sovrano-ai/industrial_manipulation_egocentric_sample

Industrial Manipulation — Egocentric Video (Sample)

Head-mounted video of industrial manipulation tasks, recorded by workers during real production shifts. Deformable materials and metal handling.

This is a 16-clip sample from an ongoing collection. It exists so you can judge the footage for yourself and tell us whether it's useful.

What's here

Clips	16
Total duration	8 minute 14 seconds
Camera	Head-mounted, DJI Osmo Action 5 Pro
Resolution	[1920x1080]
Frame rate	[30] fps
Format	MP4 (H.264)
Audio	[Removed]

Per-clip metadata is in metadata.csv: task description in natural language, material category, duration, and outcome.

What's NOT here

Stated plainly so you don't have to find out by downloading:

• No action labels. Video only. No end-effector poses, joint states, or hand pose annotations.
• No force or tactile data.
• No depth. Monocular RGB.
• No camera calibration or intrinsics.
• Head camera only in this sample. Synchronised wrist-camera views exist in the full collection — available on request.

Collection

Footage is captured by workers wearing head-mounted cameras during their normal shifts in [an industrial facility / industrial facilities]. Tasks are not staged or performed for the camera: this is unscripted work with the variability, occlusion, and failure cases that implies.

Domains in the wider collection: Cloth deformable material handling, sheet metal work, packing and repackaging.

Current production capacity: approximately 100 hours of footage per week, scalable.

Licensing

Released under CC BY-NC 4.0.

• Academic and non-commercial research use is free. Attribution required.
• Commercial use requires a separate license. If you want to train models for a commercial product, contact us we're straightforward to work with.

Footage is anonymised: Faces are blurred.

What we're trying to learn

We're actively deciding what to add next, and we'd rather be told than guess. Two directions we're weighing:

• Action labels via hand-pose and trajectory extraction from existing footage.
• Contact force, which several people have told us matters more than anything else for deformable and metal handling.

If you work on video pretraining, world models, latent action models, or learning from human video: we'd value fifteen minutes of your read on this, whether or not it's useful to you.

Contact

Sovrano AI, robotics@sovrano.ai, https://sovrano.ai/robotics

Citation

@misc{sovrano_industrial_egocentric_2026,
  title  = {Industrial Manipulation: Egocentric Video (Sample)},
  author = {[AUTHORS]},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sovrano-ai/industrial_manipulation_egocentric_sample}
}