Map vendor data, identify coverage gaps, and prioritize acquisition
Controlled acquisition
monocular
stereo
multi-camera
wrist-mounted
Language annotation
Task taxonomy
Match vendor proposals to tasks
language
video
action
trajectories
Internet ego video
Segment selection
Check quality and hand visibility
Cleaning
Deduplicate and validate recordings
Motion reconstruction
Recover dense hand and body motion
Data distribution
Guide sampling and collection
Kubernetes
Schedule and scale parallel pipelines
Maintain reliable execution across stages
Lakehouse
Unify videos, sensors and processing artifacts
Manage tens of PB across pipeline stages
350,000 hours per week
Process ego data for training
Scale linearly with added compute