Dyna Robotics Details Infrastructure Behind Training Dyna-2 on 1M+ Hours of Video

On August 18, Dyna Robotics, the team behind the Dyna-2 robot model, published a long-form retrospective on how they reproducibly trained models on over 1 million hours of egocentric video. The post was widely reshared on X/Twitter by JasonMa2020, ZeYanjie, chrisjpaxton, joshwills and Scobleizer. The team had also appeared on RoboPapers episode 99, addressing the industry complaint that world-action models lack concrete scaling evidence, showing predictable accuracy gains on held-out data as pretraining data approached 1 million hours.

Confirmed

• Scale: Dyna-2 was trained on more than 1 million hours of egocentric video, with a reproducible pipeline; the team claims this is the first predictable scaling demonstrated in robot learning.
• Small-scale approaches failed at scale: the old pipeline processed only 14,000 hours per week (over a year for a million hours); manifest construction took 48 hours; reading petabytes from cloud storage left GPUs idle.
• After rebuilding the stack: 31x faster data ingestion, training startup time cut from 48 hours, and 24x faster overall training, supporting the compute scaling Dyna-2 required.
• Key thesis: scaling is not a switch; bottlenecks shift from storage to ingestion, manifest generation, then the optimizer; at million-hour scale the bottleneck is the infrastructure connecting robots and GPUs, not either endpoint.
• The team stressed that infrastructure is the foundation for rapid iteration on models, data and hardware, yet is rarely discussed openly in robotics.

Why it matters

• Public accounts of large-scale training infrastructure in robotics are rare; this retrospective offers reusable engineering lessons for data throughput, training preparation and storage I/O bottlenecks, and signals that data infrastructure capability is becoming a key variable in embodied-AI competition.
• If the claimed first predictable scaling in robot learning is further validated, it would supply the missing evidence for the scaling path of world-action models.

2026-08-18 ~ 2026-08-19 · 9 related posts

• Dyna Robotics: training on 1M hours of egocentric video broke everything that worked at 10K — ZeYanjie · 2026-08-18
• Dyna Robotics scaled ingestion 31x and cut training runs from 16 months to 3 weeks — JasonMa2020 · 2026-08-18
• Dyna shares infra practices for 1M-hour video training — JasonMa2020 · 2026-08-18
• Dyna Rebuilds Infra for 1M+ Hours Robot Data, 31x Faster — JasonMa2020 · 2026-08-18
• DynaRobotics: Infrastructure for Training on 1M Hours of Video — JasonMa2020 · 2026-08-18
• Scaling ML Infra for Million-Hour Robot Training — JasonMa2020 · 2026-08-18
• Dyna Robotics nears 1M hours of training data, showing predictable scaling in robot learning — chris_j_paxton · 2026-08-18
• Training Dyna-2 on 1M hours: Infrastructure challenges in robotics — josh_wills · 2026-08-18
• Dyna-2 trained on 1M+ hours of egocentric video; data infra detailed, open-sourced — Scobleizer · 2026-08-19