Robot Failure Detection

Robot failure detection studies how to recognize failed, unsafe, or off-course behavior during tasks such as manipulation, locomotion, inspection, and industrial operation. Approaches range from sensor and trajectory anomaly detection to action monitoring, uncertainty estimates, conformal prediction, world models, and vision-language models that identify or explain failures. Methods are evaluated on benchmarks such as LIBERO, RoboFAC, and BotFails, as well as real-robot tasks, with growing attention to warning timeliness, unseen tasks, and sensor limitations. A central tension is that accurate alarms must also avoid false positives and lead to useful interventions, since detection alone does not ensure recovery or safer operation.