Dyna

Dyna Robotics builds general-purpose robots for physical work like hotel housekeeping, restaurant prep, and laundromat operations, where its earlier DYNA-1 model is already running in production. The company was founded by Lindon Gao and York Yang, who previously sold Caper AI for $350 million, along with former DeepMind research scientist Jason Ma, and is backed by CRV and First Round. This launch introduces DYNA-2, a world-action model pre-trained on over one million hours of egocentric human video, roughly 170 years of continuous waking experience. The reason this matters right now is the data bottleneck that has held back generalist robotics. Teleoperation footage is expensive and slow to collect, so most robot foundation models plateau quickly. DYNA-2 replaces that pipeline with a dual next-frame and next-action architecture trained on ordinary human video, and Dyna claims it demonstrates a smooth human-to-robot scaling law across four orders of magnitude of training data. In practice that means physical intuition and spatial reasoning learned from people transfer to robot hardware without the model ever seeing robot data during pre-training. The reported results give some sense of what changed between generations. In head-to-head evaluations against DYNA-1, DYNA-2 completed tasks 1.55 times more often, hit an 87 percent pass rate at a customer site where the prior model scored 46 percent, and lifted high-precision manufacturing success from 20 percent to the 80 to 90 percent range through pre-training alone. For founders and operators tracking the humanoid and manipulation space, this launch is worth reading as a bet that video pre-training, not more teleoperation, is the path to commercial-grade robot autonomy.