Jason Ma, co-founder and chief scientist of Dyna Robotics, discusses Dyna-2, a world action model that predicts future video to control robots. Trained on 1 million hours of data, it enables dexterous, generalist models that adapt to new tasks with minimal fine-tuning. The episode covers the full-stack robotics challenge: data, hardware, and deploying hundreds of robots in real-world locations like Din Tai Fung.
Summarized by Podsumo
Dyna-2 is the first world action model trained on 1 million hours of data. It predicts future video to guide robot actions, acting as a 'visual chain-of-thought' for robotics.
The model uses a diffusion transformer with flow matching, reducing inference steps from ~100 to 4, which is critical for real-time robot control (e.g., safely handling a knife).
Data strategy: pre-train on diverse human egocentric data (1M hours) for broad prior knowledge, then fine-tune with minimal robot-specific data (as little as 45 minutes) to master new tasks.
The company is full-stack, building its own semi-humanoid hardware (wheels, torso, arms) to optimize data transfer from human demos and control actuator quality.
Real-world impact: robots are already deployed at Din Tai Fung, folding thousands of napkins per day with high precision, and the company plans to scale to hundreds of robots next year.
"The mental framework to think about how do we build models for robotics is data and how to make a model learn as much as possible from that data. Under the hood is a diffusion transformer... diffusion helps you to refine the outcome to get to the smooth, precise, fine-grained motion."
"Jason Ma"
"The bitter lesson for robotics is that scaling will work... just keep the algorithm simpler and scale the right kind of data. But in robotics, to scale and how to scale is just much harder because there is no internet."
"Jason Ma"