Learning to Switch Gait Randomly and Continuously of Bipedal Robots via Stage-Wise Reward Shaping

Bipedal robots capable of continuously and randomly switching gaits acrossvarious terrains are of great significance. This paper proposes a stage-wise reward shaping method that enables a single policy to perform random switching between jumping and walking locomotion modes. Dedicated reward functions are designed for the jumping and walking locomotion, respectively, and during training, the locomotion mode is randomly switched at fixed intervals. For the complex jumping task, we further divide it into four consecutive sub-stages, each shaped with a different reward function. In the walking mode, omnidirectional velocity command tracking is used to train the robot, enabling omnidirectional locomotion at various speeds. In addition, a long short-term memory (LSTM) network is integrated into the Actor-Critic framework to help the model utilize historical information and make better decisions. Simulation results demonstrate that the proposed method enables the bipedal robot to switch between jumping and walking gaits smoothly and naturally in arbitrary order, allowing it to traverse complex terrains by switching gait modes.