Blog
Notes on building the data layer for physical AI.
Research and engineering notes on how egocentric human demonstrations are collected, refined, rights-cleared and made usable for training vision-language-action models.
All
Research
Data
Community
Company
Research
Why public teleop data isn't enough
Teleoperation corpora are narrow, lab-bound and rights-ambiguous — and why real-world demonstrations close the gap.
[date placeholder]
DataWhat makes a demonstration hour usable for VLA training
Not every recorded hour trains a model. What separates a usable demonstration hour from raw footage.
[date placeholder]
DataInside our QC pipeline
How submissions move through automated checks and human review before anything reaches a delivery bucket.
[date placeholder]
CommunityMeet our first collectors
Portraits from the collector community recording the demonstrations that our refinement pipeline turns into training data.
[date placeholder]