Dyna Robotics Trains DYNA-2 on Human Video

Dyna Robotics has unveiled DYNA-2, a robotic foundation model pre-trained on more than one million hours of egocentric human video—equivalent to roughly 170 years of continuous waking experience. Unlike traditional models that rely on expensive teleoperated robot data, DYNA-2 utilizes a World-Action Model architecture. By running a dual objective to predict both the next video frame and the next action, the model develops spatial reasoning and contact physics that transfer to various physical embodiments, including stationary arms, humanoid prototypes, and five-fingered hands, without having seen robot data during pre-training.

Adapting DYNA-2 to specific hardware requires only hours of local fine-tuning. For example, the company used just 13 minutes of data to train a pair of five-fingered hands to twist open a bottle cap. On high-precision manufacturing tasks, scaling up the pre-training data alone boosted task success rates from 20% to between 80% and 90%. Additionally, a video co-training algorithm yielded a 133% improvement on instruction-following tasks requiring distinct motions.

In real-world head-to-head evaluations, DYNA-2 completed 1.55x more tasks than its predecessor, DYNA-1, which is currently deployed commercially in hotels, laundromats, restaurants, and gyms. At a customer deployment site, DYNA-2 achieved an 87% pass rate compared to 46% for DYNA-1. During dexterous tasks like clearing workspaces and chopping food, DYNA-2 successfully recovered from physical disturbances without human intervention, whereas DYNA-1 failed and required manual resets.

The company's existing commercial footprint serves as a data flywheel. Its older DYNA-1 model already folds more than 40 shirts per hour and operates 16 hours a day with a success rate of over 99% during 24-hour non-stop operations. Having demonstrated the viability of learning from human video, Dyna Robotics now plans to scale its training corpus to 10 million hours of video.