Dyna Robotics nears 1M hours of training data, showing predictable scaling in robot learning

chris_j_paxton · x · 2026-08-18

Episode #99 of RoboPapers features the Dyna Robotics team addressing the lack of concrete scaling evidence in robot learning despite the hype around world-action models. As their pretraining data approaches 1 million hours, they report predictable, statistically significant gains in accuracy on held-out data — a rare explicit scaling demonstration in robotics. They also share lessons learned and show how the approach generalizes to many different problems.

Related event: Dyna Robotics Details Infrastructure Behind Million-Hour Robot Training(7 posts)→

Original post →

More from Embodied

Embodied channel →