A Biped Robot Learning to Walk like Human by Reinforcement Learning

It's challenging to make a biped robot walk like a human. Many researches have been made such as gait planning, stable walking controller and so on and achieve great progress. Reinforcement learning methods are used in biped robots recently due to their powerful ability to deal with high-dimensional computing problem. However, it's hard to design good reward function to guide the robot to walk and behavior like human. This paper builds a biped robot model and presents a control framework of reinforcement learning based on Isaac Gym simulation platform. The designed reward function considers the velocity tracking, the symmetry of hip angle and leg lifting to simulate human motion. The training process only lasts for 2 hours from the very beginning. The results show that after training the biped robot has a good performance of velocity tracking and attitude control and shows good symmetries in joint angles especially in hip. The results also prove the designed reward function is effective and hopeful to be available on other applications.