Dyna-2 uses one million hours of human video to test a new scaling path for robot learning

Dyna Robotics has introduced Dyna-2, a world-action model pretrained on more than one million hours of first-person video showing people performing everyday physical tasks. The company reports that increasing the amount of human video improved predictions not only on held-out human examples but also on robot data the model had never seen during pretraining. Its central claim is a human-to-robot transfer scaling law: more human experience in the training set produced progressively better results across the embodiment gap.

Robot-learning teams often gather data through teleoperation or specialized capture equipment. Those methods provide valuable action labels, but every hour must be deliberately recorded with a robot or controlled device. Dyna’s alternative begins with egocentric video of people cooking, tidying, folding and assembling objects. The company built a cleaning and hand-pose extraction pipeline that turns suitable footage into wrist trajectories and grasp signals. It then created nested training sets of exactly 1,000, 10,000, 100,000 and one million hours so each larger run added data without replacing the earlier distribution.

Dyna-2 jointly models future video and future actions using a video-diffusion backbone and a mixture of transformers. Video and action inputs have distinct layers that can exchange information, while language instructions condition the video stream. The design lets the video objective shape a shared representation of how scenes and objects change, even though the deployed policy does not need to generate a future video before acting. Dyna says that shallower action layers connected early to the video stream improved real-time latency without sacrificing performance in its architectural tests.

On held-out human data, the company reports monotonic improvement across continuous error measures and thresholded accuracy as training video grew. To examine transfer, it evaluated the same checkpoints on 39 tasks drawn from two stationary, bimanual YAM robot platforms. The tasks included cloth handling, knot tying, packing, cleaning, food service and assembly. None of the checkpoints had trained on a trajectory from that robot evaluation set. Dyna says every metric ranked monotonically with the scale of human pretraining, with an inflection between 10,000 and 100,000 hours.

The researchers then post-trained the four checkpoints on the same robot-task datasets and ran blind physical tests across 14 tasks and three embodiments. Mean normalized performance rose from 20 percent at the smallest human-video scale to 28 percent, 45 percent and 53 percent as the scale reached one million hours. The largest model was best on nine of the 14 tasks. A lockbox key-turning task remained unsolved through the 100,000-hour checkpoint but reached a 90 percent success rate at one million hours, according to the company’s results.

The data-efficiency example that draws the clearest line between human video and robot learning involved two five-fingered hands opening a bottle cap. Dyna used roughly 10 minutes of robot demonstration data for post-training, while performance increased from 10 percent at smaller pretraining budgets to 40 percent and then 50 percent as human-video scale grew. Targeted drink retrieval also rose from 58 percent to 83 percent. These are company-run experiments, and independent replication across other laboratories, robots and data sources will be important before treating the relationship as universal.

Controlled comparisons in the report suggest that predicting future video is doing more than adding training volume. A joint video-and-action objective beat action-only training on all 39 offline robot tasks at every tested action-data scale, while adding video-only data produced further monotonic gains in cross-embodiment generalization. That result supports a practical path for robotics: abundant human video could provide broad physical knowledge, then a smaller amount of robot-specific data could adapt the model to a particular body and task. The remaining questions concern data rights, the diversity and quality of recorded activity, performance outside the tested manipulation settings and whether the reported curve continues as models, compute and robot forms change.