Dyna-2's Million-Hour World-Action Model

What Dyna Robotics announced

On August 11, 2026, Dyna Robotics introduced Dyna-2, a world-action model pre-trained on more than one million hours of human manipulation video. The company reports monotonic improvement as pre-training data grows from 1,000 to 10,000, 100,000, and one million hours.

Dyna-2 uses a video-diffusion backbone with separate video and action token stacks. It is trained to predict future video and future action chunks, while proprioception and language provide additional context. Dyna says this shared world-and-action objective improves transfer without requiring the policy to generate future video during deployment.

• More than one million hours of mostly head-mounted, first-person human manipulation video
• Nested training subsets at 1,000, 10,000, 100,000, and 1,000,000 hours
• A fixed 100-hour held-out human validation set
• Zero-shot offline evaluation on 39 robot tasks across two bimanual robot data sources

The result that matters crosses the embodiment gap

Dyna evaluated the human-only pre-trained checkpoints on robot trajectories that were absent from pre-training. Across 39 manipulation tasks, the company reports that offline robot action-prediction metrics improved monotonically with the amount of human video. That is a stronger claim than showing that a model gets better at predicting held-out human motion.

The company then applied the same robot-only post-training recipe to checkpoints from each data rung. Across 14 physical tasks, mean normalized performance rose from 20% at 1,000 hours to 53% at one million hours. The largest model was best on nine of the 14 tasks. These policies still required task-specific robot demonstrations, so the result supports data-efficient transfer rather than zero-shot mastery of every robot task.

World modeling, not action prediction alone, unlocked the trend

The sharpest experiment fixes the architecture and action-labeled data, then changes the objective. Dyna compared action-only training, joint future-video and action prediction, and joint training with additional video-only data. It reports that the joint objective beat action-only training on all 39 robot tasks at every tested action-data scale.

More importantly, the cross-embodiment scaling trend appeared only when abundant video was added through the future-prediction objective. With action-labeled data held fixed, increasing video-only data produced monotonic improvement on held-out robot evaluation. This suggests that unlabeled interaction video can add value, but only when the model is explicitly trained to learn how scenes evolve.

The data pipeline is the real moat

Dyna describes a corpus dominated by first-person recordings of cooking, tidying, folding, assembling, and other everyday manipulation. Clips that pass its quality bar receive 3D hand-pose tracks. Wrist motion becomes pseudo-action supervision for end-effector trajectories, while thumb-index aperture becomes a continuous grasp signal.

That pipeline turns passive pixels into approximate action trajectories. It also exposes the next bottleneck. Recording video is easier than producing reliable pose tracks, synchronized pseudo-actions, task labels, provenance, and rights evidence at million-hour scale. Dyna explicitly notes that annotation quality and processing capacity lag behind raw video collection.

• Keep raw video, derived pose, pseudo-action, and confidence as separate layers
• Preserve source, task, environment, capture system, and transformation provenance
• Validate hand visibility, pose continuity, timing, grasp semantics, and failure cases
• Treat video-only and action-labeled hours as different inventory classes

What the announcement does not prove

The evidence is a Dyna Robotics technical post, not an independently reproduced release. The company does not provide public model weights, the million-hour corpus, or enough sample-level documentation to audit its source mix, rights, annotation error, and coverage claims. Its customer-site pass rates use company-reported acceptance criteria.

Offline action prediction is also not the same as closed-loop task success. Dyna supplies a separate post-training study for real robots, and individual tasks do not improve monotonically even though the 14-task aggregate does. The responsible conclusion is that Dyna has presented unusually strong evidence for human-video scaling, not that robot data, careful evaluation, or embodiment-specific adaptation are obsolete.

What data teams should do next

Build data programs around two linked scaling axes: broad interaction video for learning world dynamics, and smaller action-rich datasets for grounding, post-training, and evaluation. Do not collapse both into one hours figure.

For buyers, the minimum useful release record now needs to describe capture viewpoint, task and object coverage, pose-extraction method, pseudo-action semantics, confidence, timing, source rights, held-out splits, robot embodiments, and the exact evaluation protocol. One million undocumented hours is a marketing number. One million traceable hours with measurable transition coverage is infrastructure.