Dyna Robotics introduced Dyna-2, a world-action model for robot manipulation, pre-trained on more than one million hours of egocentric human video. That equals roughly 170 years of continuous waking experience. The team tested whether ordinary human video can stand in for costly action-labeled data. They built a “ladder” from 1,000 to 1,000,000 hours and measured what scales. The results include a scaling law on human data, its first transfer to unseen robot data, and evidence that video prediction drives the transfer.
Is Dyna-2 deployable today?
Yes, but only as a vendor-operated system, not as downloadable weights. The company has announced no public checkpoints, API, or licenses for Dyna-2. So deployment means buying a Dyna robot cell rather than self-hosting the model.
Which companies benefit most? Dyna-1 already runs in hotels, restaurants, and laundromats, per the company. That points at mid-market service operators and multi-site enterprises with repetitive, stationary manipulation work. It is not a fit for solo builders or research labs seeking local inference.
The target industries are straightforward. Hospitality, commercial laundry, food service, light assembly and kitting, and facilities cleaning are in scope. The platform addresses scenarios where reliable, repeatable manipulation yields the strongest productivity gains.
The 14 post-training tasks map cleanly to real work. They include trash tray clearing, first-aid kitting, tote construction, food scooping, rope tying, hanger preparation, and targeted drink retrieval from a fridge. This set highlights applied capabilities rather than demos.
Deployment today: no weights and no public API, only a vendor-run Dyna robot cell.
What is Dyna-2 and how does a world-action model work?
Dyna-2 is a world-action model (WAM): a single generative model that denoises future video and a future action chunk, jointly or separately, on a video-diffusion backbone. It was pre-trained on more than one million hours of egocentric human video, roughly 170 years of waking experience.
Architecturally it is a mixture of transformers. Video and action are tokenized separately and receive distinct DiT layer stacks that attend to each other. Proprioception feeds directly into the action transformer, providing the manipulator’s internal state.
Attention regimes differ across modalities. Video tokens use causal masking; action tokens use bidirectional self-attention and attend to contextual video tokens. Video tokens also cross-attend to text, but text does not directly influence action tokens. This decouples visual state rollout from policy computation.
Training uses flow matching. A video loss and an action loss share a trunk as two separate marginal velocity fields. Because the action network never takes the noised video latent as an argument, the policy stays reactive at inference — it neither generates nor attends to predicted future video. The action transformer is deliberately shallower and joins the video stream early, which the team says improves real-time latency without costing performance.
Key idea: learn video and actions together, yet keep policy reactive and independent of generated video.
What scaling results did the team demonstrate?
The team cut nested subsets at 1,000, 10,000, 100,000, and 1,000,000 hours with identical source proportions. A larger budget only added data, so curve differences are not due to distribution shift. A fixed, disjoint 100-hour validation set scored every rung on the ladder.
First: a scaling law holds on human data to one million hours. All four metrics improve monotonically and fit power laws: held-out MSE = 0.0691·D^-0.0184 (R²=0.919), accuracy@0.5 = 0.357·D^+0.0203 (R²=0.865). Across the ladder, accuracy@0.1 rises 51% versus 12% for MSE, indicating different sensitivities of the metrics to scale.
Second: that law transfers to robot data the model never saw. The same checkpoints were scored zero-shot on 39 tasks across two stationary bimanual YAM platforms — 12 internal, 27 from xdof ABC. Zero-shot action MSE = 0.306·D^-0.0713 (R²=0.884). The team reports an inflection between 10k and 100k hours, after which gains accelerate.
Third: the objective matters, and video is a separate axis. Joint denoising beat action-only on 39 of 39 tasks at every action scale. With action-labeled data fixed at 50,000 hours, adding video-only hours drops zero-shot robot MSE from 0.340 to 0.120. Notably, held-out human error does not improve — the benefit from video is specifically cross-embodiment generalization.
The scaling law holds, transfers to robots, and is powered by joint video-plus-action denoising.
How did Dyna-2 perform directly on robots?
Each rung was post-trained on 14 tasks, at most 10 hours of robot data per task, across three embodiments: 6-DOF YAM arms with parallel-jaw grippers, the same arms with WUJI-2 20-DOF dexterous hands, and a semi-humanoid prototype. Post-training used robot data only — no human-robot alignment and no co-training.
Mean normalized score rose 20% → 28% → 45% → 53% across the ladder, with best results on 9 of 14 tasks at one million hours. Lockbox Key Turning is the threshold case: 0% up to 100,000 hours, then 90%. Bottle Cap Untwisting was post-trained on roughly 10 minutes of demonstrations and still climbed to 50%.
Against Dyna-1 — the company’s production VLA initialized from Qwen3-VL-4B — an early Dyna-2 reached 1.55× success rate and 1.12× grade, pooled over 7 tasks and 3 checkpoints. At unseen customer sites, Dyna-2 passed production criteria 87% versus Dyna-1’s 46%, though both pass near 100% in house.
In addition, a distillation pipeline cut video sampling time from 10,203 ms to 110 ms on one H100. This directly supports real-time behavior and reduces compute costs without sacrificing performance, according to the team’s results.
Limited robot post-training plus distillation delivered a leap: 53% mean score and 110 ms video sampling.
What are the key takeaways and next steps?
The bottom line is clear: Dyna-2 is a world-action model pre-trained on 1M+ hours of egocentric human video. Scaling laws hold on held-out human data across four orders of magnitude, and, for the first time, that law transfers zero-shot to robot data never seen in pre-training. This is a structural signal for the direction of robotics.
The main driver of cross-embodiment generalization is video co-training rather than more action labels. The practical implication is that cheaper video datasets can improve policy transfer even when metrics on human sets do not move. This opens rational scaling without a linear rise in teleoperation costs.
On deployment, the picture is unambiguous. No weights and no public API: today it is a vendor-operated system shipped as a Dyna robot cell. Therefore, near-term beneficiaries are mid-market service operators and networked enterprises with repetitive, stationary manipulation.
The task list, industries, and metrics already map to real work. If you operate in hospitality, commercial laundry, food service, light assembly, or cleaning, the 14-task map shows a direct path from experiment to operations.
Based on the original source.
