Computed-torque control of a simulated bipedal robot with locomotion by reinforcement learning

This paper presents the development of a hybrid control for an Atlas humanoid robot moving forward in a static locomotion regime. The experiments are carried out in the Gazebo simulation environment. The developed system consists of the modeling of the mechanics of the robot, including the dynamic equations that allow controlling the joints by computed-torque control. Locomotion is planned by agents through the approximate Q-Learning algorithm. A reduced simulated environment is used in the training stage, providing the agents with prior knowledge before the application in a real environment. The proposed approach results in an effective training in few interactions, produces good results and ensures the integrity of the robot.