Robots learn from clips. People learn from the whole job. We record humans doing real work, start to finish, every cause linked to its consequence, so a model at minute 120 still knows what happened at minute 12. We want capable robots in people's lives as much as you do, and this is the data that gets them there.
Each arc is a cause and its consequence, labeled at both ends and linked. Interruptions, mistakes and plan changes stay in.
Acting is solved. Planning is close. Keeping track across hours of real work is not, and a clip cannot teach it: a clip is labeled for what happens inside it, and carries nothing about what it owes to the past. That gap is the data we make.
Two pixels wide at the scale of the job. Label: "places cup on shelf"
Every placement linked to its retrieval. Label: "glasses back to the cupboard · parked 00:49:40"
The tray soaks, the oven heats, the boxes wait by the door. Both ends labeled and linked, fifteen minutes to two hours apart.
The phone rings, the trash overflows. We record the suspended goal, the resume point and the return.
The cupboard is full, so the glasses wait on a shelf. The placement opens an obligation; the retrieval two hours later closes it.
The states clean demonstrations never reach, and the paths back from them. Nothing is re-shot.
The same job in the same place, twenty or more times, with the starting state changed on purpose. A distribution of decisions, not one lucky run.
The same footage cut into 30 to 60 second clips, shuffled.
The same clips, each with a written summary of what came before.
Full continuity plus our decision, obligation and state annotations.
Ask a model, at moments the current frame cannot answer: what is unfinished, where is the tray, what should I resume. If C wins, and wins by more as the gap widens, this data does something clips cannot. If it doesn't, you have lost nothing.
episode_0412/
├─ head.mp4 1920×1080 · 30 fps · 3 h 14 min
├─ wrist_l.mp4 synced to head
├─ wrist_r.mp4
├─ decisions.jsonl one record at every branch
├─ state.jsonl goal · suspended · owed · where things are
├─ links.jsonl cause → consequence · fail → recover
├─ subtasks.jsonl {t0, t1, phase, label}
├─ hands.npz 21 keypoints × 2 hands, per frame
└─ meta.json run, initial state, consent, known gaps
Name the job and the environment. We record one full episode in it, annotate it, run the test and send you the folder. No cost, no call required.
We reply within five hours.