Dyna Robotics just released DYNA-2, a world-action model trained only on more than one million hours of human egocentric video. That equals about 170 years of continuous waking time. No robot action data went into the pre-training stage. The company says this scale alone lifted high-precision manufacturing task success from 20 percent to 80-90 percent. The model transfers to stationary arms, humanoid prototypes, and multi-fingered hands after just a few hours of local fine-tuning. In one test, 13 minutes of data let a pair of five-fingered hands twist open a bottle cap. Across 15 benchmarks, more human video meant better results. In a real customer deployment, DYNA-2 hit an 87 percent quality pass rate versus 46 percent for the earlier DYNA-1 model. It also recovered from physical disruptions during tasks like chopping food without needing human help. The robots already run in hotels, restaurants, and laundromats. Founders include former DeepMind researcher Jason Ma.
Dyna Robotics just released DYNA-2, a world-action model trained only on more than one million hours of human egocentric video. That equals about 170 years of continuous waking time. No robot action data went into the pre-training stage. The company says this scale alone lifted high-precision manufacturing task success from 20 percent to 80-90 percent. The model transfers to stationary arms, humanoid prototypes, and multi-fingered hands after just a few hours of local fine-tuning. In one test, 13 minutes of data let a pair of five-fingered hands twist open a bottle cap. Across 15 benchmarks, more human video meant better results. In a real customer deployment, DYNA-2 hit an 87 percent quality pass rate versus 46 percent for the earlier DYNA-1 model. It also recovered from physical disruptions during tasks like chopping food without needing human help. The robots already run in hotels, restaurants, and laundromats. Founders include former DeepMind researcher Jason Ma.