WasserMan documentation¶
Build and extend the simulation benchmark for underwater manipulation. Ten demonstrated environments include nine with learned-policy evaluations. The shared six-task protocol compares ACT, DP and chunked BC; the SmolVLA study covers three tasks. All results retain their declared contracts.
Get started¶
Use and reproduce¶
- Train a reference policy.
- Run the complete source workflow.
- Read the frozen evaluation protocol.
- Compare recorded results.
- Inspect data and model availability.
- Understand observations and action interfaces.
Configure and extend¶
- Configure robots, controllers and physics.
- Read physical-model assumptions.
- Add a task, robot or controller, or policy.
- Troubleshoot.
The code and documentation live in separate repositories. Protocols, runtime references and scientific summaries are bundled with the code; the full internal research archive is not part of installation. Precomputed canonical training packages are not included in 0.1.0; generate data and train with the source workflow.