Found-RL: foundation model-enhanced reinforcement learning via asynchronous VLM feedback for autonomous driving

Reinforcement learning (RL) is promising for end-to-end driving, but suffers from low sample efficiency and limited semantic interpretability. Vision-language models (VLMs) provide rich knowledge, yet their high inference cost hinders integration into RL training. We propose Found-RL, a platform for enhancing AD RL with foundation models. Its core is an asynchronous batch inference framework that decouples VLM reasoning from the simulation loop, reducing latency and enabling learning from VLM feedback. On this platform, we use VMR and AWAG to distill VLM action guidance into the policy, and adopt CLIP-based reward shaping with Conditional Contrastive Action Alignment for dense supervision. Experiments show that lightweight RL policies with millions of params can approach billion-parameter VLMs while maintaining real-time speed (~500 FPS). Code, data, and models are available at https://github.com/ys-qu/found-rl.