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.
More from Embodied
- EmbodiedEvo tops RoboDojo as agents inherit reusable skills and lessons — Aiden_Tech_Ai · 2026-10-16
- Sunday Robotics is hiring a Senior Infra PM to scale consumer robot deployments — saranormous · 2026-10-16
- Bond's Submarine Lotus Esprit Spotted at Event, Fueling Steering-Wheel-Free Robotaxi Speculation — PTrubey · 2026-10-16
- Maker wires Grok and Claude into ROS2 robot with RealSense stereo cameras — chrismatthieu · 2026-10-16
- NEURA Robotics launches Neuron AI, a unified Perception-Language-Action model for robots — Distinct-Question-16 · 2026-10-16
- Dev demos open wearable AI passport running multilingual flashcards on tiny memory — ezshine · 2026-10-16