Rethinking Safety for Generalist Robots

Safety for versatile robots requires rethinking beyond traditional collision and force-based hazards. The paper argues that as robots take on diverse tasks like meal preparation, laundry, and infrastructure inspection, safety frameworks must expand to address context-dependent risks, unspoken user expectations, and unpredictable physical consequences that current robotics safety standards don't adequately cover.

Context and intent matter critically for robot safety decisions. The paper highlights that the same action can be safe or unsafe depending on circumstances—for example, turning off electricity is only safe during scheduled maintenance—and robots must understand user intent beyond literal instructions, such as recognizing that "clean the kitchen" implicitly means avoiding dangerous chemical combinations like bleach and ammonia.

"Embodied AI safety" requires integrating digital and physical safety considerations throughout a robot's entire lifecycle. The paper emphasizes that safety cannot be separated into "bits" (software/AI decisions) and "atoms" (physical actions), but must be considered holistically from design through deployment, addressing hard-to-predict physical consequences like burning food during cooking.

A comprehensive taxonomy and research agenda is needed to guide safe deployment of generalist robots. Rather than applying existing safety frameworks, the paper calls for a new, full-stack approach to robotics safety that systematically addresses the broader range of emerging risks introduced by versatile AI systems operating in human environments.