Deep Learning rooted Potential piloted RRT* for expeditious Path Planning

Abstract

Randomised sampling-based algorithms such as RRT and RRT* have widespread use in path planning, but they tend to take a considerable amount of time and space to converge towards the destination. RRT* with artificial potential field (RRT*-APF) is a novel solution to pilot the RRT* sampling towards the destination and away from the obstacles, thus leading to faster convergence. But the ideal potential function varies from one configuration space to another and different sections within a single configuration space as well. Finding the potential function for each section for every configuration space is a grueling task. In this paper, we divide the 2 dimensional configuration space into multiple regions and propose a deep learning based approach in the form of a custom feedforward neural network to tune the sensitive parameters, upon which the potential function depends. These parameters act as a heuristic and pilots the tree towards the destination, which has a substantial effect on both the rate of convergence and path length. Our algorithm, DL-P-RRT* has shown the ability to learn and emulate the shortest path and converges much faster than the current random sampling algorithms as well as deterministic path planning algorithms. So, this algorithm can be used effectively in environments where the path planner is called multiple times, which is typical to events such as Robo-Soccer.