or insecure data connection—can undermine safety across the entire system, meaning piecemeal solutions are insufficient.

Attacks can chain together across system boundaries in dangerous ways: The survey systematizes how attackers can exploit multiple vulnerabilities in sequence to cause physical harm. This means security assessments must consider not just individual component failures, but how failures in one layer (e.g., a cloud data breach) could enable attacks that manifest as dangerous physical behavior.

Current safety approaches need to bridge the gap between software alignment and hardware constraints: The paper advocates for unifying AI alignment efforts (ensuring systems behave as intended) with rigorous physical safety engineering. This suggests responsible embodied AI requires collaboration between AI researchers and traditional safety engineers rather than siloed approaches.