X-EraAI/e05-robotwin-data

E05 processed RoboTwin data

• Robot benchmark tasks: 50; episodes: 2,500 (50/task).
• Frames: 549,787; LeRobot format: v2.1; robot: aloha-agilex.
• Metadata FPS: 15. This is the processed dataset's timing metadata, not a claim about the simulator's internal physics/control rate.
• Language table: 158,984 entries. LeRobot total_tasks counts language entries, not the 50 robot benchmark tasks.
• Cameras, in order: cam_high, cam_left_wrist, cam_right_wrist; RGB 240×320.
• State/action: 14 joint/gripper components in the existing dataset order. Preserve left-arm joints/gripper followed by right-arm joints/gripper as consumed by RoboTwin's joint-action interface. Do not interpret this as end-effector delta.
• Source modality metadata records state[t]=joint[t], action[t]=joint[t+1].
• Geometry cache: depth in metres, valid pixels, intrinsics [3,3,3], and OpenCV world-to-camera extrinsics [3,3,4]. Future target offset: 32 frames; targets beyond an episode are excluded by the existing padding mask.

All 2,500 episodes are used for training; there is no internal held-out loss split. Normalization uses the exact source run's statistics, verified equal to the named dataset statistics file. Geometric diagnostics on these frames are training-set diagnostics, not held-out generalization measurements.

The code expects data that have already been processed. The dataset and caches are external resources and are not included in the Git repository. RGB/Parquet/meta files retain their content and indexing. Cache shards contain original .pt files compressed losslessly; ZIP readers return the same bytes. There is no raw-data reconstruction pipeline. Only language entries referenced by this dataset are included; unrelated experimental text caches and incomplete temporary files are excluded. Novel inference instructions use the public text encoder online, without modifying the published cache.

The reference layout contains RGB/action files and metadata in the LeRobot directory, geometry cache shards by global-frame ranges of 1,000, and text cache shards by SHA256 prefix. Each cache manifest records sizes and hashes; ZIP shards retain their companion JSON entry manifests. scripts/validate_data.py --full verifies complete coverage and cache archive hashes after the data are placed in the configured directory. See the resource layout for the expected paths.

The processed-data shards in this Hub repository can be installed with scripts/install_data.py; no raw-data regeneration is required.

Provenance: processed from the locally supplied RobotTwin2.0-Depth aloha-agilex_clean_50 dataset. This is a project-specific derivative, not the upstream official dataset release. Upstream data, language and asset terms must be retained; the source code MIT license does not relicense these materials.

Download and install

The download contains about 378 GiB of losslessly compressed data and caches. Allow space for both the download and installed copy. Preserve the repository layout. Run from the e05 code repository:

source scripts/env.sh
hf download X-EraAI/e05-robotwin-data --repo-type dataset --local-dir downloads/robotwin-data
python scripts/install_data.py downloads/robotwin-data --output data
python scripts/validate_data.py --data-root data --full

The root manifest records 1,665 payload files and their SHA256 checksums: 51 RGB/action/metadata ZIPs, 1,101 reconstruction cache files, and 513 text cache files. Cache files include ZIPs, entry JSONs, and manifests. Use the e05 installer and loader, not datasets.load_dataset.

Resources

• e05 code and instructions
• RoboTwin 109k policy
• Embodied-pretrained VGGT
• Processed RoboTwin training data

Terms

See third-party notices. The code repository's MIT license does not relicense weights or data.