Dyna Robotics unveiled DYNA-2 on August 10, 2026, a world-action foundation model pre-trained on one million hours of human egocentric video — roughly 170 years of continuous waking experience and, the company claims, the first robot foundation model at that data scale.
The Scaling Claim
DYNA-2 is pre-trained using a dual next-frame and next-action world-modeling architecture. Dyna reports a smooth scaling law across four orders of magnitude, from 1,000 to 1,000,000 hours of human data, with no observed plateau — the first published human-to-robot scaling curve. Notably, the model was pre-trained without a single frame of robot data; robotic manipulation performance emerges after a much smaller fine-tuning step on real hardware.
Why Human Video?
Robot demonstration data is expensive to collect and platform-specific. Human egocentric video is comparatively abundant and captures the same physics and object interactions robots ultimately need to reproduce. DYNA-2 treats humans as the pre-training substrate, letting physical intuition and spatial reasoning transfer to robot hardware — an approach echoed by, but ahead of, similar work at NVIDIA and Meta.
Reported Performance
Independent write-ups cite roughly 90% success on evaluated manipulation tasks after fine-tuning, with performance improving predictably as the human-video pool grows — a departure from the plateau-heavy behavior of prior robot foundation models trained on smaller, robot-only corpora.
Why It Matters
If the scaling law holds, DYNA-2 changes the economics of building generalist robots: instead of asking hardware makers to collect billions of teleoperation hours, model builders can lean on human video that already exists on the open web. That is a very different roadmap than the one behind Google DeepMind's Gemini Robotics 2 or NVIDIA's Cosmos 3 stack, and it hands smaller robotics companies a shot at competitive foundation models without hyperscaler data budgets.
Reporting based on coverage from PR Newswire, Interesting Engineering and The AI Insider.