Environmental Features Assessment Network Aided Deep Reinforcement Learning for Quadrupedal Locomotion in Tough Terrain

Quadrupedal robots have more powerful locomotion capabilities compared to other types of mobile robots. However, the complexity of their controllers and the need to adapt to a wide variety of terrains poses a huge difficulty in the design of the controllers. Recently, deep reinforcement learning (DRL) has been widely used in quadruped robot controller design but still relies on complex and reliable sensing frameworks. In this paper, we propose a motion controller training framework. The framework includes an Environmental Feature Assessment Network (EFANet) to guide the action output of the policy network, and an asymmetric actor-critic structure to help the policy network infer terrain features. We also introduce a strategic trajectory generator into our framework to prevent the robot from generating abnormal gaits. The proposed framework for learning quadrupedal locomotion allows quadrupedal robots to traverse challenging terrains with limited sensors. Finally, the proposed approach is tested and evaluated in a simulation environment using the A1 quadruped robot model.