Robot Arm Grasping based on Multi-threaded PPO Reinforcement Learning Algorithm

In recent years, the problem of robotic arm grasping has been the research focus in the field of robotic arm. This thesis carries out research on robotic arm grasping based on the multi-threaded PPO (proximal policy optimization) algorithm, and establishes a robotic arm grasping system model by using the multi-threaded PPO algorithm. The original PPO algorithm is difficult to select the appropriate policy, and the difference between the old and the new policy varies a lot during the training process, which affects the training effect, whereas the current PPO utilizes the CLIP function to reduce the gradient calculation process. Therefore, the reward function is designed based on the current PPO algorithm, and the method of multi-threaded parallel computing, dominant value regularization and reward scaling is proposed, which further improves the training stability and scalability of PPO, and accelerates the sampling efficiency and convergence of the algorithm.