Motion planning is an essential element of autonomous driving systems. By integrating perception, trajectory prediction, map data, and localization outcomes, it is anticipated to enable the safe and efficient navigation of autonomous vehicles (AVs) through complex dynamic environments. This necessitates the development of planning systems and algorithms capable of effectively managing complex interactions among traffic agents, while adhering to various constraints such as traffic regulations and vehicle dynamics. However, the concept of interactions is rarely explored in classical planning methods. Many of these methods overlook interactions in their methodologies and solely pursue collision-free planning results, which ultimately limits the planning performance.

In this thesis, I aim...[ Read more ]

Motion planning is an essential element of autonomous driving systems. By integrating perception, trajectory prediction, map data, and localization outcomes, it is anticipated to enable the safe and efficient navigation of autonomous vehicles (AVs) through complex dynamic environments. This necessitates the development of planning systems and algorithms capable of effectively managing complex interactions among traffic agents, while adhering to various constraints such as traffic regulations and vehicle dynamics. However, the concept of interactions is rarely explored in classical planning methods. Many of these methods overlook interactions in their methodologies and solely pursue collision-free planning results, which ultimately limits the planning performance.

In this thesis, I aim to design safe and efficient motion planning methods for autonomous driving, taking into account traffic interactions and various constraints. I start by presenting the design of a campus autonomous driving system, which facilitates commuting in a campus environment and highlights the challenges encountered. This provides inspiration for research on motion planning methods. Following this, I explore questions concerning interactions and their significance in planning, focusing on two primary issues. Question 1: Is there a unified interaction model that can explain different interaction situations among traffic agents? Question 2: Traffic agents play different roles during interaction, how can we model it in a spatio-temporal planning framework? For Question 1, our focus is on modeling driving patterns of traffic agents around the minimum conflict space: a single conflict point. We introduce the interaction point model to uniformly describe various forms of interactions in traffic and propose an efficient planning framework to utilize this model, demonstrating its effectiveness. For Question 2, we propose a prediction-based interactive planning framework, where traffic agents’ different roles and behavior patterns are integrated into a traditional spatio-temporal (s-t) search method. To validate these proposals, a series of experiments has been conducted to illustrate how the proposed models and frameworks can improve the performance of autonomous driving.