Dyna Robotics Unveils Dyna-2 World-Action Model Pre-Trained on 1 Million Hours of Human Video

Dyna Robotics introduced Dyna-2, a world-action model designed for robot manipulation, pre-trained on more than one million hours of egocentric human video, equivalent to roughly 170 years of continuous waking experience. The model employs a mixture of transformers with separate tokenization and processing of video and action, using a video-diffusion backbone with flow matching. The company reports three key scaling results: a scaling law holds on human data up to one million hours, this law transfers zero-shot to unseen robot data across 39 tasks, and video prediction, rather than action data, drives cross-embodiment generalization. On-robot post-training on 14 tasks across three embodiments showed mean normalized scores increasing from 20% to 53% across the data ladder, with specific improvements such as Lockbox Key Turning reaching 90% success. Dyna-2 is not available as downloadable weights; deployment requires purchasing a Dyna robot cell. The model outperformed Dyna-1, the company's production VLA, in on-robot tests and at unseen customer sites, where it passed production criteria 87% of the time versus 46% for Dyna-1. The technical report is available for further reading.