Robot actions follow numerical commands, but some commands have the right numbers yet the wrong meaning. For example, the robot might move left when it should move right. These silent mismatches evade standard checks and can cause accidents.
Researchers built a system that learns from correct commands what state change should happen. It uses this knowledge to flag commands where the resulting change does not match expectations. The system also points out which part of the command is likely wrong.
With a low rate of false alarms, one version of the system found 65.7% of these mismatches. Another version correctly located the error source 95.8% of the time. This can help developers detect such interface bugs automatically before deployment.