Benchmark for Underwater Manipulation Policy Learning.
To our knowledge, WasserMan is the first multi-task simulation benchmark focused on visuomotor policy learning for manipulating articulated underwater mechanisms. The suite combines contact tasks, underwater dynamics and paired execution studies.
Fixed-policy evaluations vary water appearance, lighting, RGB feedback, contact parameters, actuator delay, arm properties and mechanism pose. Matched expert references establish feasibility under the same physical task tests. Controlled training ablations compare observation history and expert trajectory mixtures; graded placement tests expose their condition-dependent effects.
Complete expert candidate streams make demonstration selection inspectable. Contact, tracking, actuator and full-episode contract records explain execution beyond a completion score. The study guide links every comparison to its public datasets, models and evaluation records.
- 10 tasks — contact, recovery, delivery and underwater intervention; nine have learned-policy evaluations.
- Two robot platforms — BlueROV2 Heavy + Alpha 5 and RexROV2 + Oberon7, plus bimanual valve manipulation.
- Visual baselines — ACT, Diffusion Policy and chunked BC on six tasks; SmolVLA on three tasks, each with three training seeds.
- Underwater physics and control — currents, buoyancy, thruster response, PID/PD comparisons, optional wall/seabed effects and camera imaging.
- Auditable evaluations — fixed task contracts and splits, paired policy studies, complete expert candidate streams, checkpoints and per-episode physical records.
Requires native Linux, Python 3.12 and a compatible NVIDIA GPU/driver. The open core uses Isaac Sim 6.1, Isaac Lab and procedural arm geometry.
git clone --branch main https://github.com/dancher00/wasserman.git
cd wasserman
./scripts/bootstrap.sh --asset-profile open-procedural-v1
.venv/bin/python scripts/setup_policy_dependencies.py
.venv/bin/python scripts/setup_training_backbones.py
uv pip install --python .venv/bin/python --no-deps -r research/media-requirements.txt
export WASMAN_ASSET_PROFILE=open-procedural-v1The installer verifies the scientific runtime files bundled with this checkout. See Installation and Verify installation.
Step an environment with zero actions:
OMP_NUM_THREADS=4 .venv/bin/python scripts/rollout_button_visual.py \
--mode zero --purpose development --seeds 31000 --steps 16 \
--output-dir artifacts/installation-zeroRecord a development demonstration:
OMP_NUM_THREADS=4 .venv/bin/python scripts/benchmark.py collect \
--task PressButton --purpose development --seeds 31000 \
--output-dir artifacts/button-demoUse a fresh output directory for each run. The 16-step check verifies loading and stepping; it does not measure success. Follow the training guide and frozen protocol for full experiments. No pretrained visual-policy download is needed to start.
Guides cover collection, training, evaluation and extending the benchmark. See models and evidence, policy results, including SmolVLA, bimanual studies and Contributing. The website source lives in a separate repository.
Please cite the WasserMan paper on arXiv:
@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},
doi = {10.48550/arXiv.2610.04536},
url = {https://arxiv.org/abs/2610.04536}
}Machine-readable citation metadata: CITATION.cff.
Apache License 2.0. Third-party components retain their own terms; restricted robot CAD and Isaac Sim binaries are not redistributed. See Third-party notices.