Animal-Like Eye Vision Assisted Locomotion of a Quadruped Based on Reinforcement Learning

Abstract

The legged robot is only satisfied with walking on flat ground, it obviously does not take advantage of its locomotion performance, especially in the real environment where the robot may encounter various complex terrains. It is a highly nonlinear system for the quadruped robot, so it is very hard to model the dynamics accurately and achieve high performance locomotion control. In recent years, with the emerging of reinforcement learning, there are more possibilities to improve the locomotion ability of legged robots. To address the problem of how to improve the ability of a quadruped robot to negotiate complex terrains, this paper proposes a method to provide an animal-like eye for a quadruped robot to obtain a control strategy based on reinforcement learning. The method uses only the terrain information in front of the quadruped robot as the input state of the robot, and uses a curriculum training method to make the quadruped robot negotiate complex terrains such as the stairway terrain and the gap terrain constructed in the simulation environment smoothly. Compared with the motion strategy based on proprioception only, the vision-assisted motion strategy is safer and smoother, and we utilize less and simpler visual information than other methods based on visual information.