How Neural Networks Speed up a Randomized Incremental Graph-based Motion Planner

Using graph based approaches save trajectories for manipulators can be planned fast. It is favorable to use techniques that allow to plan at least some motions from the beginning of the graph construction process, and that can be improved incrementally. We introduce an approach that fulllls the above requirements using random conngurations for graph construction (unless speciic tasks are given) in connguration space. Graph nodes serve as subgoals and graph edges as collision free sub-trajectories. We show the high performance of this approach with respect to preprocessing and trajectory generation time, as well as planning success in a realistic simulation of a real world manipulator task.