Reinforcement Learning-Based Collaborative Manipulation Method for Robotic Arms

With the development of artificial intelligence technology, robotic grasping has become increasingly widely applied in industrial automation and intelligent service robots. This paper proposes a solution for robotic grasping tasks based on reinforcement learning. By designing two core actions (pushing and grasping) and optimizing the reward mechanism, this approach significantly improves the operational efficiency of robotic arms. Each action is trained by a neural network with identical structure. In the reward design, a mixed dispersion coefficient is introduced to evaluate the benefits of scene changes. Different reward values are set based on whether objects are pushed out of the operable area. Additionally, extra constraints are added during network training to prevent robotic arms from pushing objects out of the executable range, further optimizing the reward allocation strategy. Simulation results demonstrate the effectiveness of the proposed method.