Dyna-2: Scaling Laws for World-Action Models at the 1 Million Hour Scale

Dyna-2 is a World-Action Model (WAM) pre-trained on over 1 million hours of first-person human video. It adopts an approach that learns scene changes, object contact responses, and hand-object interactions by observing tasks performed by humans. 1 million hours corresponds to approximately 170 years of continuous human activity data.

In experiments scaling human video training data from 1,000 hours to 1 million hours, Dyna-2 showed a consistent scaling trend where performance on held-out human data improved. This confirms existing scaling laws where increased training volume on human data leads to improved model performance.

Notably, performance improvements were observed on robot data despite no robot data being used during pre-training. As the volume of human video training increased, the MSE for robot action prediction decreased and [email protected] increased, which the authors explain as the first demonstration of the existence of human-to-robot transfer scaling laws.

Dyna Robotics argues that sensorized videos capturing humans performing actual economic activities are more suitable as pre-training data for robot learning than data created via teleoperation or specialized capture equipment. This is because such data can be acquired virtually unlimitedly and is directly connected to the tasks that general-purpose robots must perform.

However, an embodiment gap remains between humans and robots due to differences in morphology and implementation. Given the confirmation of human-to-robot transfer scaling laws, the researchers propose that future efforts should focus on continuously expanding human video data while simultaneously developing robot morphologies and capabilities to be closer to those of humans.