@article{Qu2026, 
author = {Yansong Qu and Zihao Sheng and Zilin Huang and Jiancong Chen and Yuhao Luo and Tianyi Wang and Yiheng Feng and Samuel Labi and Sikai Chen},
title = {Found-RL: Foundation model-enhanced reinforcement learning via asynchronous VLM feedback for autonomous driving},
year = {2026},
journal = {Communications in Transportation Research},
volume = {6},
number = {3},
pages = {9640027},
keywords = {autonomous driving (AD), vision-language models (VLMs), reinforcement learning (RL), asynchronous inference},
url = {https://www.sciopen.com/article/10.26599/COMMTR.2026.9640027},
doi = {10.26599/COMMTR.2026.9640027},
abstract = {Reinforcement learning (RL) has emerged as a dominant paradigm for end-to-end autonomous driving (AD) with real-time inference. However, RL typically suffers from sample inefficiency and a lack of semantic interpretability in complex scenarios. To mitigate these limitations, foundation models (particularly vision-language models (VLMs)) can be integrated because they offer rich, context-aware knowledge. However, deploying such computationally intensive models within high-frequency multienvironment RL training loops is severely hindered by prohibitive inference latency and the absence of unified integration platforms. To bridge this gap, we present Found-RL, a specialized platform tailored to leverage foundation models to efficiently enhance RL for AD. A core innovation of the proposed platform is its asynchronous batch inference framework, which decouples heavy VLM reasoning from the simulation loop. This design effectively resolves latency bottlenecks, supporting real-time or near-real-time RL learning from VLM feedback. Using the proposed platform, we introduce diverse supervision mechanisms to address domain-specific challenges: We first implement value-margin regularization (VMR) and advantage-weighted action guidance (AWAG) to effectively distill expert-like VLM action suggestions into the RL policy. Furthermore, for dense supervision, we adopt high-throughput contrastive language-image pre-training (CLIP) for reward shaping. We mitigate CLIP’s dynamic blindness and probability dilution via conditional contrastive action alignment, which prompts discretized speed/command and yields a normalized, margin-based bonus from context-specific action-anchor scoring. Found-RL delivers an end-to-end pipeline for fine-tuned VLM integration with modular support and shows that a lightweight RL model with millions of parameters can achieve near-VLM performance compared with billion-parameter VLMs while sustaining real-time inference (~500 FPS).}
}