Dyna Robotics unveiled DYNA-2, a robot foundation model trained on one million hours of first-person human video, and published research claiming the first human-to-robot scaling law spanning four orders of magnitude. DYNA-2 is built on a World-Action Model (WAM), which jointly predicts the next video frame and next motor command from the same representation, unlike the Vision-Language-Action (VLA) approach used by Physical Intelligence, NVIDIA, and Skild AI. In head-to-head testing against DYNA-1, its VLA-based predecessor, DYNA-2 completed tasks 1.55 times more often under matched training conditions. At one unnamed customer deployment site, DYNA-2 achieved an 87% pass rate while DYNA-1 achieved only 46%; both figures are company-reported. The foundational research is published at dyna.co. VLA models emerged as the dominant architecture around 2023 with Google's RT-2, followed by Physical Intelligence's π0 in 2024. NVIDIA's GR00T N1.7, released in April 2026, built EgoScale pretraining on 20,854 hours of human egocentric video into a VLA architecture, documenting a doubling of task completion rates as data scaled from 1,000 to 20,000 hours. DYNA-2 claims to have run the same experiment at 50 times the scale—one million hours versus 20,854—with a different