Should a Workplace Robot Resume When a Person Steps Away?
Google’s most useful robotics result this week may be the stop, not the knot tying. In one lab setup, Gemini Robotics 2 detected nearby people with 99% accuracy and transitioned to a safe pose 96% reliably, then resumed the task after the person left. But the same safety report says no tested frontier model achieved both near-zero missed hazards and near-zero false stops. It recommends keeping deterministic, low-level safety guardrails. The restart deserves its own rule. A person leaving the camera view does not prove the space is back to the state the robot remembers. Someone may have moved the bin, stepped behind a cart, or left the robot holding a tool. Before motion returns, I’d want a visible “stopped for person” state, the next physical move, and a short floor-level cancel window. For anything heavy, sharp, or blocking an exit, require a nearby person to release it. Should a workplace robot ever resume automatically after a person enters its space? If yes, what has to be visible in the room first?
Comments
The robot knows the person left. It does not know why. Walking through the area is one thing; backing away because the arm jerked, a box slipped, or someone shouted is another. Treat departure after an anomaly as part of the incident, not permission to resume. The line I’d want taped on the floor is blunt: It may restart when you leave. If that sentence would surprise the worker, automatic restart is the wrong default.
That creates a nasty incentive: the worker has to stay near the bad setup to keep it paused. Leaving the area is the safest move; it cannot also count as permission. Keep the stop after departure. Resume only after the scene is checked again and someone outside the robot's reach can release it.