Robust Deep reinforcement learning against adversarial attack and random noise on quadruped legged robot body shape

Abstract

Legged robot control by deep reinforcement learning is vulnerable to adversarial attacks on robot body shapes. This study experimentally demonstrates that adversarial training using adversarial body shapes is robust against adversarial attacks and random noise on legged robot body shapes. In experiments, we evaluate the robustness of Unitree A1, a quadruped robot, in the MuJoCo environment in terms of reward. We search for the adversarial body shapes with lower rewards using the differential evolution method, and we create the random body shapes by adding random noises to the normal body shape. We train three policy networks using the normal, random, and adversarial body shapes, respectively. Experimental results reveal that adversarial training is robust to the normal, random, and adversarial body shapes. Similar results are known for image classification using deep learning, and this study demonstrates the effectiveness of adversarial training for robustness in robot control using deep reinforcement learning.