Motion Acquisition of Vertical Jumping by a Bio-inspired Legged Robot via Deep Reinforcement Learning

Achieving animal-like agility in legged robots is one of the challenging tasks. Motions such as those generated in a simplified or ideal environment to reduce the complexity of the model cannot adapt to changes in the environment specially in the case of dynamic motions. Deep reinforcement learning (DRL) has been attracting attention as an approach to generalize and robustify robot motions. In this paper, we focused on DRL as an approach to achieve dynamic motions for bio-inspired legged robots, and used it to learn a vertical jumping motion, which is one of the dynamic motions. By training the policy while randomizing the values of the robot's initial posture and environmental parameters, we acquired the general controller. The controller enabled the robot to jump in various situations without having to rerun the optimization routine whenever those values change, as in the optimization approach. The controller also enabled the robot to utilize the dynamical interference of the body to achieve high jumps.