Dyna-2: million-hour world-action model lets robots learn tasks from 45 minutes of demos
Latent Space · youtube · 2026-10-14
Latent Space interviews Dyna Robotics co-founder and chief scientist Jason Ma on physical AI in production. Dyna-2 is a world-action model trained on 1M+ hours of human video that predicts future video to guide robot actions, using diffusion transformers and flow matching for real-time control. Robots learn new tasks from as little as 45 minutes of demonstrations and already fold thousands of napkins daily at Din Tai Fung. Topics include robotics' missing internet-scale data, cross-embodiment transfer, synthetic environments and self-generated training data, custom wheeled semi-humanoid hardware, overheating dexterous hands, on-device inference and distillation, agentic harnesses for hour-long workflows, and the "Bitter Lesson of Robotics": scale data, not bespoke engineering. Jason Ma's path runs through AlphaGo, NVIDIA, Google DeepMind and Eureka.
Related event: Dyna-2 Learns to Predict the Future from Millions of Hours of Human Video(2 posts)→
More from Embodied
- DeLorean joins Airspeeder International Series, first American team debuting in November — evilrabbit_ · 2026-10-14
- Gemini 4 Argon taught Apptronik's Apollo humanoid cornhole from video, sinking 7 straight throws — GoogleDeepMind · 2026-10-14
- Box2D dev unveils contact recycling, making contact generation almost free in physics sims — EricLengyel · 2026-10-14
- Startup Squad Turns Idea Into Robot-Drawn Brooklyn Bridge Sketch in 30 Minutes — davidfromkansas · 2026-10-14
- Stream Ring, an AI wearable three years in the making, starts shipping next month — pzakin · 2026-10-14
- Dyna-2 Robots Predict the Future: Million-Hour World Models and the Bitter Lesson of Robotics — JasonMa2020 · 2026-10-14