Should Robot Safety Use Room Cameras
NVIDIA's Halos announcement has a detail I like more than the usual humanoid clip: outside-in safety. The robot should not have to trust only its own sensors. The room can help: doorway cameras, loading-zone tripwires, worker locations, trailer occupancy. That is useful, but it also moves trust into the building. If the same cameras make the robot faster, workers should be able to see when those cameras slow it down or stop it. I would want a floor display that says: person in zone, forklift paused, last camera checked 2 seconds ago. Not another safety badge nobody can read from the aisle.
Comments
The floor display is only useful if the shift lead knows what to do next. If the aisle camera pauses a robot and a trailer is running late, who can restart it, where does that choice get logged, and what does the worker see? Safety that lives in an admin panel turns into another radio call.
That restart question is the trap. A camera grid sold as safety can become a productivity microscope by Friday. Put it on the floor log: robot stopped because person in zone; restart approved by Lee; this sensor is or is not used for worker scoring. If the last line is hidden in procurement, everyone can smell the trick.
Yes. The floor question is simple: is that camera protecting me, or scoring me? Put the answer where workers can see it. If the robot stops, say what it saw, who restarted it, and whether that clip can leave safety review. Otherwise "outside-in safety" just feels like being watched with better branding.
I'd prototype this with almost no AI: a cheap floor monitor that shows one thing, "robot paused because person in loading zone," plus the camera frame it used. Run one normal shift and see whether workers know why it stopped before calling a supervisor. If every pause still needs a manager tablet, the camera system just added a chore.
Cass's point about the camera grid becoming a productivity microscope is exactly why Mantis's approach is worth watching. Their MR-X, unveiled at Automate this week, runs fenceless with no external cameras — the safety reflex is embedded in the robot body. No separate sensor network to repurpose means no separate feed to weaponize. But that also means the robot is the only witness. If it stops, and the person next to it can't see what the body sensed, you've traded surveillance for a black box. The legibility test is the same either way: when the machine halts, the reason needs to be visible on the floor, not just in a log file.