Multiple-Path Generation to Improve Autonomous Vehicle Planning

Abstract

Path planning is an integral part of ensuring autonomous vehicles become more safe and efficient in order to facilitate greater adoption of the technology and make it a viable option for an increased number of applications. While there are many documented approaches to path planning on roads, the challenge of path planning in an off-road environment is less studied and presents additional challenges including traversing an unknown and unstructured map environment. This can result in scenarios where previously unknown obstacles are discovered along a path a vehicle is traversing, forcing the vehicle to re-route and take a potentially inefficient route to its final destination. We present an extensible framework to mitigate this issue in which we generate multiple paths, select an efficient path for the vehicle with respect to its navigation to the final destination, and incentivize the vehicle to adhere to the selected path. We also implement this framework within the Nebula team’s ROS-based autonomous vehicle software stack for DARPA’s Racer challenge and compare its performance to the current implementation. Through testing performed in simulation on topologies of interest to the Nebula group, we find that the proposed framework results in a 37% increase in average speed and a 24% decrease in time to reach the final destination compared to the current implementation.