Dyna-2 World-Action Model Scales Robot Learning to 1M Hours

Dyna Robotics published its technical report for Dyna-2 on August 10, 2026, releasing an architecture description and three quantified scaling results built from a data ladder spanning 1,000 to 1,000,000 hours of egocentric human video. The core claim is that scaling laws previously observed in language and vision transfer to robot manipulation — first on held-out human data, then zero-shot to robot platforms the model never encountered during pre-training. For practitioners tracking the agentic era's implications for embodied systems, this is one of the first results to put hard power-law exponents on that transfer.

Dyna-2 ships without public weights, an API, or a license. It is a vendor-operated system tied to Dyna robot cells already running in production in hotels, restaurants, and laundromats.

Architecture: One Trunk, Two DiT Stacks

Dyna-2 is a world-action model (WAM): a single generative model that denoises future video and a future action chunk simultaneously, trained under flow matching as two separate marginal velocity fields sharing one trunk. Video tokens and action tokens each have distinct Diffusion Transformer (DiT) layer stacks that cross-attend to each other. Proprioception feeds directly into the action transformer. Video tokens use causal masking; action tokens use bidirectional self-attention and attend to context video tokens. Text cross-attends to video tokens only — it does not directly influence action tokens.

Because the action network never takes the noised video latent as an argument, the policy remains reactive at inference rather than waiting on video generation. The action DiT is shallower and joins the video stream early; Dyna Robotics reports this improves real-time latency without a performance penalty. A distillation pipeline compresses video sampling from 10,203 ms to 110 ms on a single H100 — roughly 90× — holding FVD at 121 and flicker at 1.94 against 2.37 for real recorded footage.

The Three Scaling Results

Human-data scaling law. Dyna Robotics cut nested subsets of exactly 1,000, 10,000, 100,000, and 1,000,000 hours, maintaining identical source proportions at each rung. A fixed, disjoint 100-hour validation set scores every rung. All four metrics 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 full ladder, accuracy@0.1 rises 51% while MSE improves only 12% — tight-threshold accuracy captures scale gains that aggregate error metrics understate.

Zero-shot transfer to robot data. The same four checkpoints were scored on 39 tasks across two stationary bimanual YAM platforms — 12 internal, 27 from xdof ABC — with zero robot trajectories in pre-training. Zero-shot action MSE fits a power law: 0.306·D^−0.0713 (R²=0.884), with a reported inflection between 10,000 and 100,000 hours. Each rung was then post-trained on 14 tasks using 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 or co-training. Mean normalized score rose from 20% at 1,000 hours to 28%, 45%, and 53% at successive rungs. Lockbox Key Turning was the threshold task: 0% at every checkpoint up to 100,000 hours, then 90% at one million. Bottle Cap Untwisting, post-trained on roughly 10 minutes of demonstrations, reached 50% at the top rung.

Video prediction as an independent axis. Holding action-labelled human data fixed at 50,000 hours and adding video-only hours with no action labels drives zero-shot robot MSE from 0.340 to 0.120. Held-out human error does not improve across the same sweep — the gain is exclusively cross-embodiment generalization. Joint denoising beat an action-only objective on all 39 tasks at every action data scale tested.

Head-to-Head Against Dyna-1

The comparison used matched pre-training data, post-training data, and hyperparameters across three distinct pre-training checkpoints. Dyna Robotics explicitly frames the result as a lower bound for the WAM because the evaluation pipeline was tuned for the VLA. The 41-percentage-point site-transfer gap — 87% versus 46% on production criteria graded by on-site operators not involved in model development — carries the most direct commercial signal.

The power-law transfer exponent on zero-shot robot MSE (−0.0713, R²=0.884) is measurably steeper than the human-data exponent (−0.0184), suggesting robot generalization benefits from scale more than in-domain human prediction does. Whether that curve extends beyond one million hours, or holds for mobile or contact-rich manipulation, remains open — but Dyna Robotics has given the field concrete numbers to target.