Hierarchical learning-based control system for agile locomotion of quadrupedal robots on discontinuous terrain

Abstract

This paper presents a hierarchical control framework that enables agile locomotion in a quadrupedal robot on discontinuous terrain. The first part introduces a hierarchical motor control strategy that supports an importable learning-based controller. This method helps the neural network trained in the simulation environment be easily used in real robots. The second part presents a technique for training a control policy with hierarchical reinforcement learning to overcome the stepping stone environment. To this end, the controller is hierarchically composed of a low-level controller trained for jumping locomotion and a high-level controller trained for jumping commands. A fixed gap test and a variable gap test are used for verification in a simulation. As a result, the hierarchical learning controller overcame a gap of up to 0.73 m in the fixed gap test. Also, in the variable gap test, the controller achieved a gap of 0~0.33 m with a 90% success rate and a gap of 0~0.5 m with a 70% success rate.