Dyna Robotics unveiled Dyna-2 on August 10, 2026 — a world-action model trained on more than one million hours of first-person human video. The company claims it is the first documented human-to-robot scaling law: more human video predictably improves a robot's dexterity on hardware the model has never seen.
Key takeaways
- Trained on 1,000,000 hours of egocentric video (roughly 170 years of continuous human experience)
- 87% pass rate in zero-shot customer deployments, versus 46% for the earlier Dyna-1
- 1.55x higher success rate than Dyna-1 in controlled comparisons
- Scaling inflection point between 10,000 and 100,000 hours of footage
- Video generation latency dropped from 10,203 ms to 110 ms after distillation (90x faster)
A model that predicts video and motion at once
Dyna-2 is a world-action model (WAM) — a single generative model that predicts future video frames and robot actions at the same time. The architecture is a mixture-of-transformers built on a video-diffusion backbone with DiT layers. The action transformer is deliberately shallow and joins the video stream in early layers to cut latency without losing quality. The model was pre-trained purely on head-mounted camera recordings — cooking, tidying, folding clothes, assembling — from which wrist trajectories and a continuous grasp signal were extracted.
Dyna-2's architecture is a single generative model with three parts:
The key result: the model saw zero hours of robot data during pre-training, yet its zero-shot?zero-shot: A model performing a new task or on new hardware with no additional training or examples. scores across 39 tasks on two bimanual YAM platforms rose monotonically with the amount of human video. Dyna Robotics calls this the first human-to-robot transfer scaling law.
Video as the new scaling axis
The company stakes out a position directly opposed to the dominant Vision-Language-Action (VLA) models, which rely on expensive robot teleoperation?teleoperation: Collecting data by a human remotely controlling the robot — expensive and slow. data. According to the report, scaling video data alone — without action labels — monotonically improved generalization to the robot. Language-following ability rose from 35% to 67% thanks to video co-training, reaching 96% at full scale. The effect is clearest on hard tasks: on a lockbox key-turning task the model scored 0% through the first 100,000 hours, then 90% at one million hours. The company sums it up with the line that "video is the new scaling axis."
| Ability | Starting point | 1M hours of video |
|---|---|---|
| Instruction following | 35% | 96% |
| Lockbox key-turn | 0% | 90% |
Why it matters
If the result holds, it changes the economics of teaching robots. Teleoperation data is expensive and slow to gather, while first-person human video exists in near-unlimited supply. A scaling law means progress can be bought with a larger video corpus, not only a new architecture. That shifts the competitive edge toward companies with access to large footage archives and cheap compute. The approach also undercuts the assumption that dexterity requires data from the specific robot.
What's next?
- The full technical report with scaling curves and formulas is available at dyna.co/dyna-2
- Tests included WUJI-2 hands with 20 degrees of freedom and a semi-humanoid prototype — pointing toward multi-finger manipulation
- Key caveat: the results come from a company report, not an independently peer-reviewed publication
Sources
- Humanoids Daily — Dyna-2 Proves Scaling Laws for Robotics: 1 Million Hours of Human Video Unlocks Zero-Shot Dexterity
- Dyna Robotics — Dyna-2 Technical Report
