Dyna Robotics' Dyna-2 robot foundation model

Dyna Robotics released Dyna-2, a World-Action Model pre-trained on over one million hours of egocentric human video, replacing the Vision-Language-Action architecture of its predecessor Dyna-1 with a model built on video generation. The model uses a dual training objective, predicting next frames and next actions simultaneously, which Dyna Robotics says enables cross-embodiment transfer across stationary robot arms, humanoid prototypes, and dexterous robotic hands. In customer deployments at matched post-training budgets, Dyna-2 recorded an 87% quality pass rate versus 46% for Dyna-1, completing tasks 1.55 times more often. Across 15 benchmark tasks, models pre-trained on more human video consistently outperformed those trained on less, reaching 90% task success. In zero-shot deployment at new customer sites without site-specific retraining, the model achieved 87%. In high-precision manufacturing evaluations, task success rates climbed from roughly 20% to between 80% and 90% as pre-training data volume increased with the post-training dataset held constant. The model recovered from physical disturbances during tasks like food chopping without human intervention, while the VLA baseline required manual recovery. With 13 minutes of fine-tuning data, Dyna-2 learned to command five-fingered robotic hands to twist open a bottle cap. The measured scaling law spans four orders of magnitude, from 1,000 to 1,000,000 hours of human video, and Dyna Robotics says performance improves smoothly with each additional hour without plateauing. The company is targeting 10 million hours as its next data scale milestone. Dyna-2 follows a model lineage that began with Dyna-1 in June 2025.