A system-theoretic approach to modeling driver behavior in shared-autonomy vehicles

Abstract

Shared autonomous vehicles pair human drivers with autonomous controllers. Shared autonomy aims to pair human intuition and adaptability with autonomous precision. Over the past decade, an estimated 98 million vehicles in the United States have operated a primitive form of shared autonomy called advanced driver assistance systems or ``ADAS." However, while ADAS reduces accidents, current shared autonomy methods generally treat humans as providing perturbations, rather than as sources of valuable information. This thesis aims to fill this gap by developing methods to use human-provided information--specifically, a suggested steering command--to improve the state estimation and control of shared autonomous vehicles. The first contribution introduced by this thesis is developing the human-as-advisor archetype, a new method for using human input in shared autonomy. Shared autonomous vehicles using the human-as-advisor archetype take driver-suggested steering commands as inputs for state estimation. The state estimator then uses a driver model to infer the vehicle's state according to the driver from the suggested steering command. Using the driver's suggested steering input allows the state estimator to reliably estimate the vehicle's state even under challenging conditions, such as duplicative lane markings that are indistinguishable to a lane-centering camera. However, using driver-suggested steering commands for state estimation or control requires understanding the relationship between the driver's steering input and the vehicle's state. Thus, the second contribution introduced by this thesis is the generalized two-point visual control model of steering, a new driver steering model. The generalized model accurately predicts driver steering across participants, vehicle velocity, and in both straight and curved roads. The generalized model is also accurate enough for closed-loop state estimation in the face of duplicative lane markings. When used in closed-loop control in a shared autonomous vehicle, though, driver steering dramatically changes. This thesis's third contribution is therefore an exploration and explanation of how driver steering changes and can be predicted in closed-loop control with a shared autonomous controller. Specifically, drivers use their steering to regulate the total output steering applied to the vehicle. This steering input accounts for both the vehicle's actual position in the lane--that is, drivers use a steady-state steering input to cause the vehicle to be off-center in the lane--and the vehicle's transient lane-centering behavior. While this analysis revealed a means to predict driver steering input, it does not account for driver preferences, trust, or complex behavior like obstacle avoidance. The fourth contribution made by this thesis is the development of a shared autonomy vehicle simulator, along with a large-scale obstacle avoidance test examining how driver trust is related to specific controller characteristics. The simulator is entirely open-source and built on a modular structure designed to make future tests easy to develop and deploy. Initial analysis of the results from the large-scale test indicates that driver trust is significantly related to driver control authority in shared autonomous vehicles. Trajectory traces show that behavioral metrics like the magnitude of input steering may allow trust measurement.