Validation of RL-Based Highway Merging Policies Under Traffic Stress and Observation Noise

Read full text Summarise with VIRAT
Authors: Funmilayo Adeeye
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Reinforcement learning (RL) has shown strong promise for autonomous driving tasks, but high average performance alone is not enough for deployment in safety-critical settings. Learned policies must also be evaluated for robustness under uncertainty, stress, and rare but safety-relevant scenarios. In this project, we validate two RL-based highway merging policies in the merge-v0 environment of HighwayEnv: one trained under baseline traffic conditions and another trained under aggressive traffic conditions. Both policies are trained using Proximal Policy Optimization (PPO) and implemented with Stable-Baselines3. Validation is carried out through large-scale rollout evaluation under both nominal and stressed traffic profiles, with additional robustness testing under observation noise. Policy performance is assessed using success rate, collision rate, lane-change behavior, and safety-oriented surrogate metrics such as time-to-collision (TTC) and minimum inter-vehicle distance. Results show that the baseline-trained policy performs slightly better under nominal traffic, whereas the aggressive-trained policy is substantially more robust under aggressive traffic conditions, achieving lower collision rates and higher success rates. These findings highlight the importance of validation under distribution shift and demonstrate that the training environment strongly shapes policy robustness in autonomous driving tasks.

Fields of Study & Subjects

Autonomous Vehicle Technology and Safety Reinforcement Learning in Robotics Traffic and Road Safety

Publication details

Journal / source
Stanford Digital Repository
Publication date
25 March 2026
Language
EN