Robust Quadruped Jumping on Sloped Terrains using Reinforcement-Learning-Augmented Model Predictive Control

Abstract

Legged robots can navigate in much more complex environments than wheeled counterparts due to the added mobility of the limbs. Although much research has been done on walking and running control, real-time dynamic jumping of these robots on 3D sloped terrains remains underexplored. In this work, we aim to control a quadruped robot to jump in real time using a hierarchical framework, consisting of a Model Predictive Control (MPC) policy coupled with Reinforcement Learning (RL) modifier that offsets the torques of the MPC. Our method showed a success rate of 93.3% and position tracking error of 0.3157 ± 0.00565 m, as compared to 69.8% and 0.4167 ± 0.00793 m respectively for pure MPC. This approach will enable legged robots to jump across complex terrains, paving the way for the deployment of these robots in challenging environments such as disaster response, construction, or forestry.