Active Affordance Exploration For Robot Grasping

Robotic grasp in complicated un-structured warehouse environments is still a challenging task and attracts lots of attentions from robot vision and machine learning communities. A popular strategy is to directly detect the graspable region for specific end-effector such as suction cup, two-fingered gripper or multi-fingered hand. However, those work usually depends on the accurate object detection and precise pose estimation. Very recently, affordance map which describes the action possibilities that an environment can offer, begins to be used for grasp tasks. But it often fails in cluttered environments and degrades the efficiency of warehouse automation. In this paper, we establish an active exploration framework for robot grasp and design a deep reinforcement learning method. To verify the effectiveness, we develop a new composite hand which combines the suction cup and fingers and the experimental validations on robotic grasp tasks show the advantages of the active exploration method. This novel method significantly improves the grasp efficiency of the warehouse manipulators.