Stable Imitation of Multigait and Bipedal Motions for Quadrupedal Robots Over Uneven Terrains

Imitation learning has proven effective for quadrupedal robot control by demonstrating capabilities such as precise motion tracking and executing complex maneuvers like backflips and rapid trotting while achieving faster convergence compared to traditional reinforcement learning. Despite these advancements, challenges remain. Existing methods struggle with stability, generalization, and real‐world adaptation, particularly in the task of dealing with disturbances and varying terrains. To overcome these obstacles, we propose a novel framework that integrates a long short‐term memory (LSTM) data regularization module, model‐based stability reward functions, and an interpolation‐based dataset augmentation technique. This comprehensive approach refines motion data, enhances stability, and improves the robot's ability to handle disturbances. Our method supports a variety of data inputs from different robots, simulators, and controllers, and enables the robot to perform diverse motions, including tripod walking, trotting, pacing, bounding, pronking, and even bipedal walking, while maintaining stability over a variety of uneven terrains. By implementing variable frequency imitation learning, we significantly improve the generalization of learned behaviors across different gaits and speeds. By involving the dynamics modeled reward, we enhance the robot's adaptability and robustness in dynamic environments. These contributions mark an important step toward deploying quadrupedal robot imitation learning policies in more unpredictable and challenging real‐world scenarios.