Experience Based Motion Planning using Scene Clustering

Abstract

Sampling based motion planners are widely used in the industry for generating fast and optimum solutions. This thesis investigates ways to improve planning time and quality of generated paths by effectively using the experience from past problems to optimize solution generation for later plans. It is assumed that the robot's environment affects each planning run and planning experience can be reused for similar scenes. 3D scene geometry is parsed into a graph model and custom kernel functions are developed to compare the scenes. A clustering based approach is then followed to extract similar scenes. Solutions of the planning problems from the similar scenes are then given to the planner which uses it to generate a new plan for a similar problem. Experiments have shown that this approach considerably improves the planning duration and quality of the generated path.