Dyna Robotics Introduces Dyna-2: A World-Action Model Pre-Trained on 1 Million Hours of Human Video

Details

Dyna-2 is architecturally a world-action model (WAM): a single generative model that jointly denoises future video and future action chunks using a video-diffusion backbone. Pre-training used 1M+ hours of human egocentric video; the team trained data ladders from 1,000 to 1,000,000 hours to measure what scales. Three findings emerge: a scaling law on human data, the first transfer of that law to unseen robot data, and evidence that video prediction drives the transfer.\n\nThe architecture uses a mixture of transformers with separate tokenization for video and action. Video tokens use causal masking; action tokens use bidirectional attention. Proprioception feeds directly into the action transformer. At inference time, the policy does not generate or attend to predicted future video — making it reactive rather than planning-based, which improves real-time latency.\n\n14 post-training manipulation tasks were demonstrated: trash tray clearing, first-aid kitting, food scooping, rope tying, hanger preparation, and targeted drink retrieval. Deployment today requires purchasing a Dyna robot cell — no public weights, API, or license exist. Dyna-1 robots already operate in hotels, restaurants, and laundromats, giving Dyna-2 an immediate commercial fleet to upgrade.