Dyna-2 Deep Dive: World-Action Models Trained on 1M Hours of Human Video, 99.4% 24-Hour Autonomy
Latent Space · rss · 2026-10-14
Latent Space interviews Dyna Robotics co-founder and Chief Scientist Jason Ma on the Dyna-2 world-action model and the Dyna-2.1 physical agent:
- Dyna-2 was pre-trained on over 1M hours of human video, combining future video prediction with robot actions (diffusion transformers + flow matching), with claimed scaling laws for human-to-robot transfer across four orders of magnitude;
- Live demo: a robot folded restaurant napkins for 24 hours straight at a 99.4% success rate, recovering after knocking over the whole stack; new tasks need as little as 45 minutes of demonstrations;
- Real deployments at Best Western and Din Tai Fung fold thousands of napkins daily;
- Jason covers why Dyna builds its own wheeled semi-humanoid hardware, bottlenecks like overheating dexterous hands and missing sensory data, world-action models vs. VLAs, the "Bitter Lesson of Robotics" (scaling data over specialized engineering), and agentic harnesses enabling hour-long physical workflows;
- Plus GPT-6, coding agents, synthetic environments, and self-generated training data as the next wave.
More from Embodied
- RigidFormer accepted at NeurIPS: mesh-free Transformer beats GNN simulators for rigid dynamics — camenduru · 2026-10-14
- Science Corporation implants vision-restoring chip in first Stargardt patient — TechCrunch AI · 2026-10-14
- Mind Robotics, Rivian's robotics spinoff with $1B raised, hires ex-Ford EV chief Doug Field — Sethwinterroth · 2026-10-14
- AlphaNIST robot training: Newton physics plus sparse reward did the heavy lifting — milesmacklin · 2026-10-14
- Industrial Next wants factory robots to switch tasks in minutes, not weeks — TechCrunch AI · 2026-10-14
- Nature Will unveils Natura-1: 50B-parameter MoE robot foundation model trained on 1M hours of egocentric human video — rohanpaul_ai · 2026-10-14