Learning to Walk on Low Friction Terrain by Reinforcement Learning

Locomotion control of legged robots has been an important topic in the field of robotic research. Its practical application in complex terrain environment is the focus and difficulty of research. In order to overcome the nonlinear dynamic characteristics of the legged robots walking on low friction terrain and obtain robustness, adaptability, the traditional controller design is too complex. We solve these problems by Reinforcement Learning (RL). Firstly, we make robot learn to walk on ordinary terrain by RL. And then with the usage of curriculum training, our robot can maintain balance on low friction terrain and adapt to terrains with the different friction coefficients. At last, we conduct a series of physical experiments on UnitreeA1 to verify the effectiveness of the algorithm.