dancher00/WasserMan-Review-v3

WasserMan: simulation studies

To our knowledge, WasserMan is the first multi-task simulation benchmark focused on visuomotor policy learning for manipulating articulated underwater mechanisms.

Simulation data and independently audited experiment records for WasserMan. The original demonstration release is dancher00/WasserMan.

Each subset index lists the released study scope, checksums and audit. Engineering smoke runs are excluded.

selection-audit/ describes all retained original attempts. The coverage/ subsets contain the PressButton, RotateValve and OpenHatch candidate streams; ras-closure-v4-20261012/ also supplies the CollectShell, PushSlider and PullLever streams and the combined six-task initial-state analysis. diverse-demonstrations/ contains RGB, commands, measured proprioception and physical hatch evidence for the specified expert trajectory mixture, including failed attempts. RGB is lossless H.264 RGB in Matroska; rgb-storage.json provides both the compressed-file hash and the exact decoded RGB hash. Training and validation splits are whole episodes. evaluations/ contains closed-loop physical traces and complete cohort outcomes; models/ contains fixed-budget review checkpoints. Each subset index specifies its episode scope and checksums.

Policy and execution studies

Study descriptions and results cover visual transfer, RGB feedback controls, contact parameters, actuator delay, arm properties, mechanism pose, matched history and expert-trajectory ablations, complete candidate streams and full-episode task analysis.

Records	Location
Visual, physical, history and diagnostic evaluations	Evaluation traces
Matched history and trajectory-mixture training	Models and demonstrations
RGB controls, graded placement, six-task candidate coverage and full-episode contract analysis	Control and coverage records
Reset generators and exchange-pair cluster intervals	Protocol and statistical records

The reproduction guide specifies the pinned runtime, selective input restoration, full evaluation commands and independent physical scoring. Keep a fixed HF revision and the declared ordered cohort when reproducing a study.

See protocols/ for the prospective protocols, and tools/ for bounded RGB reading and verification. Simulation/controller prerequisites follow the project documentation. Evaluation scripts are versioned extensions around the original runtime.