For wide and flat objects, such as books and package boxes, the robot arm cannot grasp directly because of the limit of the gripper width. Therefore, pre-grasp manipulation is needed to make them graspable, e.g., pushing it to the table side and grasp from the side. In this paper, we propose a novel method to learn the pre-grasp manipulation based on deep reinforcement learning. Instead of directly feeding the raw images, we generate the binary masks as states according to the shape and current position of the object and table. Meanwhile, we use adaptive resets to accelerate the training process, i.e., adaptively reset the initial states to demonstration states. After training on a large scale of automatically generated objects and tables of different shapes, our proposed method achieves a 97.6% success rate on novel objects.
Demo for seen objects and tables
Demo for unseen objects and tables
Pushing: a RL policy learned by our proposed method (after making the object graspable, there is still a duration to verify it won't fall)
Grasping: a motion planner according to the desired grasping point (the red point in the video)