Dyna Robotics Unveils Dyna-2 Trained on 1 Million Hours of Video

Dyna Robotics is developing foundation models aimed at giving robots broader manipulation skills. Progress in the field has been constrained by the cost and scarcity of robot demonstrations labeled with actions, limiting the scale of conventional training. Dyna-2 takes a World-Action Model approach, using ordinary human video to learn how objects, people and environments change during physical interactions, potentially widening the pool of useful pre-training data.

Dyna Robotics unveiled Dyna-2 on Aug. 10, 2026, after pre-training the model on more than 1 million hours of human video. The company said its research showed that videos without robot-action labels can still help overcome the sector’s data bottleneck. Tests also indicated favorable scaling as training resources increased and transfer to new tasks and operating settings, supporting the case for human video as a practical foundation for robot-control models.