Dyna Robotics Unveils Dyna-2: A World-Action Model Trained on 1 Million Hours

Dyna Robotics has officially introduced Dyna-2, a cutting-edge world-action model (WAM) designed to denoise future video and upcoming action chunks either jointly or separately using a video-diffusion backbone. The system was pre-trained on more than one million hours of egocentric human video, which roughly equals 170 years of continuous waking human experience.

From an architectural standpoint, the model integrates a mixture of transformers where video and actions are tokenized separately. They are processed through distinct DiT layer stacks that attend to each other, while proprioception feeds directly into the action transformer. Video tokens utilize causal masking, and action tokens employ bidirectional self-attention.

The training procedure implements flow matching, sharing a trunk as two separate marginal velocity fields for video and action losses. Because the action network never processes noised video latents as an input argument, the policy remains highly reactive during inference. The action transformer is built intentionally shallower to improve real-time latency without sacrificing overall task performance.

Stock photo for illustration only, not from the actual event

The engineering team structured nested subsets containing exactly 1,000, 10,000, 100,000, and 1,000,000 hours, maintaining identical proportions across all sources. Across the scaling ladder, mean normalized scores increased progressively from 20% to 28%, 45%, and finally 53%, performing best on 9 out of 14 tasks at the one-million-hour threshold.