- Published on
A World-Action Model Pre-Trained on 1 Million Hours of Human Video
- Authors
- Name
- DeepHorus
Dyna Robotics has unveiled Dyna-2, a world-action model for robot manipulation, pre-trained on more than one million hours of egocentric human video—roughly 170 years of continuous waking experience. The model aims to overcome the bottleneck of action-labelled data, which teleoperation must deliberately produce. By testing whether ordinary human video can substitute, Dyna-2 could reshape how robots learn complex manipulation tasks.
Traditionally, robot learning has been limited by the scarcity of action-labelled data. Dyna-2 addresses this by using human video at scale. The research team trained a data ladder from 1,000 to 1,000,000 hours and measured what scales. Three key results emerged: a scaling law on human data, transfer of that law to unseen robot data, and evidence that video prediction drives the transfer.
All four metrics improved monotonically, fitting power laws such as held-out MSE = 0.0691·D^-0.0184 (R²=0.919). Zero-shot action MSE on 39 tasks across two stationary bimanual YAM platforms followed a similar scaling law, with an inflection between 10,000 and 100,000 hours. Joint denoising of video and action beat action-only on all 39 tasks. Adding video-only hours while holding action data fixed at 50,000 hours dropped zero-shot robot MSE from 0.340 to 0.120.
The benefits are specifically cross-embodiment generalization, not improved human video prediction. The 14 post-training tasks map to real work, including trash tray clearing, food scooping, and hanger preparation. However, Dyna-2 is not available as downloadable weights or via API. It ships as part of a vendor-operated Dyna robot cell, with no public checkpoint or license announced.
Dyna-1 robots already operate in production in hotels, restaurants, and laundromats, per the company’s August 10, 2026 announcement. This positions Dyna-2 for mid-market service operators and multi-site enterprises handling repetitive, stationary manipulation. Industries such as hospitality, commercial laundry, food service, and light assembly are primary targets, while solo builders and research labs seeking local inference are not.
The company claims the architecture—a mixture of transformers with a shallow action transformer joining the video stream early—improves real-time latency without performance loss. The scaling law suggests that more human video could further improve robot performance, but the closed deployment model limits independent verification. Researchers may need to wait for an open release to reproduce the results.
No public checkpoint, API, or license for Dyna-2 has been announced, signaling a commercial-first strategy. The success of the video-prediction approach could influence other robotics labs to explore large-scale pretraining on human video. For now, Dyna-2 remains a demonstration of what scaling human data can achieve, constrained by its vendor-operated delivery.
