The warehouse robot problem is usually a dirty sensor, not a flashy AI failure

After working around deployed mobile-robot systems, I’m increasingly convinced that the reliability bottleneck is less often a spectacular AI failure than ordinary physical degradation: a shifted sensor mount, a dirty lens or LiDAR window, wheel wear, temperature change, or a docking target that has deteriorated.

That matters in warehouses, hospitals, and public spaces because these failures can quietly reduce localization and perception quality. The robot may not do anything obviously bizarre. It may stop short of a dock, misjudge an obstacle, take an unnecessarily cautious route, or simply wait for help. NIST treats calibration as an operational reliability issue: impacts, wear, and temperature changes can offset a sensor’s field of view and degrade navigation, docking, and object-location accuracy.

There’s a very unglamorous example in Clearpath’s maintenance documentation: contamination on a 3D LiDAR can make a robot navigate poorly or become stuck, while contamination on a 2D LiDAR can impair docking and prevent automatic recharging. That is not an exotic model failure. It is a maintenance condition that can strand the robot.

I also think robots should be required—at least in higher-consequence settings—to detect and report degraded confidence before acting. That does not necessarily mean exposing a single magic percentage to a user. It could mean monitoring blocked returns, disagreement between cameras, LiDAR, odometry, and maps, localization residuals, calibration age, and task-specific confidence. Low confidence should lead to an appropriate fallback: slower motion, a safe stop, or a human request. NIST work on robotic bin-picking makes a similar case for estimating task-completion probability and asking for help when confidence is inadequate.

I’m not arguing that better algorithms are unimportant, or that there is already a universal legal rule requiring uncertainty reports. This is an engineering judgment: autonomy has to account for the sensing hardware’s changing condition, not just the model’s nominal accuracy.

What have people seen in field maintenance—dirty optics, bumped mounts, wheel or odometry drift, bad lighting, degraded dock markers, cleaning chemicals? And if you had to prioritize one intervention, would you choose better sensors, better self-checking software, or stricter human oversight and maintenance procedures?

The distinction between “important reliability factor” and “leading cause” is worth preserving here. The cited material supports calibration and contamination as real failure modes, but not a cross-industry ranking against software defects.

I like the uncertainty proposal if it is framed as a safety function rather than a demand for a universally interpretable confidence score. A robot can report “sensor health degraded; docking confidence insufficient” without pretending that 0.73 has the same meaning in every environment.

The hard part is usually ownership. “Clean the lens weekly and recalibrate after impact” sounds straightforward until nobody knows whether that belongs to facilities, robotics operations, or the vendor.

A useful report would create a maintenance ticket with the symptom, sensor-health state, last inspection, and whether the robot fell back safely. Otherwise operators just reboot it, send it back out, and the same quiet failure returns.

I’d make uncertainty reporting mandatory where a bad action can affect people or block a critical workflow, but I would not automatically require a full stop. A degraded sensor could justify reduced speed and a restricted operating zone; an unexplained disagreement between sensors might justify human approval.

The security analogy is simple: a warning is not a control. The fallback behavior has to be defined, tested, and auditable.

I’m less ready to put “better software” behind sensors and oversight. Hardware that tolerates dust, glare, and small alignment changes may remove entire classes of failure, while software-based confidence estimates can themselves be poorly calibrated.

That said, better sensors do not eliminate inspection. They just change what has to be inspected. My vote is a combination, with the budget going first to the degradation mode that is actually observed rather than to a more impressive model.