An Object Recognition Grasping Approach Using Proximal Policy Optimization With YOLOv5

Aiming at the problems of traditional grasping methods for mobile manipulators, such as single application scenarios, low accuracy, and complex grasping tasks, this paper proposes an object recognition grasping approach using Proximal Policy Optimization (PPO) with You Only Look Once v5 (YOLOv5), which combines a vision recognition algorithm with a deep reinforcement learning algorithm to achieve object recognition grasping. First, YOLOv5 is adopted to identify the object and obtain the location information. Second, the PPO algorithm is used for object grasping to obtain the grasping strategy. Third, the PPO algorithm is compared with the Soft Actor-Critic (SAC) and Trust Region Policy Optimization (TRPO) algorithms in batches 16 and 128, respectively. The average reward training results of the PPO, SAC, and TRPO algorithms are obtained in our work. Experimental results show that the proposed method, in terms of object recognition speed, outperforms the original YOLOv4 model. The YOLOv5 model achieves 96% precision on our own built recognition dataset, which has higher detection precision and lower hardware requirements than the YOLOv4 model. Our proposed method outperforms SAC and TRPO algorithms in object grasping, and the average reward of the PPO algorithm is improved by 93.3% and 41% compared to SAC and TRPO algorithms, respectively. Finally, through the comparison of ablation experiments, our method has the highest accuracy and mean average precision (mAP)@0.5 value of 92.3%. We demonstrate in actual physical experiments that the grasping success rate under our proposed approach reaches 100%, providing a new research strategy for object grasping by the robot manipulator.