Speed tracking and natural gait control with the aid of adversarial prior data for legged robots

This study proposes a new reinforcement learning-based method to enhance the motion control of quadruped robots by combining adversarial prior data, terrain training curricula, and specialized reward function design. Through training with adversarial prior data (AMP), the robot successfully learns stable and natural gaits, enabling it to traverse various terrains with higher adaptability and efficiency. The introduction of the terrain training curriculum allows the robot to progressively adapt to increasingly complex terrains, accelerating the training process and enabling it to quickly pass through the highest difficulty levels of terrain. Additionally, the reward and punishment mechanism for low-speed commands enhances the robot's sensitivity to low-speed operations while maintaining agility during high-speed movement. This algorithm enables the robot to quickly complete terrain course training, accurately track speed commands, and naturally and flexibly navigate through complex terrains.