Fast Adaptation Dynamics Model for Robot’s Damage Recovery

Abstract

In the process of operating, robots will inevitably encounter damage due to external or internal factors, such as motors blockage. For the legged robot, when the motors of joints are failing, if other motors still act according to the original instructions, it will cause the robot to deviate from the predetermined trajectory, which is unacceptable for legged robots. Inspired by the fact that the model trained by supervised learning on the training set can be generalized to the testing set, our goal is to obtain a dynamic model that can be generalized to all kinds of motor damage situations. It can predict what state will be reached in the next step when an action is applied in the current state. With this dynamics model, we use the Monte Carlo particles to optimize the feasible actions in a model predictive control (MPC) fashion and achieve the expected goal (such as making the robot walk in a straight line). The comparison experiment adopt two meta-learning model and vanilla dynamics model approaches, the results show that the proposed method is superior to the three baselines, which proves the effectiveness of the proposed method.