Distributed Motion Planning for Industrial Random Bin Picking

The task of bin picking is to automatically unload objects from a container using a robotic manipulator. A widely used solution is to organize the objects into a predictable pattern, e.g., a workpiece carrier, in order to simplify all integral subtasks like object recognition, motion planning and grasping. In such a case, motion planning can even be solved offline as it is ensured that the objects are always at the same positions. However, there is a growing demand for non-structured bin picking, where the objects can be placed randomly in the bins. This arises from recent trends of transforming classical factories into smart production facilities allowing small lot sizes at the efficiency of mass production. Due to unknown positions of the objects in the non-structured bin picking scenario, trajectories for the manipulator cannot be precomputed, but they have to be computed online. Sampling-based motion planning methods like Rapidly Exploring Random Tree (RRT) can be used to plan the trajectories. In this paper, we propose a modification of RRT for distributed motion planning aiming to reduce the runtime. The planning task is first simplified by computing several guiding waypoints. The waypoints are distributed to a set of planners running in parallel and each planner computes a short trajectory between two given waypoints. Connecting the waypoints is easier than solving the original task, therefore each planner runs fast. In comparison to other parallel motion planning techniques, the proposed approach does not require any communication among the computational nodes, which is more suitable for cloud-based computing. The proposed work has been verified both in simulation and on a prototype of a bin picking system. (C) 2018 The Authors. Published by Elsevier B.V.