How it works
From spec to scale in four steps. No ambiguity, no surprises.
Define the spec
We don't start with vague requirements like "kitchen tasks." We start with a spec that your training code cares about.
- —Robot embodiment and control interface
- —Action space definition (EE delta pose vs joint commands)
- —Camera count, placement, fps, resolution
- —State and action rate (Hz)
- —Calibration requirements
- —Required labels (success/fail, steps, objects)
- —File format (HDF5, RLDS, Parquet, etc.)
- —Reject rules
If you don't have a spec yet, we'll help you write one. Download our spec template.
Build a sample pack
A sample pack is not "demo videos." It's a tiny version of the real deliverable.
- —100-300 episodes
- —3 tasks
- —Full schema
- —QA report + rejection reasons
- —Loader script
Your team validates that the data loads, trains, and matches expectations. If it doesn't, we iterate until it does.
Pilot
A pilot is what turns us into a vendor.
- —2 weeks
- —50 hours of data
- —5-15 tasks
- —Weekly drops
- —Fixed acceptance tests
The pilot proves we can hit your acceptance tests on a schedule. It also surfaces any schema issues before you commit to scale.
Scale
Scaling is a process problem:
- —Operator training
- —Environment resets
- —Calibration drift checks
- —Continuous QA
- —Versioning and provenance
If you can't keep quality stable, volume is worthless. We've built the systems to maintain quality at scale. Same schema, same acceptance tests, more data.
What about existing data?
If you already have raw robot logs, we can convert them. Send us a sample of your existing format. We'll tell you:
- —What's recoverable
- —What needs manual review
- —What has to be recollected
Often 60-80% of existing data can be cleaned up. The rest usually has fundamental issues (missing timestamps, dropped frames, calibration errors) that can't be fixed after the fact.
Ready to start?
Send us your spec and we'll reply with confirmation, sample pack timeline, and pilot quote.