Improvement of fault tolerance of quadruped robots by detecting correlation anomalies in sensor signals

Abstract

Autonomous robots that rely on sensors for operation require fail-soft strategies to continue tasks despite partial sensor failures. We propose a sensor anomaly detection method that monitors changes in sensor data correlations. Our method eliminates the need for pre-defined programming to determine abnormal states for each individual sensor. Furthermore, real-time anomaly detection is possible through sparse structure learning. In the experiment, we evaluated this method on a quadruped robot in a simulated environment. We perturbed the sensor readings by adding two types of large or small noise at one of the robot’s leg joints. When an anomaly was detected, the robot estimates the actual value of the noisy joint using a pre-trained multiple regression model. With our proposed anomaly detection method, the robot successfully completed the walking task in most trials. Specifically, without anomaly detection, adding a large noise to any of the twelve joints resulted in a 0 % success rate. However, with anomaly detection, the success rate improved to over 89 % in seven of the twelve joints.