Behavior Generation Approach for Quadruped Robots Based on 3D Action Design and Proximal Policy Optimization

To improve quadrupedal robots' movement diversity and capacity to learn a wide range of behavioral skills, a quadrupedal robot behavior generation method based on 3D animation software and imitation learning is proposed. First, 3D animation software is used to provide expert experience of feasible control strategies for specific task skills; then, a motion skill optimization method based on imitation learning algorithm is designed to achieve the optimization of the designed actions. Experiments conducted on simulation and physical prototypes demonstrate that the Panda5 quadruped robot can perform extremely dynamic and agile behaviors like hopping in place and trot in place, verifying the efficacy of the method provided in this work.