Learning Graphic Convolutional Network Based Quadruped Locomotion on Complex Terrain

The similarity between the body structures of legged robots and those of humans and animals endows these robots with the capability to navigate rugged terrain. However, designing controllers for locomotion on various terrains involves complex terrain perception and dynamic modeling. Deep reinforcement learning methods can efficiently integrate external visual perception to create versatile controllers for robots. Nevertheless, in real-world walking tasks, certain lighting and terrain conditions can render visual perception unreliable, compromising the stability of the controller. In the absence of visual perception, proprioception, including body orientation and joint state information, proves more durable but requires effective utilization to achieve stable locomotion on complex terrains. Typically, learned controllers employ multilayer perceptrons to process proprioceptive data, often overlooking the interactions between joints and between joints and the body. This work proposes a control framework based on graph convolutional networks, which incorporates the morphological structure of the robot to more effectively organize proprioceptive information. This approach enables robust locomotion control on complex terrains without relying on external perception.