Dyna Robotics scaled ingestion 31x and cut training runs from 16 months to 3 weeks
JasonMa2020 · x · 2026-08-18
Dyna Robotics, the team behind Dyna-2, has published an infrastructure retrospective on training robots on over 1,000,000 hours of egocentric video. Core thesis: scaling isn't one switch — the bottleneck keeps shifting from storage to ingestion, to the manifest, to the optimizer, and a relentless team must keep finding and fixing it.
Key results:
- Ingestion throughput scaled from 14k to 440k episode-hours/week (31x)
- First batch prep went from 48 hours to under a minute
- Episode storage 68% smaller with 2.9x faster reads
- Full runs compressed from 16 months of processing to under 3 weeks
They note that at the million-hour scale, most of what worked at ten thousand hours did not hold up — e.g., reading a petabyte from cloud storage during training left GPUs exposed to latency and packet loss. The team is hiring across ML infra, robotics and data ops.
Related event: Dyna Robotics Details Infrastructure Behind Million-Hour Robot Training(5 posts)→
More from Embodied
- Investment thesis on the robotics market: inflection point and scale — Rewkang · 2026-08-18
- Tiangong Robot Formation Prepares for Beijing Games — davidpattersonx · 2026-08-18
- HONOR's Humanoid Robot Runs So Fast the Camera Can Barely Keep Up — davidpattersonx · 2026-08-18
- Bedrock launches first fully autonomous excavators on commercial sites in TX and NV — Sethwinterroth · 2026-08-18
- Dyna Rebuilds Infra for 1M+ Hours Robot Data, 31x Faster — JasonMa2020 · 2026-08-18
- HUI360: 1M-sample dataset for human-robot interaction prediction released — jonstephens85 · 2026-08-18