InternRobotics/WET-OP-Bench-Train

WET-OP Bench — Training Dataset

This dataset contains 7,000 simulated robot demonstrations for 14 scientific experiment tasks, with 500 episodes per task. The robot is Lift2. Data are stored in LeRobot v2.1 format, with four RGB camera streams, robot states, actions, timestamps, and language task descriptions.

Each top-level task folder is an independent LeRobot dataset root. Its data/, meta/, and videos/ directories must remain together. Episode indices are local to each task; this release preserves the original indices and metadata. Images are 640 × 480 and the recorded video rate is 15 FPS.

Task folder	Episodes	Frames
task_01_pour_flask_to_cylinder	500	1082240
task_02_pour_liquid_and_stir	500	1250790
task_03_magnetic_beaker_stirrer	500	924400
task_04_pour_liquid_with_funnel	500	1732500
task_05_remove_container_lid	500	364780
task_06_remove_centrifuge_tubes	500	1825520
task_07_insert_centrifuge_tubes	500	1290560
task_08_place_container_in_oven	500	2066180
task_09_solid_weighing	500	1062420
task_10_titration_experiment	500	484260
task_11_remove_container_from_oven	500	1729620
task_12_fehling_test	500	2211640
task_13_blue_bottle_reaction	500	2154840
task_14_fixed_weight_weighing	500	1474800

Download and use

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

To download one task, add allow_patterns=["task_01_pour_flask_to_cylinder/**", "README.md", "LICENSE", "manifest.jsonl"]. Point a LeRobot v2.1-compatible reader at that task folder. This repository contains multiple task roots, rather than one concatenated dataset. Feature names, shapes, camera keys and per-episode lengths are described by each task's meta/info.json and meta/episodes.jsonl.

The release was checked for file integrity, episode/frame index consistency, referenced video presence, and sample video decoding. These checks do not constitute a fresh replay or task-success evaluation of every trajectory.

Related dataset

WET-OP-Bench-Validation provides one task-01 simulator initialization example without recorded trajectories. It is not the full benchmark evaluation set.

Provenance and license

The scene and model assets were created internally by the AI Lab team. The released training data are distributed under the MIT License; see LICENSE. This repository does not redistribute the simulator SDK.

Release status

Upload in progress. Wait for the release status to change to verified before using a complete snapshot.