Traly/Robotwin2.0-clean-lerobot

RoboTwin 2.0 official clean — project LeRobot v3 conversion

Downloaded, converted and validated on 2026-09-14 for the current CDLM data path in /data/dev/code/lerobot (conda lerobot). The final completion evidence is meta/validation_receipt.json.

Contents

Embodiment	Episodes	Frames
Franka	2,300	335,325
ARX-X5	2,500	388,516
Aloha AgileX	2,500	552,287
Piper	1,850	278,377
UR5	2,350	338,773
Total	11,500	1,893,278

The source contains 50 semantic tasks and 230 published task/embodiment combinations. The metadata task table contains 11,088 distinct embodiment/task/instruction labels. All selected source episodes are included. This is a source trajectory count, not a deduplicated increment over other RoboTwin datasets.

The payload consists of 12 numeric Parquet files and 37,000 H.264 MP4 files, totaling 1,456,057,898 bytes before metadata and provenance. Episode metadata occupies 12 additional Parquet files.

Representation and loader

• Format: project LeRobot v3.0, 320×240 RGB video, H.264/yuv420p, CRF 30.
• Raw joint state/action: 16D, two arms with 7 joint slots plus one gripper each. Six-joint embodiments use a zero padding slot per arm.
• Raw EEF state/action: 16D, [left xyz3, quaternion wxyz4, open-gripper1, right xyz3, quaternion wxyz4, open-gripper1].
• Actions use the next observed state; the final action repeats the final state, matching the existing project conversion contract.
• The 30 FPS value defines the conversion timebase. Physical simulation/control frequency was not independently established.
• Embodiment one-hot order: franka, arx-x5, aloha-agilex, piper, ur5.
• Head, left-wrist and right-wrist cameras are present in all 11,500 episodes. Front-camera video exists in 2,500 episodes; the other 9,000 episodes contain explicit null references and source_missing_cameras metadata.

Use the current project RobustLeRobotDataset path, including the missing-camera reader fix. Its placeholder and missing-view mask preserve absent-camera semantics. Compatibility with an unmodified external LeRobot checkout is not asserted.

meta/rotation_stats.json provides verified canonical 20D rotate6d action/state statistics with closure=1-open_gripper. Mean/std/min/max/count and q01/q10/q50/q90/q99 use the complete dataset; the quantiles are exact global linear quantiles. These are clean-only statistics. Combining clean with randomized requires the mixture's own normalization selection or aggregation.

Validation and provenance

• All 230 upstream archive SHA256 values and all extracted ZIP-member CRCs passed.
• All 23,000 consumed HDF5/instruction files were independently rehashed.
• All 11,500 episodes and 1,893,278 numeric frames matched the float32 conversion contract exactly.
• All 37,000 MP4 files and 6,232,121 camera frames were decoded; frame counts and PTS passed with zero timestamp error.
• Source RGB comparison sampled first/middle/last frames from the first episode of every task/embodiment group: 2,220 points. Minimum PSNR was 28.526 dB; the predefined failure threshold was 25 dB. Two below-30 dB reflection-detail warnings remain recorded in meta/provenance/media_warning_review.json; a near-neighbor image ambiguity is retained rather than claimed resolved by pixel similarity alone.
• Canonical statistics passed independent full-data verification. The real CDLM loader/processor chain passed 30 anchors from 10 episodes across all five embodiments, including T50 padding, normalization, gripper handling and action roundtrip.

meta/provenance/ contains upstream selection and episode lineage, component receipts, numeric details, complete payload SHA256 manifest, and statistics/readback evidence. The complete working report is /data/dev/code/lerobot/continue_dlm/experiments/results/reports/data_quality/20260914_robotwin_clean_conversion/README.md.

Source: TianxingChen/RoboTwin2.0 at revision 785feb15aa4a4f532395ad2b1d2be5f28cb561ad, selecting only the five named embodiment clean_50 variants. Raw archives and extracted data remain at /data/dataset/Robotwin2.0-clean-raw-785feb15. Legacy demo_clean archives were excluded because of substantial overlap with the Aloha selection.

Conversion used the existing scripts/convert_robotwin_to_lerobot.py --variant clean_50; the exact command and frozen code hashes are in provenance. Use a new output root for a changed conversion recipe and validate the new result. This operation did not launch training or benchmark evaluation, or alter the existing randomized dataset or its training configuration.