Learning Agile, Robust Locomotion Skills for Quadruped Robot

Locomotion control of legged robots has been a key topic in the field of robotic research. Due to the highly nonlinear dynamic characteristics and interaction with the environment at high frequencies, conventional controllers are too complicated to design for the sake of obtain robustness and adaptability. We avoid these problems by reinforcement learning (RL). Without consider the dynamics and contact problems, ours RL-based approach makes the robot understand itself and really learned how to locomotion, just like newborn quadrupedal do. Through the guidance of reward and Life-Long learning, the trained policy has extraordinary performance in agile and robust locomotion. As a result, our robot has strong robustness to the instantaneous impact or continuous drag during running, and adaptive the slippery ground which not encountered in training. The robot capable of reaching speed of 4m/s, and naturally, the gait of robot automatic transitions from walking-trot to flying-trot as the speed increases.