Gunter, George (Vanderbilt University), Bunting, Matt (Vanderbilt University), Sprinkle, Jonathan (Vanderbilt University), Work, Daniel (Vanderbilt University)

Experimental Deployment of a Reinforcement Learning Controller Supervised by a Control Barrier Function

Scheduled for presentation during the Regular Session "S42b-Safety and Risk Assessment for Autonomous Driving Systems" (FR-EA-T42), Friday, November 21, 2025, 14:10−14:30, Broadbeach 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

In this work a reinforcement-learning (RL) policy for controlling an automated vehicle (AV) is fielded in live experiments under supervision from a control barrier function (CBFs). This work focuses on using the CBF to deploy the RL policy in an environment different from its original training while maintaining safety. The CBF provides a formal safety filter to maintain a minimum time-gap while allowing the RL policy to control the AV when safety is not of concern. The RL controller is trained in an environment designed for freeway traffic control, but is deployed onto urban arterial roads. The primary contribution of this work is to show through deployment that the CBF can maintain safety, allowing the RL to policy to be deployed to a new environment. We also show that while the CBF intervenes in some contexts, overall metrics of performance from training such as reward and estimated traffic flow energy remain close to theoretical expectations without a supervisory CBF. From this we conclude that CBFs may be an effective tool for safely deploying RL policies into new testing environments.