Dyna-2 trained on 1M+ hours of egocentric video; data infra detailed, open-sourced

Scobleizer · x · 2026-08-19

Robotics firm Dyna published how they trained Dyna-2 repeatably on over 1,000,000 hours of egocentric video, noting that what worked at 10k hours broke at scale: ingestion capped at 14,000 episode-hours/week (a million hours would take over a year), training manifests took 48 hours to build, and reading a petabyte from cloud storage exposed GPUs to latency and packet loss.

A YC team resharing the thread called it a gold mine and open-sourced HFlow, a reference implementation of the described infrastructure: Airflow for DAG orchestration, MCAP for episode storage with random access, and DuckDB for analytical queries and dataset curation over columnar metadata. Their takeaway after two months at YC: at scale, AI problems become software engineering problems—fix one bottleneck and it moves elsewhere.

Related event: Dyna Robotics Details Infrastructure Behind 1M-Hour Video Training of Dyna-2(9 posts)→

Original post →

More from Embodied

Embodied channel →