Learning Motion Planning Assumptions

Abstract

The performance of a motion planning algorithm is intrinsically linked with applications that respect the assumptions being made. However, the mapping of these assumptions to actual environments is not always transparent. For example, a gradient descent algorithm is capable of tackling a complex opti- mization problem if some assurance of absence of bad local minimas can be ensured - however detecting the local minimas beforehand is very challenging. The state of the art technique relies on an expert to analyze the application, deduce assumptions that the planner can leverage and subsequently make key design decisions. In this work, we make an attempt to learn a mapping from environments to specific planning assumptions. This paper presents a diverse ensemble of planners that exploit very different aspects of the planning problem. A classifier is then trained to approximate the mapping from environment to performance difference between a pair of planners. Preliminary results hints at the role played by convexity, whilst also demonstrating the difficulty of the classification task at hand.