---
base_model:
- Qwen/Qwen3-VL-4B-Instruct
language:
- en
license: apache-2.0
pipeline_tag: image-to-image
library_name: transformers
tags:
- autonomous-driving
- vision-language-action
- chain-of-thought
- trajectory-prediction
- VLA
---

# OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

**[📄 Paper (arXiv)](https://arxiv.org/abs/2604.18486)** | **[💻 GitHub](https://github.com/xiaomi-research/onevl)** | **[🌐 Project Page](https://Xiaomi-Embodied-Intelligence.github.io/OneVL/)**

OneVL is a Vision-Language-Action (VLA) framework for autonomous driving that achieves state-of-the-art trajectory prediction accuracy while matching the inference latency of answer-only autoregressive models.

## Overview

OneVL addresses the limitations of prior latent Chain-of-Thought (CoT) methods by introducing **dual-modal auxiliary decoders**. These decoders force compact latent tokens to encode both human-readable reasoning and future scene dynamics. During inference, these decoders are discarded, and the latent tokens are prefilled into the context in a single parallel pass, achieving high performance at answer-only speeds.

### Key Architecture Components

- **Latent Token Interface**: 4 visual and 2 language latent tokens inserted before the response.
- **Visual Auxiliary Decoder**: Acts as a world model, predicting future-frame visual tokens (at t+0.5s and t+1.0s).
- **Language Auxiliary Decoder**: Reconstructs explicit CoT reasoning text from language latent hidden states.
- **Prefill Inference**: Enables 1.5× to 2.3× speedup over explicit autoregressive CoT.

## Usage

### Requirements

- Python 3.10+, CUDA GPU (≥16 GB VRAM recommended)
- `transformers >= 4.57.0` (required for `Qwen3VLForConditionalGeneration`)

```bash
# Environment Setup
uv venv venv/onevl --python 3.12
source venv/onevl/bin/activate
pip install -r requirements.txt
```

### Inference (Trajectory Prediction Only)

```bash
python infer_onevl.py \
    --model_path /path/to/OneVL-checkpoint \
    --test_set_path test_data/navsim_test.json \
    --image_base_path "" \
    --output_path output/navsim/results.json \
    --device cuda:0 \
    --num_latent 2 --num_latent_vis 4 \
    --max_new_tokens 1024 --answer_prefix "[" --prefix_k 0
```

For full inference options, including language and visual explanations, please refer to the [GitHub repository](https://github.com/xiaomi-research/onevl).

## Results

OneVL is the first latent CoT method to surpass explicit autoregressive CoT across all major autonomous driving benchmarks.

| Benchmark | Metric | AR CoT+Answer | OneVL |
|---|:---:|:---:|:---:|
| **NAVSIM** | PDM-score ↑ | 88.29 | **88.84** |
| **ROADWork** | ADE (px) ↓ | 13.18 | **12.49** |
| **Impromptu** | ADE (m) ↓ | 1.42 | **1.34** |
| **APR1** | ADE (m) ↓ | 2.99 | **2.62** |

## Citation

```bibtex
@article{lu2026onevl,
  title={OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation},
  author={Lu, Jinghui and Guan, Jiayi and Huang, Zhijian and Li, Jinlong and Li, Guang and Kong, Lingdong and Li, Yingyan and Wang, Han and Xu, Shaoqing and Luo, Yuechen and others},
  journal={arXiv preprint arXiv:2604.18486},
  year={2026},
  url={https://arxiv.org/abs/2604.18486}
}
```

## License

This project is released under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0).
Model weights are built on [Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct) and the visual tokenizer is from [Emu3.5-VisionTokenizer](https://huggingface.co/BAAI/Emu3.5-VisionTokenizer).