World models & VLA
event_type · action · tools
the data startswhen the job breaks.
Robots learn the task from demonstrations. They learn recovery from nothing. We capture the part everyone throws away.
- capture
- egocentric + fixed, synced
- labelled
- failure · diagnosis · intervention · recovery
- validated
- by the technician who did the work
a successful demonstration ends where the hard part begins.
Task-demonstration datasets record the clean run and stop. Four things happen in real industrial work that they have no representation for at all.
01 / 04 · failure
failureFailure
The moment the job stops going to plan, with the seconds either side of it. A rounded bolt head, a dropped bead, a jam that returns four cartons later.
diagnosisDiagnosis
What the technician checks, in what order, and what they rule out. This is the expensive part of the expertise and the part no procedure document contains.
interventionIntervention
The off-procedure work. Tool changes, improvised fixtures, heat, force, patience. Often the longest stretch of the episode and always the least documented.
recoveryRecovery
Getting back to the procedure, not just getting it working. A model trained only on clean runs has no policy for the return path.
every episode: failure, diagnosis, intervention, recovery.
what gets built from records like this.
Robotics & humanoids
repeated_attempt · attempts · failed_attempts
Industrial copilots
symptoms · root_cause_provenance · confidence.causal
Inspection & anomaly
off_nominal · evidence_count · grounded
two ways in.
for robotics and physical AI
Request sample data
A full episode with video, labelled timeline and structured record, plus the schema. Enough to judge whether recovery data changes your policy.
for industrial operators
Capture your data
We fit capture around work that is already happening. Your technicians keep working; the expertise stops leaving when they do.