End-to-End Robotics Fails on Hour-Long Industrial Workflows
Dyna Robotics Chief Scientist Jason Ma on why long-horizon automation requires separating task reasoning from continuous motor control.
Robots That Predict the Future: Dyna-2 & Physical AI — Jason Ma, Dyna Robotics
With swyx, Jason Ma, George · Sunday, October 18, 2026
Dyna Robotics co-founder and Chief Scientist Jason Ma joins swyx on-site to discuss Dyna-2, a world-action model trained on over one million hours of video data that predicts future visual states while generating continuous robot actions. Ma breaks down why Dyna builds custom semi-humanoid hardware instead of bipedal humanoids, how agentic harnesses enable multi-hour industrial workflows like laundry and napkin folding at Din Tai Fung, and why scaling data beats hand-crafted robotics heuristics.
Key takeaways
- Short robotics demonstrations fail in commercial deployments because factories and kitchens measure success in continuous, unsupervised hours rather than 30-second clips. Read more →
- Vision-Language-Action (VLA) models map camera frames straight to motor commands without visualizing intermediate physics, which causes them to fail when contact dynamics drift. Read more →
- Five-finger dexterous hands overheat during sustained physical tasks, forcing robotics teams to place ice bags onto actuators during active training runs. Read more →
- Dyna trains its base model, Dyna-2, on over one million hours of egocentric human video to give it a physical world model before touching robot hardware. Read more →
4 articles from this episode
Why Dyna-2 Uses World-Action Models Over Plain VLAs
Jason Ma explains how Dyna-2 uses visual chain of thought and flow matching to generate continuous robot actions across 100 degrees of freedom.
Why Physical AI Runs on Ice Bags and Simple Grippers
Dyna Robotics co-founder Jason Ma explains why fragile robot hands and missing sensory data stall physical AI in the real world.
How Dyna Trains Robot Models on 1 Million Hours of Video
Jason Ma explains how Dyna trains robot models on 1 million hours of human video and fine-tunes new tasks in 45 minutes.
More Latent Space episodes
- How to kill your SaaS: what we learned replacing a $43K/yr tool with AI agentsOct 18 · 4 articles
- The Bitter Lesson of Biology: Why Scaling AI Isn't Enough — Pushmeet K, DeepMind & Sal C, BiohubOct 11 · 4 articles
- AI Scientists Are Here: Autonomous Labs & Synthesis Superintelligence — Periodic LabsOct 11 · 5 articles
- The Future of Claude Code: Mods, Mutable Software, & Multiplayer Agents — Thariq Shihipar, AnthropicOct 4 · 4 articles