Yanlj/CodeRobo-DROID-Segmentation
CodeRobo DROID Segmentation Private research dataset of frame-aligned DROID real RGB and full-scene instance/region segmentation with URDF-derived Franka/Robotiq point-line skeletons. Each episode package is stored under episodes/<episode-id>/<package-name>/<run-id>/. New deliveries use lossless H.264/RGB lossless MP4 standalone real RGB and segmentation-plus-skeleton videos per camera, under training/real_rgb/ and training/conditioning_rgb/. Clip pairs are listed in… See the full description on the dataset page: https://huggingface.co/datasets/Yanlj/CodeRobo-DROID-Segmentation.
CodeRobo DROID Segmentation
Private research dataset of frame-aligned DROID real RGB and full-scene instance/region segmentation with URDF-derived Franka/Robotiq point-line skeletons.
Each episode package is stored under episodes/<episode-id>/<package-name>/<run-id>/. New deliveries use lossless H.264/RGB lossless MP4 standalone real RGB and segmentation-plus-skeleton videos per camera, under training/real_rgb/ and training/conditioning_rgb/. Clip pairs are listed in training/pairs.jsonl, with canonical/source-frame timestamps in frame_pairs.jsonl. comparison.webm is only a review preview, never a training condition. Use the latest MP4 training run per episode/package. Superseded MKV/tar runs have been removed from the current dataset tree, not counted as extra examples. Lossless RGB H.264 is readable through OpenCV/FFmpeg; browser/hardware codec support can vary. The WebM comparison remains the browser preview.
New uploads exclude PNG intermediates, API responses, worker logs, duplicate previews and original unsampled MP4/H5 backups. Canonical real RGB is preserved losslessly, with official download provenance when the source cache exists. Local generated packages and source caches are removed after training upload verification. Superseded package/raw tar backups are removed from the current tree; ordinary Hub commit history has not been purged.
Canonical frame, original source frame, timestamp, camera identity, per-frame states, and pairing/quality records are preserved in the package. Default training conditions use the two exterior cameras. All canonical frames remain, including invalid frames with quality flags. These are not Blender or physics simulations. Object trajectory files are included only when available; the segmentation-only pipeline does not fabricate object 3D poses.
Segmentation and semantic labels are model predictions, not independent manual ground truth. completed_needs_review and not_evaluated labels must remain intact. Storage verification is not a claim of segmentation or reconstruction accuracy.
Source data: DROID. Code: hku-sail/CodeRobo. Consult the original DROID data terms before any redistribution or downstream use; this card does not assign a new license to source data or generated annotations.
Local generated packages are removed only after their pinned Hub commit has been verified. Credentials, model weights, and unrelated project data are never included.