Deep Reinforcement Learning with PSO Algorithm in Agile Dodge Actions of Quadruped Robot

This research aims to solve the highly dynamic obstacle avoidance task using Deep Reinforcement Learning (DRL) for quadruped robots. The proposed method integrates collision-free commands generation and reward design to enhance the dodge action' learning ability of DRL policy. By deploying parabolic collision point predict and Particle Swarm Optimization (PSO) algorithm to generate rough move commands, then we use DRL policy to actively follow the command while being passively affected by collision reward. Our method enables efficient obstacle avoidance in dynamic ball shooting scenarios. Experiments show that our method has high training efficiency and a success rate of over 90% when the shooting speed of the ball is $5\sim 15\mathrm{m}/\mathrm{s}$ and the reaction time is limited to 800ms.