AE-Reorient: Active Exploration Based Reorientation for Robotic Pick-and-Place.

Finding objects in dense clutter and placing them in specific poses play an important role in robot manipulation in fields like warehousing and logistics, and have a significant influence on the automation of these fields. However, most methods that perform well in simple clutter do not hold up well in dense clutter because of severe stacking and occlusion between objects, which prevents the target objects from grasping successfully. In this paper, we propose an interactive exploration framework called AE-Reorient based on reinforcement learning to let robots learn the interactive exploration autonomously. AE-Reorient enables robots to actively find the target objects and accomplish the pick-and-place task. Our AE-Reorient generates iterative and interactive actions to improve the visibility of the target objects which is an important factor affecting the success of grasping. Specifically, we first collect the scene information through an RGB-D camera. Then, using the point cloud reconstruction and pose estimation methods to obtain the height map of the clutter, through which we generate the grasp pose or predict the position. We apply interaction to change the state of the scene by pushing if the grasp is unsuccessful. The generated push action is applied to disrupt and recreate a new state of the clutter with the target objects more exposed, which greatly improves the probability of grasp success. Robots can successfully perform pick-and-place tasks in dense clutter by applying our proposed framework. We test our AE-Reorient in 130 dense clutters and it enhances the success rate and the visibility of the target object by 7.60% and 4.10% compared with the baseline method.