Learning robust quadrupedal locomotion under disturbances via reinforcement learning with an autonomous evolutionary mechanism

Maintaining stable locomotion in complex environments is critical for the practical application of quadrupedal robots. Existing learning-based motion control methods often rely on artificially designed simulation parameters and training schedules. However, such methods tend to fail to meet the dynamic needs of the robot at different stages of training, resulting in insufficient robustness of the learned policy when facing disturbances. This study proposes an autonomous evolutionary mechanism that enables the robot to dynamically adjust its training process through comprehensive self-assessment, ensuring an optimal training environment and curriculum. Furthermore, the mechanism incorporates our proposed adaptive rewards and disturbances, providing a relaxed training environment in the early stages to encourage the robot to explore the policy space. As training progresses, it gradually increases control constraints and disturbances to ensure the robot remains robust under more challenging conditions. We evaluate the robustness of our method through both simulations and real-world deployments, conducting experiments on various quadrupedal locomotion tasks and more challenging scenarios.