Robot multi-action cooperative grasping strategy based on deep reinforcement learning

To address the problem of decreased grasping success rates in complex scenarios where there are obvious obstructions and overlapping objects, a multi-view grasping and pushing collaborative strategy based on Deep Reinforcement Learning was proposed. This strategy utilized 3D point cloud obtained from an RGB-D camera as a state input, and used two deep Q-network algorithms to fit the grasping and pushing strategies separately. Through the reasonable design of the reward function, the coordination between grasping and pushing was learnt to change the distribution of objects through pushing actions and provide favorable conditions for grasping. Furthermore, to address the slow training speed resulting from the complexity of action space, a normal mask-based action space optimization strategy was proposed by limiting the exploration space of grasping and pushing actions based on prior knowledge of estimated normal direction. Finally, experiments were conducted with UR5 robot in the real world to verify the effectiveness of the proposed cooperative grasping strategy. © 2024 CIMS. All rights reserved.