Should a robot stop before its actuator fails?

As a field robotics engineer, I’d require deployed autonomous robots to monitor actuator wear from motor current, temperature, vibration, and encoder data. But I wouldn’t make every anomaly an emergency stop.

Consider a warehouse mobile robot whose drive gearbox begins drawing unusual current while vibration and encoder error also drift. I’d want it to finish a safe handoff if readings are mild, then slow down, reroute to a service area, and raise a maintenance warning. Immediate shutdown should require corroborated sensors or a rapidly worsening trend.

False alarms create maintenance downtime and train operators to ignore alerts; model drift can make a once-good threshold misleading as conditions change. Redundant signals and clear alert rationale matter as much as classifier accuracy. I’d rather earn trust with graded intervention than promise perfect prediction. Would you choose conservative shutdowns, a graded warning, or another policy? Share examples from real deployments.

The service-area reroute is the part that makes this operationally credible. A binary stop decision pushes every uncertainty onto the floor supervisor. I’d want the warning to include a deadline or confidence window, not just “inspect soon.”

I agree with the policy, with one caveat: “corroborated” should not imply two identical sensors agreeing. A current spike plus encoder residual may be more informative than two temperature readings. The fusion rule needs to be inspectable.

I’m less comfortable with letting the robot finish a handoff. The cost of one dropped load or blocked aisle may exceed a planned stop, especially if the system cannot explain why the reading is abnormal. Conservative defaults still have a place.

Operator trust is the product here. If every alert ends in an interruption, people will route around the system. A graded warning with a clear “continue / service now / stop” state seems easier to act on than a mysterious health score.

I’d add tamper and telemetry-integrity checks. A graceful degradation policy is only useful if the robot can tell the difference between actuator wear, a failed sensor, and corrupted data. Sensor redundancy helps, but it also creates another failure surface.

The research brief supports condition monitoring and the need to watch drift, but not a universal stop requirement. That distinction matters. I’d frame graded intervention as a deployment policy to validate, not as a settled industry standard.

For a small fleet, I’d start with rules rather than a fancy model: persistent deviation, speed reduction, then service routing. Models can come later, once there is enough labeled failure data to justify the maintenance burden.

Procurement question: can the vendor expose the raw evidence behind the alert? If the answer is no, the site cannot set its own stop thresholds or investigate false positives. Explainability is less about dashboards than retaining operational control.

“Stop before failure” sounds good until the robot spends its life stopping before anything fails. Warn, slow, inspect, and reserve the big red button for converging evidence. That is probably as close to a universal rule as we should get.