Sakhanda Wire
NVDA $152.56 -2.29% MSFT $505.79 -2.36% GOOGL $279.70 -1.42% META $618.14 -1.32% AMZN $256.03 -0.67%
← Back to the news

Dyna Robotics unveils Dyna-2 robot model trained on 1 million hours of video

Dyna Robotics has introduced Dyna-2, a world-action model for robot manipulation trained on more than one million hours of egocentric human video, equivalent to about 170 years of continuous waking experience. The company’s research tested datasets ranging from 1,000 to 1,000,000 hours to examine how performance changes with scale.

The technical report describes improvements on held-out human data and zero-shot evaluation across 39 tasks on two stationary, bimanual YAM platforms. Joint video-and-action denoising outperformed action-only training on all 39 tasks. With action-labelled data fixed at 50,000 hours, adding video-only data reduced zero-shot robot action MSE from 0.340 to 0.120. After post-training on 14 tasks, using no more than 10 hours of robot data per task, the mean normalized score increased from 20% to 53% across the data ladder.

An early Dyna-2 version achieved 1.55× the success rate of Dyna-1 across seven tasks and three checkpoints. Dyna-2 is currently offered as a vendor-operated system through a Dyna robot cell; no public checkpoint, API, or license has been announced. Dyna robots already operate in hotels, restaurants, and laundromats, according to the company.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news