---
base_model: Qwen/Qwen3.5-9B
datasets:
- TVRBench/tvr-sft-va
language:
- en
license: apache-2.0
pipeline_tag: robotics
library_name: transformers
tags:
- embodied-ai
- spatial-intelligence
- vision-language
- tvr
---

# TVR-Qwen3.5-9B-VA-SFT-RL

Best-performing model from the paper **"Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?"** [[arXiv]](https://arxiv.org/abs/2606.01247).

Code: [https://github.com/aim-uofa/TVRBench](https://github.com/aim-uofa/TVRBench)

## Model Description

Fine-tuned from [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) for **Target Viewpoint Reproduction (TVR)** — the agent receives a target image and navigates in a 3D indoor environment to reproduce the exact viewpoint through active exploration.

- **Base model**: Qwen3.5-9B (8.95B parameters, full fine-tuning)
- **Training**: Visual-Action SFT + Online Multi-turn GRPO
- **Overall success rate**: 51.4% on TVRBench (500 tasks)
- **Image resolution**: 640×360 (~220 visual tokens per image)
- **Action space**: 9 actions (MoveAhead/Back/Left/Right, RotateLeft/Right, LookUp/Down, Stop)

## Usage

Evaluation requires setting up the AI2-THOR simulator. See the [TVRBench repository](https://github.com/aim-uofa/TVRBench) for detailed evaluation instructions and scripts to serve the model using vLLM.

## Citation

```bibtex
@misc{li2026lookfoundationmodelsreach,
      title={Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?}, 
      author={Liyang Li and Muzhi Zhu and Zhiyue Zhao and Hengyu Zhao and Ke Liu and Linhao Zhong and Hao Chen and Chunhua Shen},
      year={2026},
      eprint={2606.01247},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.01247}, 
}
```