Predicting Partial Paths from Planning Problem Parameters

Many robot motion planning problems can be described as a combination of motion through relatively sparsely filled regions of configuration space and motion through tighter passages. Sample-based planners perform very effectively everywhere but in the tight passages. In this paper, we provide a method for parametrically describing workspace arrangements that are difficult for planners, and then learning a function that proposes partial paths through them as a function of the parameters. These suggested partial paths are then used to significantly speed up planning for new problems.