Towards Robotic Picking of Targets with Background Distractors using Deep Reinforcement Learning

Robotic manipulation using vision-based learning algorithms has benefits many tasks where the robot needs to interact with a desired target. However, when a desired target is placed with other objects, there remains several technical challenges. Firstly, the target might be occluded by other objects in terms of the camera field of view. Secondly, prioritizing the selection of the objects changes the polices and rewards of the learning algorithm which might deteriorate the successful rates. Thirdly, the occlusion by other objects increases the difficulty of manipulation as it requires complex synergies between pushing and grasping. Therefore, we propose a value-based deep reinforcement learning with Mask R-CNN to address the issues of robotic manipulation for multiple object. In the proposed three rewards are proposed, namely: 1) success of grasping the desired target; 2) removal of the distractors; and 3) effective pushes. This method aims to enable the manipulation to grasp a desired target among distractors. Simulation has been conducted to demonstrate the effectiveness of the proposed method. The results show that the robot is able to effectively pick the desired targets out despite the physical and optical occlusion by other objects. All tasks are completed with 100% successful rate within 30 episodes.