Improving RRT with Context Sensitivity 15-780 Grad AI

The RRT algorithm represents an extreme in the design space of planning algorithms. It forsakes optimality and explores the space through randomness and a small bias towards goals. This is in marked contrast to other planning strategies such as visibility graphs or voronoi diagrams which carefully analyze and partition the space to find paths. This difference makes RRT very efficient even in higher dimensional spaces because its computational overhead is small in comparison to other algorithms which attempt to precisely characterize the world. However, this also means that RRT does not take advantage of information that it learns about the world while exploring. The central idea of our project is to introduce learning of the environment into the RRT algorithm for online use in the search and also for re-planning purposes. We focus our efforts on the extension length for two reasons. First, the RRT algorithm gets information about good and bad extension lengths from given points when it tests to see if a given extension fails. Second, the optimal extension length varies based on the obstacles surrounding a given point. The goal of this project is to extend the RRT planning and the ERRT re-planning algorithm to store and use information learned about good extension lengths while searching the space. Our hope is that we can realize significant improvements in how fast and often RRT-based planning finds the goal, without sacrificing its key principles. Our main goal is to explore a set of variations in the standard RRT (and ERRT) algorithms to try and offer a better algorithm for these heterogeneous worlds both in planning and re-planning scenarios. Some other papers describing extensions to the RRT algorithm which include [Lav98],[LL04],[JYLS05], and [YL09].