Title:Multi-Arm Bin-Picking in Real-Time: A Combined Task and Motion Planning Approach

Abstract:Automated bin-picking is a prerequisite for fully automated manufacturing and warehouses. To successfully pick an item from an unstructured bin the robot needs to first detect possible grasps for the objects, decide on the object to remove and consequently plan and execute a feasible trajectory to retrieve the chosen object. Over the last years significant progress has been made towards solving these problems. However, when multiple robot arms are cooperating the decision and planning problems become exponentially harder. We propose an integrated multi-arm bin-picking pipeline (IMAPIP), and demonstrate that it is able to reliably pick objects from a bin in real-time using multiple robot arms. IMAPIP solves the multi-arm bin-picking task first at high-level using a geometry-aware policy integrated in a combined task and motion planning framework. We then plan motions consistent with this policy using the BIT* algorithm on the motion planning level. We show that this integrated solution enables robot arm cooperation. In our experiments, we show the proposed geometry-aware policy outperforms a baseline by increasing bin-picking time by 28\% using two robot arms. The policy is robust to changes in the position of the bin and number of objects. We also show that IMAPIP to successfully scale up to four robot arms working in close proximity.