Learning Locomotion For Legged Robots Based on Reinforcement Learning: A Survey

The legged robot can adapt to almost any kind of complex terrain and overcome all kinds of obstacles. So to this day, many people are working on using leg-based robots for complex locomotion tasks. It is tractable and difficult to achieve the agile locomotion of quadruped robots. Conventional controllers always need a lot of professional experience and lots of time to debug and tune the parameters. Deep reinforcement learning(DRL) can learn the effective skills from trails directly in practice, which holds the promising to overcome the limitation of the conventional controllers. Therefore, we have surveyed the current research working on learning locomotion skills via DRL techniques; and compare two commonly used DRL algorithms to learn the locomotion skills on a constructed simulation task.