---
license: apache-2.0
library_name: transformers
base_model: Qwen/Qwen3-VL-4B-Instruct
tags:
  - object-navigation
  - safetensors
  - arxiv:2610.23841
---

# StopWatch perception

Draft model card.

The 4B perception model predicts explicit goal grounding from RGB observations, including a semantic inventory, goal visibility and goal distance. It forms one navigation pipeline with [the 2B policy](https://huggingface.co/Amazon-FAR/stopwatch-policy-2b), which reads the grounding and observations to choose discrete navigation actions.

## Base model

Initialized from `Qwen/Qwen3-VL-4B-Instruct` at revision `ebb281ec70b05090aa6165b016eac8ec08e71b17`, as recorded in the StopWatch training recipe. Weights use bf16 safetensors in the stock `Qwen3VLForConditionalGeneration` layout.

## Loading

```python
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

repo_id = "Amazon-FAR/stopwatch-perception-4b"
model = AutoModelForImageTextToText.from_pretrained(
    repo_id, dtype=torch.bfloat16
)
processor = AutoProcessor.from_pretrained(repo_id)
```

The full pipeline code is available at [amazon-far/stopwatch](https://github.com/amazon-far/stopwatch). Use the StopWatch code for prompt construction and Habitat-Lab evaluation. Perception performs full prefill of the latest 128 frames every step.

## Pipeline results

These are paired perception-and-policy results from the paper evaluation archive on the released clean open-vocabulary benchmarks. Our criterion requires STOP with geodesic distance below 0.25 m to any instance viewpoint of an accepted goal class under our vocabulary, measured on the Stretch protocol navmesh. Clean subsets exclude annotation failures and empty or unreachable destination sets.

| Benchmark | Successes / episodes | Success rate |
|---|---|---|
| OVON val_unseen clean | 1606 / 2309 | 69.6% |
| MP3D-OV | 595 / 1193 | 49.9% |
| HSSD-OV | 988 / 1751 | 56.4% |

## Training data

OVON, MP3D and HSSD frontier-exploration trajectories. See [Amazon-FAR/stopwatch-data](https://huggingface.co/datasets/Amazon-FAR/stopwatch-data).

## License

Apache 2.0.

## Citation

Paper: [StopWatch: Goal Grounding for Object-goal Navigation](https://arxiv.org/abs/2610.23841).

```bibtex
@article{liu2026stopwatch,
  title={StopWatch: Goal Grounding for Object-goal Navigation},
  author={Liu, Xinhao and Tian, Yunsheng and Qiu, Yuheng and Abbeel, Pieter and Chen, Xi and Shi, Guanya and Feng, Chen},
  journal={arXiv preprint arXiv:2610.23841},
  year={2026}
}
```