/High Dimensional Planning and Learning for Off-Road Driving

High Dimensional Planning and Learning for Off-Road Driving

Guan-Horng Liu
Master's Thesis, Tech. Report, CMU-RI-TR-37, Robotics Institute, Carnegie Mellon University, June, 2017

Download Publication (PDF)

Copyright notice: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author’s copyright. These works may not be reposted without the explicit permission of the copyright holder.

Abstract

This thesis explores both traditional motion planning and end-to-end learning algorithms in the off-road settings. We summarize the main contributions as 1) propose an RRT-based local planner for high-speed maneuvering, 2) derive a novel stochastic regularization technique that robustifies end-to-end learning in the spirit of sensor fusion, and 3) traversability analysis of the unstructured terrain using deep inverse reinforcement learning (DIRL) algorithms. We first propose a sample-based local planner that is modified to solve a minimal traveling-time trajectory problem subjected to a data-driven vehicle model in high dimensional state space. The planner is implemented on a full-size All-Terrain Vehicle (ATV), and the experimental results show that it can successfully avoid obstacles on a turnpike with the vehicle velocity up to the maximum operating speed. Secondly, we also propose a stochastic regularization technique, called Sensor Dropout, that promotes an effective fusing of information for end-to-end multimodal sensor policies. Through empirical testing on a physical-based racing car simulator called TORCS, we demonstrate that our proposed policies can operate with minimal performance drop in noisy environments, and remain functional even in the face of a sensor subset failure. Lastly, we investigate into the DIRL algorithms that infer the traversability of the unstructured terrain by leveraging a large volume of human demonstration data collected on the field. We propose several modifications to overcome issues that arise from DIRL training, such as sparse gradients and ambiguity of the demonstration optimality. The framework is tested on the full-size ATV.

BibTeX Reference
@mastersthesis{Liu-2017-26362,
author = {Guan-Horng Liu},
title = {High Dimensional Planning and Learning for Off-Road Driving},
year = {2017},
month = {June},
school = {Carnegie Mellon University},
address = {Pittsburgh, PA},
number = {CMU-RI-TR-37},
keywords = {planning, deep reinforcement learning, deep inverse reinforcement learning},
}
2017-09-13T10:38:03+00:00