Candidate: Xiaofei Wu
Date: September 2, 2025
Time: 10:00am
Location: EIT 3145
Supervisors: Drs. Michael Fisher and Stephen L. Smith
All are welcome!
Abstract:
A key factor in increasing people’s level of acceptance of autonomous driving technology is trust. Personalized autonomous driving that mimics the driver’s own driving style is a viable approach. Many existing works explore learning-based approaches to achieve this goal but may be inadequate when dealing with unseen scenarios and enforcing safety guarantees. To mitigate these difficulties, we propose a hierarchical autonomous vehicle control framework, where the upper level mimics a target driver’s driving style and the lower level performs vehicle motion planning.
The lower-level motion planner solves a trajectory optimization problem, where nonlinear dynamics of the vehicle model makes the problem challenging. Many existing approaches formulate this as multistage programs and use derivatives of each stage to obtain a local approximation at each iteration. We develop a novel approach for obtaining improved local approximations using an input-to-state reformulation of system dynamics and trajectory sensitivities, which are derivatives of the entire system trajectory with respect to control inputs. The method is proved to converge with input-affine inequality constraints and is applied to generate trajectories for an autonomous vehicle in a variety of scenarios.
The upper-level driving style mimicking problem solves for weight factors that are used to parameterize the lower-level objective. We adopt a gradient-based approach to solve this problem. As differentiability is not guaranteed given the bilevel structure and the nonconvex lower level solution mapping, we use subgradients, which are generalizations of gradients for nondifferentiable functions, and a projected subgradient update algorithm to solve this problem. Simulations show that the proposed framework is capable of solving tracking and obstacle avoidance problems while mimicking driving style.