Dyna Robotics releases Dyna-2 World-Action Model

Dyna Robotics announced Dyna-2, a World-Action Model (WAM) pre-trained on more than a million hours of egocentric human video, equivalent to about 170 years of continuous footage. The model tests whether ordinary human video without action labels can replace the traditional bottleneck in robotic learning: action data that require expensive and slow tele-operation. The company did not publish a public benchmark, API or license; deployment today means purchasing a Dyna robot cell rather than running the weights locally.

Architecturally, Dyna-2 is built as a mixture of transformers. Video tokens and action tokens are processed separately and receive distinct stacks of DiT layers that attend to each other; proprioception is fed directly to the action transformer. Video tokens use causal masking, action tokens use bidirectional self-attention and attend to the video tokens of the context. Video tokens perform cross-attention to text, but text does not directly affect action tokens. Training uses flow matching. Video loss and action loss share a common trunk with two separate marginal velocity fields. Because the action transformer never receives the video’s masked latent representation as an input, the policy stays reactive during execution and does not generate or attend to predicted future video frames. The action transformer is deliberately flattened and joins the video stream early, which, according to the team, improves real-time latency without harming performance.

The team constructed nested subsets of exactly 1,000, 10,000, 100,000 and one million hours, preserving identical proportions from each source so that curve differences do not stem from distribution shift. A fixed, detached validation set of 100 hours marked each stage. On human data a strong power law was observed: MSE decreases as ≈ 0.0691·D⁻0.0184 (R²=0.919) and [email protected] increases as ≈ 0.357·D⁺0.0203 (R²=0.865). Across the scale, [email protected] improves by 51 % versus only 12 % for MSE. The same law transfers zero-shot to robot data unseen by the model: 39 tasks on two fixed-base dual-arm YAM platforms, 12 internal, 27 × xdof ABC. Zero-shot action MSE = 0.306·D⁻0.0713 (R²=0.884), with a turning point between 10,000–100,000 hours.

The video stream acts as an independent axis, not merely an additive signal. Joint Denoising outperformed pure action training on all 39 tasks at every action-scale. Holding a constant set of 50,000 hours of labeled action data and adding only video hours reduces zero-shot robot MSE from 0.340 to 0.120. Thus, unlabeled video is not merely background noise; it constitutes a separate training axis that pushes transfer forward.

Dyna-1 has already been deployed in hotels, restaurants and laundries, according to the company's announcement on 10 August 2026. It targets medium-scale service operators and multi-site organizations with repetitive stationary manipulation jobs: hospitality, commercial laundry, food service, light assembly and packaging, facility cleaning. 14 post-training tasks map directly to real work: emptying trash trays, first-aid kits, building totes, gripping food, tying ropes, preparing coat racks, and retrieving a targeted beverage from a refrigerator. The product is not intended for a single robot or a research lab that wants local inference.