Dyna Robotics Introduces Dyna-2: A World-Action Model Pre-Trained on 1 Million Hours of Human Video
1 million hours of human video. Dyna Robotics just cracked the scaling law for embodied AI — and proved it transfers to unseen robots.
Why it matters
A major research milestone in world-action models: Dyna-2 establishes empirical scaling laws on video data and demonstrates cross-embodiment generalization, directly advancing the capability foundation for physical AI systems.
The key facts
5 to knowDyna-2: world-action model pre-trained on 1M+ hours egocentric human video
Three core results: scaling law on human data, transfer to unseen robot data, cross-embodiment generalization via video co-training
Direct evidence that video scaling drives robot generalization — a capability benchmark for embodied AI
Published as technical report (research-grade disclosure, not product launch)
Addresses the 'sim-to-real' and 'human-to-robot' transfer problem, foundational for physical AI deployment
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Dyna Robotics has released Dyna-2, a world-action model pre-trained on more than one million hours of egocentric human video. The technical report establishes three results: a scaling law on human data to 1M hours, the first transfer of that law to unseen robot data, and evidence that video…