Delivery
Your schema. Your bucket. Your pipeline.
Training-ready files in the formats your stack already reads, with the QC trail attached.
Formats
The formats your stack already reads.
- Robot learning formats
- LeRobot v3 · RLDS · HDF5 · Zarr
- Robotics logs
- ROS 2 bag · MCAP · Parquet
- Annotation formats
- COCO · JSONL · KITTI · custom
Synchronization & metadata
Measured, never assumed.
- Synchronization
- ≤ 1 frame skew · IMU ±20 ms
- Measured for every episode and written to its sync report.
- Per-episode metadata
- Scene · task · embodiment · rig · consent ID · QC verdict
- Travels with every episode.
- Transfer
- S3 · GCS · Azure · Hugging Face · encrypted disk
- To your bucket or repository, or on encrypted disk when data can’t travel over the network.
Quality
The QC trail
In every delivery
- Verdicts from two reviewers for every episode
- Reviewer agreement for every batch
- Rejections with reason codes. Nothing substituted quietly.
- Review notes: what changed and why
In production
A dedicated team and a weekly batch report: what was delivered, what was rejected and why.
See one episode
Sample episode
See what one episode looks like.
Raw capture to training file, including what the reviewer changed.