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

# 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/)**

*Xiaomi Embodied Intelligence Team*

---

## Overview

**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.

Prior latent Chain-of-Thought (CoT) methods compress reasoning into opaque hidden states — fast, but consistently underperform explicit CoT on driving tasks. OneVL identifies the root cause: purely linguistic latents encode abstract semantic labels rather than the spatiotemporal causal dynamics that govern real driving scenes. OneVL addresses this with **dual-modal auxiliary decoders** that force compact latent tokens to encode both human-readable reasoning *and* future scene dynamics simultaneously.

At inference, both decoders are discarded and all latents are **prefilled** into the prompt context in a single parallel pass — matching answer-only AR prediction speed while recovering the interpretability of explicit CoT in both vision and language.

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

---

## Architecture

OneVL augments **Qwen3-VL-4B-Instruct** with three components:

**Latent Token Interface** — 4 visual latent tokens + 2 language latent tokens are inserted in the assistant response before the answer, using existing vocabulary tokens (no new special tokens added).

**Visual Auxiliary Decoder** — Predicts future-frame visual tokens at t+0.5s and t+1.0s from visual latent hidden states (using the Emu3.5 IBQ 131k codebook). Acts as a **world model** supervision signal that forces the latent space to encode genuine physical scene dynamics — agent trajectories, road geometry, and environmental change — rather than abstract descriptions.

**Language Auxiliary Decoder** — Reconstructs explicit CoT reasoning text from language latent hidden states, conditioned on ViT visual features. Recovers 97% of explicit CoT text quality while running at answer-only speed.

**Prefill Inference** — Both decoders are discarded at inference time. All latent tokens are processed in a single parallel prefill pass; only the trajectory answer is generated autoregressively. This achieves **1.5× speedup over explicit CoT on NAVSIM** and **2.3× on ROADWork**.

### Three-Stage Training Pipeline

Training proceeds in three stages to ensure stable joint optimization:
- **Stage 0**: Main model warmup (trajectory prediction)
- **Stage 1**: Auxiliary decoder warmup (language + visual decoders independently)
- **Stage 2**: Joint end-to-end fine-tuning (all components together)

Staged training is essential — ablation shows that skipping it collapses PDM-score from 88.84 to 67.13.

---

## Results

### NAVSIM

| Method | Model Size | PDM-score ↑ | Latency (s) ↓ | Interpretability |
|---|:---:|:---:|:---:|:---:|
| AR Answer | 4B | 87.47 | 4.49 | — |
| AR CoT+Answer | 4B | 88.29 | 6.58 | Language |
| COCONUT | 4B | 84.84 | 5.93 | — |
| CODI | 4B | 83.92 | 8.62 | — |
| SIM-CoT | 4B | 84.21 | 10.86 | Language |
| **OneVL** | **4B** | **88.84** | **4.46** | **Vision + Language** |

### ROADWork

| Method | ADE (px) ↓ | FDE (px) ↓ | Latency (s) ↓ |
|---|:---:|:---:|:---:|
| AR CoT+Answer | 13.18 | 29.98 | 10.74 |
| **OneVL** | **12.49** | **28.80** | **4.71** |

### Impromptu

| Method | ADE (m) ↓ | FDE (m) ↓ | Latency (s) ↓ |
|---|:---:|:---:|:---:|
| AR CoT+Answer | 1.42 | 3.96 | 6.84 |
| **OneVL** | **1.34** | **3.70** | **4.02** |

### APR1 (Alpamayo-R1)

| Method | ADE (m) ↓ | FDE (m) ↓ | Latency (s) ↓ |
|---|:---:|:---:|:---:|
| AR CoT+Answer | 2.99 | 8.54 | 3.51 |
| **OneVL** | **2.62** | 7.53 | **3.26** |

### CoT Text Quality (NAVSIM)

| Method | Meta Action Acc. ↑ | STS Score ↑ | LLM Judge ↑ | Latency (s) ↓ |
|---|:---:|:---:|:---:|:---:|
| AR CoT+Answer | 73.20 | 79.75 | 81.86 | 6.58 |
| **OneVL** | 71.00 | 78.26 | 79.13 | **4.46** |

OneVL's language auxiliary decoder recovers 97% of explicit CoT quality at answer-only inference speed.

---

## Usage

### Requirements

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

```bash
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
```

### Inference with Language + Visual Explanation

```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_explain.json \
    --device cuda:0 \
    --num_latent 2 --num_latent_vis 4 \
    --max_new_tokens 1024 --answer_prefix "[" --prefix_k 0 \
    --decoder_explain --aux_visual_condition \
    --c_thought 2 --max_explain_tokens 1024 \
    --visual_decoder_explain --visual_aux_visual_condition \
    --c_thought_visual 4 --max_visual_tokens 2560
```

### Multi-GPU Inference

```bash
export MODEL_PATH=/path/to/OneVL-checkpoint
export TEST_SET_PATH=test_data/navsim_test.json
export OUTPUT_PATH=output/navsim/navsim_results.json
bash run_infer.sh
```

Per-benchmark scripts are available in `scripts/`:

```bash
bash scripts/infer_navsim.sh
bash scripts/infer_ar1.sh
bash scripts/infer_roadwork.sh
bash scripts/infer_impromptu.sh
```

For full documentation, evaluation scripts, and data format details, see the [GitHub repository](https://github.com/xiaomi-research/onevl).

---

## Open-Source Status

| Component | Status |
|---|:---:|
| Technical Report | ✅ Released |
| Model Weights | ✅ Released |
| Inference Code | ✅ Released |
| Training Code | 🔜 Coming Soon |

---

## 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

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); please refer to their respective licenses as well.