The authors introduce WasserMan, an open-source benchmark for underwater robot manipulation policy learning. It features 10 tasks in Isaac Sim, including turning valves and recovering objects, while simulating water forces, currents, and camera effects.

  • The benchmark evaluates ACT, Diffusion Policy, and chunked BC on six tasks, and SmolVLA on three.
  • Code is available on GitHub with setup instructions for Linux, Python 3.12, NVIDIA GPUs, and Isaac Sim 6.1.
  • Task demos include scripted experts, while learned-policy results are reported separately.

This resource provides a standardized environment for testing robot learning algorithms under realistic underwater conditions.