Benchmark contribution and literature scope¶

WasserMan provides a multi-task simulation benchmark for visuomotor learning of floating-base underwater contact manipulation: versioned tasks, physical experts, RGB demonstrations, fixed learning/evaluation contracts and controlled policy–controller–environment interventions.

To our knowledge, it is the first benchmark in this scope. This priority statement concerns the combination above, not the invention of underwater simulation, manipulation, imitation learning or bimanual robots.

Closest prior work¶

• Sanz et al., 2015 present underwater intervention benchmarking with tracking under visibility/current variations and reconstruction. This establishes prior underwater benchmarking, while addressing a different evaluation scope.
• MarineGym provides GPU reinforcement-learning simulation for underwater robots, including station keeping and trajectory tracking.
• Bi-AQUA studies bilateral imitation learning and lighting variation on underwater hardware.
• UMI-Underwater studies grasping with demonstration collection and affordance-conditioned diffusion policies on a physical platform.
• ULOHA introduces underwater bimanual hardware and evaluates ACT, DP and SmolVLA.

The related-work comparison was refreshed on 1 October 2026 using primary papers and project pages, with searches for underwater manipulation benchmarks, visuomotor benchmarks and underwater simulation learning suites. No matching multi-task simulation benchmark was identified in this bounded search; a search cannot prove an exhaustive global absence. The manuscript therefore uses “to our knowledge” with an explicit scope.

The benchmark is extensible through task and platform contracts. New versions retain past data, scores and scientific provenance; adding tasks does not change the denominator or rules of an already reported comparison.