Adversarial Body Shape Search for Walking Robots

Abstract

In the development of walking robots, it is important to design robots that are robust against small changes in body shape caused by temperature changes, foreign objects, etc. We propose the use of adversarial attacks as a method to search for small body changes that significantly reduce the reward for walking robots trained by reinforcement learning. While adversarial attacks generally seek input perturbations that reduce the loss of deep learning, the proposed method seeks body shape perturbations that reduce the reward of reinforcement learning. The computation of the gradient with respect to the body shape changes uses differential evolution, since efficient optimization algorithms such as back propagation method are not available. In the experiment, the proposed method was used to search for small changes in the length and thickness of each part of the robot. We found common body shape changes in three types of walking robots with completely different shapes.