Model-free End-to-end Learning of Agile Quadrupedal Locomotion over Challenging Terrain

In recent years, many studies using reinforcement learning and complex modules to make quadruped robots traverse complex terrain rely on complex modules and are difficult to train. How to obtain a reinforcement learning paradigm with simple structure and easy training that can traverse complex terrain has become a challenging problem. Most of the previous work relied on hand-crafted foot trajectory or other modules related to gait, and there are seldom methods to get an end-to-end reinforcement learning controller without a model and only proprioceptive observations yet. We adopt the method of combining reinforcement learning with inverse kinematics to generate foot trajectories in the model-free mode, and finally obtained the standard and smooth trot gait only depends on the robot's proprioceptive observations, and robot can also achieve steady walking by curriculum learning in the case of with no camera obtaining terrain's information.