# Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models
Dyna Robotics has unveiled **Dyna-2**, a world-action model (WAM) that demonstrates how large-scale human video pretraining can enable robot manipulation without action labels. The model was pre-trained on **more than one million hours of egocentric human video**—equivalent to roughly **170 years of continuous waking experience**—and shows that video prediction can drive cross-embodiment generalization at scale.
—
## Key Capabilities and Design
Dyna-2 is a generative model that jointly denoises future video and action chunks on a video-diffusion backbone. Video and action are tokenized separately and processed by distinct transformer stacks, with proprioception feeding directly into the action transformer. The model uses **flow matching** for training, with a shared trunk and two marginal velocity fields for video and action.
Unlike prior approaches, the action network is deliberately shallower and joins the video stream early, improving real-time latency without sacrificing performance. At inference, the policy remains **reactive**—it never generates or attends to predicted future video.
—
## Three Scaling Results
1. **Scaling Law on Human Data**
Across 1,000 to 1,000,000 hours of egocentric human video, all metrics improve monotonically and follow power laws. For example, accuracy@0.1 improves by **51%** while MSE only degrades by **12%**, confirming strong scaling behavior.
2. **Transfer to Unseen Robot Data**
The same checkpoints were evaluated **zero-shot** on 39 robot tasks across two bimanual platforms (12 internal tasks and 27 from xdof ABC). The model shows clear gains as human video data increases, with an inflection point between 10k and 100k hours.
3. **Video Prediction Enables Generalization**
Adding video-only pretraining (with fixed action labels) reduces zero-shot robot MSE from **0.340 to 0.120**. Crucially, human in-domain error does not improve, indicating that the benefit comes from **cross-embodiment generalization**, not pure action accuracy.
—
## On-Robot Results
After post-training on 14 tasks with at most 10 hours of robot data per task and across three embodiments (parallel-jaw arms, dexterous hands, and a semi-humanoid), mean normalized scores improved from **20% → 53%**. Notable breakthroughs include:
– **Lockbox Key Turning**: 0% success up to 100,000 hours, rising to **90%** at one million hours.
– **Bottle Cap Untwisting**: Achieved **50%** success after only **~10 minutes** of robot demonstrations.
When compared to Dyna-1 (the production VLA), early Dyna-2 achieved **1.55× success rate** and **1.12× grade** on 7 tasks. At unseen customer sites, Dyna-2 passed production criteria **87%** of the time versus **46%** for Dyna-1.
—
## Interactive Explainer
The accompanying interactive explainer breaks down architecture, scaling, human-to-robot transfer, video as a separate axis, and world-action models versus VLAs. It includes live visualizations of learning curves, robot success metrics, and detailed comparisons between joint denoising and action-only training.
—
## FAQ
### What is Dyna-2?
Dyna-2 is a world-action model (WAM) that jointly denoises future video and robot actions, pre-trained on over one million hours of egocentric human video. It is designed to transfer robot manipulation skills without action labels during pretraining.
### Is Dyna-2 available as downloadable weights or an API?
No. Dyna-2 is not offered as downloadable weights or a public API. Deployment is vendor-operated through Dyna Robotics robot cells.
### Which industries can benefit from Dyna-2?
Target industries include hospitality, commercial laundry, food service, light assembly, kitting, and facilities cleaning.
### What real-world tasks does Dyna-2 support?
Applications include trash tray clearing, first-aid kitting, tote construction, food scooping, rope tying, hanger preparation, and drink retrieval.
### How does video pretraining help robot performance?
Video prediction enables cross-embodiment generalization. Adding video-only data significantly reduces robot action error, even when action labels remain fixed.
### How does Dyna-2 compare to traditional Vision-Language-Action (VLA) models?
In head-to-head comparisons, an early Dyna-2 achieved **1.55× success rate** and **1.12× grade** over a production VLA across 7 tasks and multiple checkpoints.
—
## Conclusion
Dyna-2 demonstrates that large-scale human video pretraining can establish a scaling law for robot manipulation and enable strong zero-shot transfer to real robots. By treating video prediction as a core learning signal, the model achieves substantial gains in success rate, sample efficiency, and cross-embodiment generalization. While not publicly open-sourced, Dyna-2 represents a major step toward deployable, data-efficient robot control at scale.
