Teaching a Robot Dog New Tricks: Diverse Quadruped Skills via Combined Reinforcement and Imitation Learning with Adversarial Task Selection

This research introduces a three-stage method that trains a single policy to perform distinct quadruped tasks such as walking, digging, and hopping, and composes them into novel behaviors like crawling. The approach combines reinforcement learning and imitation learning with adversarial task selection, achieving better motion quality and task tracking than baselines. The method is validated on a real-world Unitree B1 quadruped robot.