hugging face · gated · ~275 hrs
Ego-exo manufacturing. Real shoe-manufacturing operations: standalone sessions with action labels, plus 40 frame-synced ego+exo pairs annotated to atomic actions with expert commentary. Apache-2.0.
Request access →samples are free
Request sample data.
The gated sets on Hugging Face you can request directly. For anything below, tell us what you are training and we send the cut that fits, usually within a day.
~5.6 h · ego + exo video · action labels
Synced ego + exo. Real production work with the head-mounted view and a fixed second camera in frame-level sync.
5 min · ego video · 72 step labels
Atomic actions, one line. A protective-gear line annotated to atomic actions, segmented against frame ranges.
25 min · ego video
Five tasks, five minutes. Lint removal, rotary cutting, metal polishing, small assembly, ironing. One clip each.
UMI · wrist + base RGB-D
UMI bimanual. Two-handed UMI with per-wrist views and head pose, delivered with evaluation splits.
bimanual gloves · tactile signals
Gloves + tactile. Instrumented-glove capture with tactile channels. Few stacks have this modality at all.
ego video
Failure episodes. Real mistakes and recoveries. Demonstration datasets usually leave these out; QA models need them.
15 fps · raw depth arrays
RGB-D. Depth as raw .npy alongside the color stream.
How this goes.
- Take a sample and train on it. A full episode set in LeRobot format, for a task you choose. Load it, look at the frames, check the labels against the video.
- Name the capability gap. Tell us the behaviour your model fails at and we will say plainly whether we have it, can capture it, or cannot.
- Specify the capture. Where the answer is “can capture,” we scope the environment, task list, camera rig and annotation schema before anyone signs anything.
If we don’t have what you need, we’ll say so.