Theses and Dissertations

Date of Award

5-1-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Qi Lu

Second Advisor

Constantine Tarawneh

Third Advisor

Wenjie Dong

Abstract

Connected autonomous vehicles improve urban driving through collaborative perception via Vehicle-to-Everything (V2X) communication. Frameworks such as V2Xverse leverage this collaboration for planning and perception, yet their controllers rely on fixed parameters that cannot adapt to varying traffic. Control Barrier Functions (CBFs) enforce safety by constraining actions within a safe set, but a fixed barrier gain imposes a single operating point: conservative settings reduce throughput while permissive settings under-react to hazards.

This research proposes an adaptive CBF framework that learns state-dependent, class-specific barrier gains via constrained Reinforcement Learning (RL). A compact policy outputs separate gains for vehicles, pedestrians, and bicycles, parameterizing a CBF quadratic program that modifies only longitudinal control. Training uses constrained PPO with a PID-based Lagrangian multiplier to satisfy a safety cost budget.

We evaluated 315 closed-loop trials across 105 routes in CARLA Town05, comparing the adaptive policy against an unfiltered proportional-integral-derivative (PID) baseline and a fixed-gain CBF filter under identical perception and planning. The adaptive method achieves the highest driving score of 75.81, and reduces collisions per route by 80.3% relative to the unfiltered baseline and by 25.9% relative to the fixed CBF filter, and increases the collision-free route rate to 59.7% while maintaining the best route completion at 90.8%. The greatest improvements appear for vulnerable road users, with bicycle collisions dropping by 87.7% and pedestrian collisions by 54.5% compared to the unfiltered baseline. Because the upstream stack is never modified, all observed improvements are attributable solely to the adaptive safety layer, supporting its portability and ease of integration into other collaborative driving platforms.

Comments

Copyright 2026 Fabian A. Hernandez. All Rights Reserved. https://proquest.com/docview/3371196577

Download

Included in

Robotics Commons

Share

COinS