Ep#99: DYNA-2: A 1 Million Hour Scaling Law for World-Action Models
Jason, co-founder of DYNA, and DYNA researchers Tianyu, Quan, and Chad discuss DYNA-2’s robotics pretraining and deployment results with the Robo Papers hosts and audience.
In this talk
- 1:50
Why scale human data
Jason frames DYNA-2 as an early study of what happens when robotics pretraining scales dramatically. The team used up to one million hours of human data, primarily video from a single top-mounted camera, and found that annotation quality and diversity become constraints as the dataset grows.
- 5:00
Video helps transfer across embodiments
The team reports that validation loss on human data fell as training data increased, and that models also had lower action-prediction error on robot datasets without robot pretraining or fine-tuning. Their ablations suggest video prediction helps prevent overfitting and improves cross-embodiment generalization, even when action-labeled data is held fixed.
- 26:10
Robot results depend on post-training and evaluation
DYNA evaluated on about 40 manipulation tasks and separately tested models after lightweight post-training with limited robot data, sometimes as little as 10 to 13 minutes. Results improved with more pretraining on some tasks, but real-world evaluations were noisy, and the reported 53% success rate was well below the team’s production performance.
- 38:21
Deployment, architecture, and open questions
The researchers describe a shallow action transformer designed to retain useful motion information while running faster. They also preview a method for generating multi-view video in under 100 milliseconds and report that DYNA-2 deployments averaged 87% success against throughput and quality requirements before adaptation. They say high-quality action labels, more data types, safe exploration, and better evaluation remain important challenges.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.