Dyna Robotics: training on 1M hours of egocentric video broke everything that worked at 10K
ZeYanjie · x · 2026-08-18
Dyna Robotics published a thread on how they trained Dyna-2 repeatably on over 1,000,000 hours of egocentric video. They note few people discuss the infrastructure that enables fast iteration across models, data and hardware — and at this scale, most practices from 10,000 hours failed:
- Ingestion throughput capped at 14,000 episode-hours per week — a million hours would take over a year;
- Building a training manifest took 48 hours before a run could even start;
- Reading a petabyte from cloud storage during training exposed GPUs to latency and packet loss.
More from Embodied
- Kevin Zakka: mink now correctly supports ball joint limits — kevin_zakka · 2026-08-18
- Demo of flying humanoid robots — NickPassig · 2026-08-18
- Dyna shares infra practices for 1M-hour video training — JasonMa2020 · 2026-08-18
- Noble Machines named to BI's 2026 list of top robotics startups — ditzikow · 2026-08-18
- Unitree previews Superman humanoid that outjumps humans and outsprints Usain Bolt — GraceToSentience · 2026-08-18
- Uber invests in Zipline, scaling autonomous deliveries to over 1M daily — AnandSwa · 2026-08-18