Deep reinforcement learning based moving object grasping

Traditional grasping methods for locating unpredictable positions of moving objects under an unstructured environment cannot achieve good performance. This paper studies the utilization of deep reinforcement learning (DRL) with a Kinect depth sensor to resolve this challenging problem. The proposed grasping system integrates the DRL algorithm, Soft Actor-Critic, and object detection techniques to implement an approaching-tracking grasping scheme. Considering the state and action space for the high-degree-of-freedom manipulator, we employ an improved Soft-Actor-Critic algorithm to speed up the learning process. The proposed system can decouple object detection from the DRL control, which allows us to generalize the framework from a simulation environment to a real robot. Experimental results demonstrate that the developed system can autonomously grasp a moving object with different moving trajectories. (c) 2021 Elsevier Inc. All rights reserved.

Automated summary

This paper investigates the use of deep reinforcement learning (DRL) with a Kinect depth sensor to tackle the problem of locating unpredictable positions of moving objects under unstructured environments. The proposed grasping system integrates DRL algorithm, Soft Actor-Critic, and object detection techniques to achieve autonomous grasping of moving objects.