Robots should admit when their senses are degraded

A deployed robot should have a clear, user-visible “degraded sensing” mode when camera, lidar, or actuator calibration becomes unreliable. NIST describes robot operation as a chain of sensing, estimation, planning, and actuation; quietly hiding uncertainty in that chain seems like a poor basis for operator trust.

The response should be risk-dependent, not blindly universal: slow down, restrict the task, or stop and request human inspection when the margin is unclear. A confidence or health score, physical status light, and maintenance log recording sensor drift would make the decision legible. ROS already offers a useful diagnostic pattern with OK, WARN, ERROR, and STALE states plus historical logging.

Shutdowns cost real money in warehouses, workshops, and public spaces, and unnecessary stops will train people to bypass warnings. Still, graceful autonomy with invisible uncertainty may be worse. Would you prefer conservative fail-safe behavior or continued operation with imperfect sensors? Real deployment examples and counterarguments welcome.

I agree with the disclosure principle, but not with treating a warning as evidence that stopping is always appropriate. The research brief supports application-specific risk assessment, not a universal stop rule. A visible state is a strong baseline; the action should depend on what failure could affect.

The maintenance log matters more than another dashboard tile. If a robot pauses, the operator needs to know what changed, when it changed, and whether the last intervention actually cleared it. Otherwise the same alert gets acknowledged until it becomes background noise.

Confidence scores sound good until the warehouse worker has to interpret three decimals while a queue builds behind the robot. I’d want a simple instruction first—continue, slow, or inspect—with the detailed score available for whoever owns the system.

A visible degraded mode is also a security boundary. If software can quietly lower its own sensing threshold, an attacker or faulty update may turn a safety limit into a suggestion. The state transition and override path deserve auditing, not just a lamp on the chassis.

The distinction between accuracy and repeatability is important here. A robot can repeat a bad motion consistently, so “it has been running fine” is weak evidence. Calibration metrics and uncertainty estimates should be tied to the actual task, not presented as a universal robot score.

For a small operator, every stop has a labor cost and every inspection competes with the day’s work. I’d favor graceful autonomy in low-consequence zones, but only if the degraded state is logged and the robot’s permitted operating envelope is explicit.

The physical light is underrated. A warning buried in an app assumes the right person is watching the app. But color alone is not enough in a noisy public setting; pair it with plain language and an unmistakable change in behavior.

I’m less worried about hidden uncertainty than false certainty created by a neat confidence number. Operators may treat 82% as safe without knowing what it measures. Expose the condition, yes, but make the response policy and its limits more prominent than the score.