InternRobotics/WET-OP-Bench-Validation

WET-OP Bench — Validation Example

This release contains exactly one episode for task 01: pouring from an Erlenmeyer flask into a graduated cylinder. It provides the simulator scene, robot assets, task configuration and saved initial state. It contains no recorded trajectory or LMDB database.

This is a single initialization example, not the complete validation set and not a basis for reporting full-benchmark performance. No other validation episodes, test episodes or private holdout episodes are included.

Contents

• saved/tasks/scientific_experiments/task_01_pour_flask_to_cylinder/validation_mixed_1/000/meta_info.pkl: saved task state and instruction. Recording counters are cleared; initial task data are preserved.
• saved/assets/scene_usds/scientific_experiments/task_01_pour_flask_to_cylinder/public_validation/: original scene and its dependencies, with portable references.
• saved/assets/robot_usds/lift2_scientific_experiments/: robot assets.
• configs/tasks/scientific_experiments/task_01_pour_flask_to_cylinder/validation/: single-case configuration and selector.
• runtime_dependencies.json: simulator and material requirements.
• manifest.jsonl: file sizes and SHA-256 digests.

The scene background, object geometry, initial state and physics settings are preserved. public_validation is the publication directory name, not a newly generated background.

Download and load

from huggingface_hub import snapshot_download
path = snapshot_download(
    repo_id="InternRobotics/WET-OP-Bench-Validation",
    repo_type="dataset",
    local_dir="WET-OP-Bench-Validation",
)

Use an isolated GenManip-Sim checkout compatible with Isaac Sim 4.1 and the scientific-experiment Lift2 tasks. Place the downloaded saved/assets, saved/tasks, and configuration entries at their matching repository-relative locations. Select scientific_experiments/task_01_pour_flask_to_cylinder/validation/mixed/mixed_1.yml. The simulator and GenManip-Sim code are not included here. meta_info.pkl is a Python pickle intended for the trusted GenManip loader.

Isaac Sim supplies the built-in OmniPBR.mdl and OmniGlass.mdl material modules. The tabletop shader references the official NVIDIA Plastic_Thick_Translucent.mdl URL recorded in runtime_dependencies.json; network access is needed on first load. This shader is byte-identical to the original runtime material and is not bundled in this dataset.

Verification and limitations

The package passed isolated Isaac Sim 4.1 initialization and 20 hold steps with four camera observations. Geometry, physics and material values were compared against the original scene. Paired reset and final images preserve its appearance. This is an initialization/packaging check, not a successful pouring demonstration or policy evaluation. The original initial camera framing is retained. The overhead view shows both task vessels; individual fixed cameras provide partial views.

Related dataset

WET-OP-Bench-Train contains 7,000 demonstrations across 14 task folders.

Provenance and license

All scene and model assets were created internally by the AI Lab team. The team's released data and assets use the MIT License in LICENSE. Bundled NVIDIA material software modules retain their embedded upstream license notices. The bundled r5a_left_arm.yml cuRobo planner configuration retains its original NVIDIA notice and LICENSE.curobo, which limits that software configuration to research/evaluation use. Robot-specific assets are team-authored. NVIDIA runtime software and the remotely referenced shader remain subject to their original terms; the dataset MIT license does not relicense them.

Release status

Verified: all published payload file sizes and content hashes were checked against the local release.