Dyna Robotics' Jason Ma on Dyna-2 and World-Action Models for Robotics
Dyna Robotics co-founder and Chief Scientist Jason Ma details the architecture and training behind Dyna-2, a world-action model pre-trained on more than one million hours of human video. The model combines future video prediction with action generation to guide robot execution, leveraging diffusion transformers and flow matching for real-time control. Ma discusses scaling laws governing human-to-robot transfer and explains how robots can learn new tasks with as little as 45 minutes of demonstrations. He also addresses engineering tradeoffs, including why Dyna chose wheeled semi-humanoids over bipedal platforms, the need for on-device inference and distillation, and hardware bottlenecks in dexterous hands. Finally, the conversation explores real-world commercial operations at Din Tai Fung and Best Western, comparing world-action models to vision-language-action architectures and examining agentic harnesses for hour-long autonomous physical workflows.
Original sources
Podcast Latent Space
Episode:Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics ↗
Newsletter Latent Space
Issue:Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.