A private evaluation sample of real-world, egocentric factory-floor manipulation recordings captured during shoe production. Provided to qualified teams evaluating the full Daneel Data corpus for robotics foundation-model training.

This is not an open-source release. Access is granted by request, for evaluation purposes only.

What's included

25 unique 3-minute clips, organized into 3 task packs (57 pack-copies total)
22 workers, 19 task subcategories across 19 production stages
Egocentric video paired with synchronized wrist + head motion sensing, pre-aligned to the camera frame timeline
Task metadata with plain-language task descriptions
Step-level annotations encoded via the task-pack / category / subtask directory hierarchy

Task organization

The same set of clips is presented through three lenses:

Capability Task Packs — grouped by manipulation capability: contact-rich adhesive application; fine bimanual alignment & attachment; machine-assisted operations; inspection / visual checking / rework.
Sole Attachment Workflow — end-to-end coverage of one critical sub-process: body_bottom_glue → sole_glue → sole_attachment → final_press → sole_inspection_rework.
Full Production Line — all 19 stages of shoe assembly, suitable for studying production-stage transitions and long-horizon dependencies.

Package layout

daneel-data-sample/
├── README.md
├── LICENSE                         (CC BY-NC-ND 4.0)
├── thumbnail.jpg
└── <task_pack>/<category>/[<subtask>/]
    ├── intrinsics.json             (camera intrinsics + axis conventions, same across all clips)
    └── <clip_id>.tar               (one clip per archive)

Each .tar archive contains, for one clip:

Egocentric video (~3 min)
Synchronized per-frame motion vector (CSV — one row per video frame)
Compact JSON metadata: clip_id, subject, task (name + plain-language description), duration_s

Loading

import csv, json, tarfile, io

clip_id = "factory_001_worker_016_001"
with tarfile.open(f"{clip_id}.tar") as tf:
    meta = json.load(tf.extractfile(f"{clip_id}/metadata.json"))
    rows = list(csv.DictReader(io.TextIOWrapper(
        tf.extractfile(f"{clip_id}/per_frame_imu.csv"), encoding="utf-8")))
    tf.extract(f"{clip_id}/{clip_id}.mp4", path=".")

License

CC BY-NC-ND 4.0. Inspection and citation for evaluation purposes are permitted. Redistribution and commercial use are not.

Commercial use & full corpus access

The full Daneel Data corpus, and commercial licensing of any subset, are available on request.

Website: https://www.daneeldata.com
Inquiries (commercial / partnership / dataset access): see website

Citation

@misc{daneeldata2026sample,
  title  = {{Daneel Data}: Real-world factory manipulation data for robotics foundation models},
  author = {{Daneel Data}},
  year   = {2026},
  url    = {https://www.daneeldata.com},
  note   = {Sample package, v1}
}

Downloads last month: 21

Total file size:

8.47 GB