World-action model from Dyna Robotics pre-trained on over 1 million hours of egocentric human video for robot manipulation.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Dyna-2 is a world-action model (WAM) from Dyna Robotics, pre-trained on more than one million hours of egocentric human video. It jointly predicts action chunks and future video, and is presented in a research post titled "A 1-Million-Hour Scaling Law for World-Action Models".
The authors report a scaling law on held-out human data across a ladder from 1,000 to 1,000,000 hours, and what they describe as the first human-to-robot transfer scaling law: more human data in pre-training improves offline prediction on robot data never seen during pre-training. They also report that this carries through post-training to on-robot performance across 14 tasks, using a few hours of robot demonstrations per task.
Ablations in the post indicate that world modeling (joint video and action denoising) and video co-training matter for the cross-embodiment scaling to appear, and the post includes a comparison against a VLA architecture. The page describes research results; no public weights, API, code or pricing are mentioned in the provided material.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
World action model trained on over 1 million hours of first-person human video, showing a scaling law where more human data improves performance on unseen robot data. Adapts to robots with little robot data.
Closest neighbours by meaning, not just by category.
Model
Action-conditioned robotics world model fine-tuned from Cosmos-Predict2.5-2B using 2D skeleton conditioning across embodiments.
Model
Real-time interactive world model with compressed memory and hybrid action and semantic control, built on Wan 2.2.
Model · Xiaomi
Open-weights Vision-Language-Action robot foundation model from Xiaomi, pre-trained on 100K+ hours of UMI trajectories.
Research
Hierarchical humanoid loco-manipulation framework that chains meta-skills via contact flow, accepted to CoRL 2026.
Research · BitRobot Foundation
Open dataset of 500+ hours of whole-body teleoperated Unitree G1 humanoid data recorded in real homes.
Model · Black Forest Labs
7B open-weights world action model from Black Forest Labs that jointly predicts video and actions for robot control.