Target Acquisition by Reinforcement Learning-Based Bin Tilting with a Robotic Arm

Target acquisition is one hot topic in the field of robotic manipulation. This task becomes more difficult especially when the target is fully non-visible that the target may be fully-occluded by other objects. How to acquire the target efficiently is one of the key issues. The existing methods leverage a series of pushing and grasping to acquire the target, the time cost, however, may have large variations that is influenced by the number of objects and the clutterness around the target. In this work, we propose a novel method which is not to re-arrange the objects surrounding the target but to tilt the container, and a deep reinforcement learning model to learn the tilt policies based on image data. Compared with other methods, the proposed method makes a significant change in the distribution of locations of the objects inside the bin after each tilting action so that to achieve efficient target acquisition. The final simulation experiments demonstrate the effectiveness and efficiency of our proposed method with a success rate of up to 92%.