Dyna Robotics unveils AI model trained on 1M hours of human video

Dyna Robotics has unveiled DYNA-2, a robot foundation model trained entirely on more than one million hours of first-person human video, an approach the company says could break the data bottleneck that has limited the development of general-purpose robots.

The Redwood City, California-based startup announced the model on Sunday, describing it as a "world action model" that learns how actions change the physical world by watching human activity rather than relying on teleoperation data collected from robot hardware.prnewswire+1

How DYNA-2 Works

The model combines two training objectives — predicting what the next video frame should look like and what action should follow — giving it information about spatial relationships, movement and how objects respond to contact. Unlike vision-language-action models that connect what a robot sees with what a person tells it to do, DYNA-2 centers on predicting how a physical scene will change before determining the robot's next movement.theaiinsider

"For years, generalist robotics has been choked by a data bottleneck: collecting physical teleoperation data manually simply cannot scale to general intelligence," co-founder Jason Ma said in a LinkedIn post announcing the model. "Action data is scarce, but video is everywhere."linkedin

Dyna said the dataset represents roughly 170 years of continuous waking experience. In tests, DYNA-2 raised task success rates in high-precision manufacturing from about 20% to 80%–90% through increased pre-training scale alone, without changing the post-training dataset. In one experiment, 13 minutes of robot-specific data was enough to teach two five-fingered robotic hands to twist open a bottle cap.interestingengineering+1

Scaling Law and Deployment

The company said it observed what it calls a "cross-embodiment transfer scaling law": as more human video was added to pre-training, robot performance improved in a predictable way, even on hardware the model had never seen. The model has been tested on stationary robot arms, humanoid prototypes and dexterous robotic hands.linkedin+1

In a zero-shot customer deployment, DYNA-2 achieved an 87% quality pass rate compared with 46% for the earlier DYNA-1 model. The newer model also recovered from physical disturbances during manipulation tasks without human intervention.theaiinsider+1

Dyna already uses its DYNA-1 model in robots deployed in hotels, restaurants and laundromats. The company raised $23.5 million in a seed round in May 2025.kddi+1