StopWatch:
Goal Grounding for Object-goal Navigation

Xinhao Liu, Yunsheng Tian^‡, Yuheng Qiu, Pieter Abbeel^†, Xi Chen^†, Guanya Shi^†, Chen Feng^†

^‡ Project lead. ^† Equal advising.

Getting Started

StopWatch has two models. The perception model is a 4B Qwen3-VL that predicts goal grounding, the semantics in view and the distance to the goal. The policy is a 2B Qwen3-VL that reads this grounding with RGB observations and selects STOP, FORWARD, LEFT or RIGHT.

1. Environment

The codebase targets CUDA 13 and PyTorch 2.14.1. FlashAttention 4 requires a Hopper or Blackwell GPU, such as H200 or B200. Install uv, then clone the repository and install the locked dependencies.

git clone https://github.com/amazon-far/stopwatch.git
cd stopwatch
uv sync --locked
bash scripts/install_habitat_sim.sh
source .venv/bin/activate

Use uv sync --locked --inexact for later syncs to keep the Habitat-sim wheel.

2. Run inference

Models load directly from Hugging Face. Download the evaluation data.

hf download Amazon-FAR/stopwatch-data --repo-type dataset \
  --include 'eval/*' --include 'vocab/*' --include 'protocol/*' --include README.md --local-dir data

Obtain scene assets from the publishers and place them under data/scenes.

Benchmark	Scene assets and official downloads
OVON	HM3D v0.2, validation Habitat meshes in hm3d-val-habitat-v0.2.tar
MP3D-OV	Matterport3D v1.0, the Habitat package mp3d_habitat.zip
HSSD-OV	HSSD-200, the static scenes/ collection in HSSD-Hab, with stage and object configs, render and collision meshes, and the publisher scene dataset config

data/scenes/
|-- hm3d/val/<scene-directory>/
|   |-- <scene>.basis.glb
|   |-- <scene>.basis.navmesh
|-- mp3d/<scene>/
|   |-- <scene>.glb
|   |-- <scene>.house
|   |-- <scene>_semantic.ply
|   |-- <scene>.navmesh
|-- hssd-hab/
|   |-- hssd-hab.scene_dataset_config.json
|   |-- scenes/*.scene_instance.json
|   |-- stages/
|   |-- objects/
|   |-- semantics/
|   |-- metadata/

Start with ten episodes.

torchrun --standalone --nproc_per_node=1 -m stopwatch.evaluate \
  --config configs/eval_ovon_val_unseen.yaml --limit 10

Run the full benchmarks with the following commands. OVON val_unseen uses the 2,309-episode clean subset described in the paper.

Benchmark	Episodes
OVON val_unseen	2,309
MP3D-OV	1,193
HSSD-OV	1,751

torchrun --standalone --nproc_per_node=1 -m stopwatch.evaluate \
  --config configs/eval_ovon_val_unseen.yaml
torchrun --standalone --nproc_per_node=1 -m stopwatch.evaluate \
  --config configs/eval_mp3dov.yaml
torchrun --standalone --nproc_per_node=1 -m stopwatch.evaluate \
  --config configs/eval_hssdov.yaml

3. Training

Both models train on OVON, MP3D and HSSD demonstrations. Follow the training guide to download trajectories, render RGB and sparse semantic counts, train, resume and export weights.

torchrun --nnodes=8 --nproc_per_node=8 --node_rank="$NODE_RANK" --master_addr="$MASTER_ADDR" --master_port=29500 \
  -m stopwatch.train --config configs/train_perception_4b.yaml
torchrun --nnodes=2 --nproc_per_node=8 --node_rank="$NODE_RANK" --master_addr="$MASTER_ADDR" --master_port=29500 \
  -m stopwatch.train --config configs/train_policy_2b.yaml

License

StopWatch code, models and data use the Apache License 2.0. Scene datasets retain their own terms. See third-party notices.

Citation

If you use StopWatch, please cite this paper.

@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}
}

Acknowledgements

We build on Habitat-sim, Habitat-Lab, HM3D-OVON, Qwen3-VL and verl.

Relevant work

• CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos
• Wanderland: Geometrically Grounded Simulation for Open-World Embodied AI
• GaussGym: An open-source real-to-sim framework for learning locomotion from pixels
• VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation
• NaVid: Video-based VLM Plans the Next Step for Vision-and-Language Navigation
• Uni-NaVid: A Video-based Vision-Language-Action Model for Unifying Embodied Navigation Tasks
• Nav-R²: Dual-Relation Reasoning for Generalizable Open-Vocabulary Object-Goal Navigation
• LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation
• ABot-N0: Technical Report on the VLA Foundation Model for Versatile Embodied Navigation
• Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System
• Embodied Navigation Foundation Model
• OmniNav: A Unified Framework for Prospective Exploration and Visual-Language Navigation

Star history