Robotic Grasping Point Recognition Model based on Reinforcement Learning

Robotic grasping has always been a hot topic in the field of control. With the rise of intelligent control methods represented by deep learning, the research focus of this topic has changed to the strategy of self-learning. Since the existing two-stage robotic arm grasping method requires recalibration of the grasping system when the shape and position of the object to be grasped changes, this is time-consuming and user-unfriendly. Therefore, this paper proposes a novel grasping scheme, which combines the grasping problem of the robot arm with the two-stage object detection of deep learning to realize that the robot arm can find the object's location and perform the grasping task. We introduce the attention mechanism into the grasping system, combine the soft attention mechanism with reinforcement learning for object grasping, and use the hard attention mechanism as a recognizer to identify whether the grasping frame is grabable. Further, we use reinforcement learning to train the agent to perform the correct action and generate a reasonable grasp frame, which is finally fed into the recognizer to obtain the final required grasp position and sent back to the robotic arm to perform the grasp action. Finally, we apply the proposed two-stage object detection model to daily supplies to evaluate the model performance. Compared with the traditional object detection models, the architecture proposed in this paper performs better object detection.