Learning Target-Oriented Push-Grasping Synergy in Clutter With Action Space Decoupling

We explore a method for grasping novel target objects through push-grasping synergy in a cluttered environment without using object detection and segmentation algorithms. The target information is represented by a color heightmap of the target object. The agent needs to implicitly find the corresponding object in the cluttered scene according to the target information and gives action instructions. We propose an action space decoupling framework to efficiently address the problem of policy learning in large state spaces, which predicts the grasping position and the grasping angle separately. Our system needs to learn two policy networks: position net and angle net. The position net infers the appropriate grasping position and the starting position of the pushing motion, and the angle net predicts the grasping angle based on the grasping position. In addition, the angle net is also a coordinator to determine whether the current state should perform the grasping action or the pushing action. A series of experiments show that the proposed method can quickly learn complex push-grasping coordinated policies in a cluttered environment and the task success rate and the grasping efficiency are greatly improved compared to the baseline methods and it has good generalization ability for novel objects. Furthermore, our method can be transferred to the real world without fine-tuning.