Learning of Leg Stiffness and Damping to Match Robot Dynamics

The spring-loaded inverted pendulum model, a simplified robot dynamics model, is regularly employed for the purpose of controlling the motion of legged robots. It has two key parameters to be determined, namely, spring stiffness and spring damping. For different robot systems, model parameters need to be set to match the robot’s dynamic characteristics. Nevertheless, it’s hard to match the model parameters with the robot’s dynamic characteristics through manual adjustment in engineering practice. Inspired by animals’ hierarchical and modular motion control features, this paper designs a hierarchical motion control framework for robot dynamics matching. The spring stiffness and damping of the SLIP model are learned through the deep reinforcement learning algorithm to match the control model with the robot’s dynamic characteristics. Finally, the performance of the hierarchical motion control framework is tested on a one-legged robot by using dynamics simulation software through comparative experiments. The simulation experimental results show that the robot with learned model parameters has a less mechanical cost of transport and better motion stability than the robot with fixed model parameters.