Dynamic motion policy learning for legged robot with single rigid body pretraining and curriculum

Abstract

This research proposes a novel control policy learning method for a quadruped robot using a Single Rigid Body model pre-training and a leg mass curriculum. This approach enables the smoother discovery of dynamic movements by performing pre-training in a simulation environment where the actual robot’s model is approximated as a Single Rigid Body. This helps avoid challenges such as leg collisions and foot slipping that may arise during dynamic movements. The learned control policy is then translated into a feasible control policy for the real robot by continuously retrieving leg mass through a leg mass curriculum method. By learning locomotions and dynamic jumping motions in the simulation environment, it has been demonstrated that pre-training in a Single Rigid Body model is effective in finding the desired dynamic movements. It has also been demonstrated that the leg mass curriculum can effectively adapt the control policies found in the Single Rigid Body model to the actual robot model.