Dyna Robotics has unveiled Dyna-2, a world-action model for robot manipulation trained on over one million hours of egocentric human video, demonstrating strong zero-shot performance across 39 robot tasks and 14 post-trained real-world applications. The research establishes a scaling law from 1,000 to 1,000,000 training hours, showing monotonically improving metrics on human video and successful transfer to unseen robot platforms, with joint video-action denoising outperforming action-only training on every evaluated task. Dyna-2 operates as a vendor-run system with no public weights or API, targeting mid-market service operators in hospitality, food service, cleaning, and light assembly, where brief on-robot fine-tuning yields substantial reliability gains over the earlier Dyna-1 model.
This update represents a notable development in the Ai sector. Organizations and founders tracking this space should evaluate potential strategic and technical implications on their operations.