Training Dyna-2 on 1M hours: Infrastructure challenges in robotics
josh_wills · x · 2026-08-18
DynaRobotics published details on training Dyna-2 on over 1,000,000 hours of egocentric video. They highlight that methods effective at smaller scales failed at this magnitude. Key challenges included capped ingestion throughput, 48-hour delays in building training manifests, and GPU latency from reading petabytes of cloud storage. The post emphasizes that infrastructure and data shoveling are the main bottlenecks in building robotics foundation models.
Related event: Dyna Robotics Details Infrastructure Behind 1M-Hour Video Training(8 posts)→
More from Embodied
- DriveTeach-VLA: Teaching Vision-Language-Action Models What to See and Where to Look — jiqizhixin · 2026-08-18
- China tests truck platooning: One human leads, autonomous trucks follow — georgemillo · 2026-08-18
- T800 humanoid robots deployed at Luxshare's Suzhou factory with three-finger dexterous hands — CyberRobooo · 2026-08-18
- Physics Simulation Fails Without SimReady Assets — eigenron · 2026-08-18
- Robotics Experiment Shows Data Diversity Beats Quantity — DominiqueCAPaul · 2026-08-18
- Robots need to know who is asking: Identity authentication in robotics — chrismatthieu · 2026-08-18