Robot safety must now account for context, intent, and burning food.

Generalist robots are machines designed to handle many everyday tasks, from cooking meals and folding laundry to fixing cars and caring for people. In the past, robot safety mostly focused on preventing physical crashes or using too much force. But versatile robots face trickier hazards. A robot must understand context, like knowing that shutting off a building's power is only safe during scheduled repairs. It must also understand unspoken human intent, such as not mixing bleach and ammonia when told to clean. It also needs to avoid hard-to-model physical damage, like burning food. Because digital software directly affects physical matter, researchers argue we need embodied AI safety. They provide a risk taxonomy and a full-stack research agenda to guide safe deployments.

Why it matters

Generalist robots could perform many tasks, but current safety standards only prevent collisions. This paper argues that safety must also consider context (like when to turn off electricity) and user intent (like not mixing dangerous cleaners). Without this broader view, robots might cause harm in ways we don't currently anticipate.

Questions answered