A Stability Training Method of Legged Robots Based on Training Platforms and Reinforcement Learning with Its Simulation and Experiment.

This paper continues the proposed idea of stability training for legged robots with any number of legs and any size on a motion platform and introduces the concept of a learning-based controller, the global self-stabilizer, to obtain a self-stabilization capability in robots. The overall structure of the global self-stabilizer is divided into three modules: action selection, adjustment calculation and joint motion mapping, with corresponding learning algorithms proposed for each module. Taking the human-sized biped robot, GoRoBoT-II, as an example, simulations and experiments in three kinds of motions were performed to validate the feasibility of the proposed idea. A well-designed training platform was used to perform composite random amplitude-limited disturbances, such as the sagittal and lateral tilt perturbations (±25°) and impact perturbations (0.47 times the robot gravity). The results show that the proposed global self-stabilizer converges after training and can dynamically combine actions according to the system state. Compared with the controllers used to generate the training data, the trained global self-stabilizer increases the success rate of stability verification simulations and experiments by more than 20% and 15%, respectively.