WasserMan: underwater manipulation demonstrations

Project website · Code · Paper

Version wm-open-v2-20260929: expert trajectories for the six-task core of WasserMan. This dataset contains 480 selected successful demonstrations (456 training, 24 validation), with all 504 collection attempts retained. These are simulated trajectories collected with feedback experts using the open-procedural-v1 assets, not real-world recordings.

The 510 original lossless archive shards occupy 35.69 GiB (38.32 GB). They contain RGB observations, measured robot state, commands, physical state traces and per-episode metadata. This upload does not include trained policy weights, test-policy rollouts, the three-task SmolVLA extension or historical CAD-profile datasets. The paper covers additional tasks and experiments.

Pretrained policies

The original paper checkpoints are available in WasserMan models: 54 ACT/DP/chunked-BC models, six additional EE-interface DP models, and nine SmolVLA models, retaining all three training seeds per configuration. The model release includes verified checkpoint hashes, original configurations, recorded result summaries, a selective downloader and native inference commands. SmolVLA checkpoints require reconstruction from the pinned pretrained base; the model card provides the setup tool and exact dependency revisions.

Tasks and fixed splits

Task	Train	Validation	Retained attempts	Download GiB
PressButton	76	4	88	0.82
RotateValve	76	4	88	4.03
OpenHatch	76	4	80	6.30
CollectShell	76	4	80	17.91
PushSlider	76	4	80	2.80
PullLever	76	4	88	3.83

Use the explicit seed lists in splits/, not a new random split. The protocol selects the first 80 successful audited episodes per task; failed/unselected attempts are retained for provenance and must not be added to the training set. These validation episodes are not the benchmark's independent online test resets.

Download and restore

Requires Python 3.11+, huggingface_hub, zstandard, and NumPy to read trajectories.

pip install huggingface_hub zstandard numpy
hf download dancher00/WasserMan --repo-type dataset \
  --include README.md dataset-index.json 'tools/*' --local-dir WasserMan
python WasserMan/tools/download_task.py --task PressButton --seed 40000 \
  --output ./wasserman-sample

Omit --seed to restore all attempts for one task. The downloader verifies the SHA-256 of each archive and its individual members. Metadata and split audits are restored alongside the episode files. Downloads are cached by Hugging Face; allow space for both compressed downloads and extracted archives. Raw RGB expansion requires substantially more space: restore one task at a time.

from pathlib import Path
import json
import numpy as np

episode = Path('wasserman-sample/archives/PressButton/train_batch00/seed_40000')
metadata = json.loads((episode / 'metadata.json').read_text())
with np.load(episode / 'trajectory.npz', allow_pickle=False) as trajectory:
    print({key: trajectory[key].shape for key in trajectory.files})

Each episode has metadata.json, trajectory.npz, and wrist.rgb.mkv. The MKV uses lossless RGB H.264, not a presentation-quality compressed video. Observation shape, recording rate, timing convention, action frame and success contract are specified per task in metadata.json. Numeric field names and shapes are task-specific; inspect the NPZ keys instead of assuming a single cross-task action representation. Diagnostic contact/object state must not be used as policy inputs in the standard protocol.

For the benchmark's native raw-RGB readers, decode with the original checked restoration tool (requires imageio-ffmpeg; it keeps a 40 GiB free-space reserve):

pip install imageio-ffmpeg
python WasserMan/tools/restore_revision_rgb.py \
  --archive wasserman-sample/archives/PressButton \
  --output wasserman-sample/data/PressButton --episode seed_40000

The tool checks decoded RGB bytes against the original manifest. Consult the code documentation for native training setup. This is an archive dataset, not a Hugging Face load_dataset() Parquet table or a converted LeRobot dataset.

Provenance and limitations

dataset-index.json is the authoritative file manifest for this HF distribution. SHA256SUMS records the same archive hashes for command-line verification. The full original package indices are retained under provenance/; they also list model/evaluation assets that are not part of this dataset. Their historical private-access statements describe the original release, not this public mirror. No trajectory, image, success criterion or train/validation split was changed.

These data support simulation benchmark reproduction. They do not establish calibrated underwater optics, physical fidelity to a particular deployment site, or successful transfer to real underwater robots. Consult the paper for scope and limitations. The retained license and third-party notices apply; proprietary Reach CAD is not bundled in the open-procedural profile.

Citation

@misc{belov2026wasserman,
  title = {WasserMan: Benchmark for Underwater Manipulation Policy Learning},
  author = {Belov, Danil and Erkhov, Artem and Parsegov, Sergei and Osinenko, Pavel},
  year = {2026},
  eprint = {2610.04536},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2610.04536}
}