Robust Control of Quadruped Robots using Reinforcement Learning and Depth Completion Network

Achieving robust control of quadruped robots in dynamic and complex terrains is still a challenging task. Although reinforcement learning-based control strategies have made great progress in simulation and reality, motion control of quadruped robots based on depth cameras is still worth studying. In this paper, we proposed a reinforcement learning framework that uses visual perception and proprioception as inputs to train a quadruped robot for robust control, and designed a new depth completion network called DRI-Net for completing missing depth visual information. The proposed network is based on fusing the depth features from depth maps with the contour features from RGB images and enabled the quadruped robot to accurately perceive external environment. Our main aim is to improve the decision making procedure of reinforcement learning controller and final evaluations in dynamic multi-obstacle environments demonstrated that our method outperformed the baselines in terms of multiple metrics.