Knowledge / Robot Safety

Robot Safety

A practical overview of safety layers for embodied AI systems operating around people.

Safety is a stack

Physical AI safety can involve mechanical limits, collision avoidance, perception of people, uncertainty handling, policy constraints, supervisory control and emergency stopping.

Uncertainty matters

Unlike a purely digital agent, a robot can cause physical consequences when its model is uncertain. Systems therefore need mechanisms to detect uncertainty and defer or stop when appropriate.

Evaluation

Safety evaluation should consider both nominal task success and failure behavior around humans, tools, objects and changing environments.

Primary source

Read the primary source →

Editorial guide · Reviewed September 2026