Efficient and scalable planning systems for autonomous driving

Abstract

Autonomous driving is steadily transitioning from concept to reality. As the pursuit of fully autonomous vehicles intensifies, there is an escalating emphasis on the planning component of these systems, which currently represents a major barrier to achieving higher levels of autonomy. This thesis is devoted to building efficient and scalable planning systems for autonomous driving, from module-based to end-to-end solutions. The discussion begins with an examination of module-based planning systems, introducing a real-time motion planning framework that effectively navigates various planning constraints and investigates the motion forecasting problem from a novel perspective using self-supervised learning techniques. Building on the insights derived from these module-based systems, the thesis then transitions to the development of more scalable end-to-end planning systems. It explores several fundamental design choices to address closed-loop issues in learning-based planning and introduces a robust baseline model with effective longitudinal planning capabilities. Furthermore, this foundation supports the development of a multi-policy neural planner designed to foster more sophisticated and human-like planning behaviors by enhancing lateral maneuver learning. Synthesizing all previous developments, these components are consolidated into a comprehensive end-to-end planning system, featuring enhancements in model architecture, loss design, data augmentation, and training framework. The system demonstrates diverse and flexible human-like driving behaviors while adhering to stringent planning constraints. Notably, it matches the performance of state-of-the-art rule-based planners for the first time, marking a significant progression in the domain of practical end-to-end planning systems. The effectiveness of our system is validated through extensive open-loop and closed-loop evaluations.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Qifeng CHEN (Supervisor)