PhD Proposal: Modeling Human Factors and Traffic Dynamics for Generalization in Autonomous Driving

Self-driving cars have long been regarded as a futuristic technology. Thanks to advancements in deep learning and large-scale computing power, this technology is becoming closer to a likely reality by the day. While state-of-the-art (SOTA) autonomous driving can operate smoothly in clear, sunny, and low-density traffic scenarios, we have not yet achieved fully reliable autonomy under edge-case conditions, such as adverse weather or rare traffic events.
The primary challenge is that end-to-end SOTA autonomous systems often rely on large, powerful, and uninterpretable deep learning models. As a result, these models are incredibly data and compute intensive. Currently, the solution to address shortages in data is simply to scale, or invest in compute and collect as much data as possible. However, such approaches only address the tail distribution problem from a statistical standpoint. In reality, traffic is a highly constrained but complex problem for intelligent reasoning. Take, for example, the simple act of "driving forward": a simple heuristic is to follow a trajectory parallel to the curvature of the lane markings, while simply not colliding with the vehicle in front. Within the bounds of feasible driving, any particular driver has the ability to disrupt the entire traffic system. Thus, what makes modeling driving behavior challenging may not necessarily be learning the traffic rules and dynamics, but rather accounting for the natural corruptions and sensor noises, abstract and uninterpretable human factors, such as driver's personality, the unpredictable reaction, and individual behavior to rare traffic events.
These factors reveal a gap in driving research: the distinction between (1) model-based dynamics with (2) human factors in deep learning for autonomous driving. In this thesis, I propose to bridge this gap by modeling perceptual sensitivity, human factors, traffic flow, and vehicle kinematics with state-of-the-art deep learning methods for autonomous driving, specifically in perception and behavior simulation. By modeling both the rigidly-defined rules surrounding traffic dynamics and the loosely-defined notion of driving style, we expect to achieve better generalization to edge case scenarios, where rigidly-defined dynamics should still remain consistent.