HFlow Gives Robot Data a Paper Trail

Field notes

The page models a robot-data pipeline as four visible stages: collect synchronized streams, process them with transformations and checks, curate metadata, and deliver version-pinned manifests. Its example uses an MCAP episode, a wrist-camera black-frame measurement, and a Parquet manifest, while the shown Python API wraps existing processing code rather than replacing it. The project also emphasizes provenance, retaining pipeline and tool versions alongside each processed episode.