README.md

---
license: apache-2.0
tags:
- autonomous-driving
- carla
- imitation-learning
- vlm
- found-rl
size_categories:
- 10G-100G
---

# Found-RL Dataset: Demonstration Data for VLM Fine-tuning

## 🚗 Dataset Overview

This dataset contains large-scale demonstration data collected from the **CARLA simulator**, designed to fine-tune Vision-Language Models (VLMs) for autonomous driving tasks. It serves as the data foundation for the paper **"Found-RL: Foundation Model-Enhanced Reinforcement Learning for Autonomous Driving"**.

The dataset comprises approximately **1.374 million state-action transitions** collected across three diverse benchmarks using expert policies.

- **📄 Paper:** [Found-RL: foundation model-enhanced reinforcement learning for autonomous driving](https://arxiv.org/abs/2602.10458)
- **💻 Code:** [https://github.com/ys-qu/found-rl](https://github.com/ys-qu/found-rl)
- **📦 Format:** Compressed `.tar.gz` archive

## 📊 Dataset Statistics & Composition

We collected demonstration data on three primary benchmarks to ensure diversity in driving scenarios. The total dataset consists of **~1.37M transitions**.

| Benchmark | Expert Policy | Episodes | State-Action Transitions |
| :--- | :--- | :--- | :--- |
| **CARLA Leaderboard** | Roach PPO Expert (Zhang et al., 2021) | 160 | ~457k |
| **NoCrash Benchmark** | Autopilot Roaming Expert | 80 | ~235k |
| **CARLA Challenge** | Autopilot Roaming Expert | 240 | ~682k |
| **Total** | - | **480** | **~1.374 Million** |

## 🛠 Data Collection Methodology

### 1. Expert Policies
- **Leaderboard Benchmark:** Data was collected using the **Roach PPO expert policy** (Zhang et al., 2021).
- **NoCrash & Challenge Benchmarks:** Data was collected using the **Autopilot roaming expert policy**.

### 2. Constraints & Filtering
To ensure high-quality training data for VLM fine-tuning, the following constraints were applied during collection:
- **Maximum Duration:** The maximum duration for each episode was set to **300 seconds**.
- **Collision Filtering:** A terminal step filtering rule was applied. A short segment of steps immediately preceding a collision event was discarded, ensuring the dataset contains only the valid, safe portion of each episode.

### 3. Usage
This data is intended to be used with open-source frameworks (e.g., *open_clip*, *LLaVA* codebases) to fine-tune VLMs, providing them with expert-level driving understanding.

If you use this dataset in your research, please cite our paper:
```bibtex
@misc{qu2026foundrl,
      title={Found-RL: foundation model-enhanced reinforcement learning for autonomous driving}, 
      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},
      year={2026},
      eprint={2602.10458},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.10458}, 
}