Detecting and Explaining Robot Policy Failures from Sparse Internal Features

This research presents MIND (Mechanistic Interpretability for Neural Diagnosis), a method that utilizes sparse autoencoders to analyze the internal activations of trained visuomotor policies. By linking these features to the robot's position, movement, and surroundings, MIND identifies specific internal concepts contributing to failures during execution. The approach flags deviations from learned patterns in both successful and failed runs, providing explanations for policy breakdowns based on meaningful feature changes. Experiments conducted in simulation and on a physical robot demonstrate that monitoring these internal features effectively detects and explains failures, connecting failure signals to understandable shifts within the policy architecture.