Using neural networks for heuristic grasp planning in random bin picking

The fast determination of collision-free grasps is a key aspect in random bin picking. Heuristic search algorithms provide a feasible solution to this problem, using statistical data on the likelihood of finding a valid solution on elements with certain parameters. In this paper, we propose the use of several neural networks in such algorithms to accelerate the search while preserving the reliability. This is done by training the neural networks on the heuristic search trees of previous situations and using the output of these neural networks as part of the heuristic function. Finally, the effect of these neural networks is experimentally analyzed with sensor data from a working bin picking system with an industrial dual arm robot and it is shown that the calculation time in this setup is reduced by up to 45%.