Amazon-FAR/stopwatch-data

StopWatch navigation data

Training trajectories for the perception and policy models, plus the released ObjectNav evaluation episodes and scoring viewpoints. RGB and semantic supervision are rendered locally from the recorded poses. Obtain scene assets separately from their publishers. Rendering writes sparse class counts, never semantic masks. See the StopWatch semantic asset instructions.

Layout

• train/{ovon,mp3d,hssd}/part-*.parquet contains training episodes.
• train/{ovon,mp3d,hssd}/manifest.json identifies the data for checkpoint resume.
• eval/{ovon_val_unseen,mp3dov,hssdov}/ contains gzipped Habitat episode files.
• protocol/instance_viewpoints/{hm3d,mp3d,hssd}/val/ contains scoring destinations.
• vocab/ contains canonical semantic names, accepted goal classes, within-class names, per-class visibility thresholds, sensor mapping tables and benchmark scoring memberships.

Training schema

One Parquet row describes one episode. T denotes its number of pre-action frames.

Column	Type	Meaning
key	string	Episode identity and local RGB video filename
scene_id, scene_dataset_config	string	Paths under the scene root
object_category	string	Goal name
category_index	int32	OVON goal index 0..378 or MP3D/HSSD semantic class 0..539
goal_category_ids	list	Zero-based semantic goal classes
n_frames	int32	T
actions	list	T actions, 0 STOP, 1 FORWARD, 2 LEFT, 3 RIGHT
poses	list	7T values, world [x,y,z,qw,qx,qy,qz], y up
geo_dists	list	T goal distances in metres

Every official training episode is included. Users render every episode, and the loader excludes STOP-invisible episodes using locally rendered semantic counts and the per-class visibility thresholds under our vocabulary. Parts preserve row order and use 2,000-row groups except for the final group.

The released policy weights were trained on an earlier generation of the OVON trajectories; the released OVON trajectories are the later generation used by the perception model. The release applies one goal vocabulary to every source, while the original training used source-specific vocabulary settings. In the original policy training on OVON, only the goal's own semantic class was accepted. Training these recipes therefore does not reproduce the paper's exact training population.

Evaluation

The populations contain 2,309 clean OVON val_unseen episodes, 1,193 MP3D-OV episodes and 1,751 HSSD-OV episodes. Episode files provide scene, episode ID, start pose and goal category. Instance viewpoint files provide destination positions by semantic class. Success requires STOP with geodesic distance below 0.25 m to any instance viewpoint of an accepted class on the Stretch protocol navmesh.

Citation

Paper: StopWatch: Goal Grounding for Object-goal Navigation.

@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}
}