Self-Healing Action Verification against Instruction Tampering and Sensor Failures in Humanoid Robots

Abstract

As humanoid robots are increasingly used in complex real-world environments, reliable verification of robot actions becomes more important. Action verification should not only check whether the executed action matches the intended instruction, but also keep reliable performance when motor-sensor observations are affected by failures. Traditional instruction authentication mainly protects the instruction itself and cannot directly verify the final physical action of the robot. In addition, the sensing information used to observe robot execution may also become unreliable. In this paper, we propose a self-healing motor-based action verification framework that combines explainable sensor selection, temporal masked reconstruction, and multiscale action recognition. First, Local Interpretable Model-Agnostic Explanations (LIME) are used to identify motor encoders that provide important information for action discrimination. Then, the selected encoder signals are processed by a temporal masked reconstruction model to recover unreliable joint observations based on temporal continuity and cross-joint coordination. After that, a multiscale gallery recognizer with zero trainable classifier parameters is used to infer the physical action executed by the robot. The inferred action is finally compared with the trusted intended instruction to detect instruction-execution inconsistency. The proposed framework is evaluated on a Booster Robotics K1 Education humanoid robot under multiple sensor-failure modes and different corruption levels, and is compared with several existing methods. The proposed method achieves average accuracies of 98.4% and 94.8% for single-action and compound-action recognition, respectively. Under different sensor-failure levels, the proposed self-healing method achieves an average recognition accuracy of 92.6%.