One of the most star-studded “fusion” papers of the year in video generation and embodied intelligence has just appeared.
A few days ago, a paper titled Masked Visual Actions for Unified World Modeling landed on arXiv. It has not yet spread through the academic circuit, but it is about to blow up.
The author list is the story. It includes Fei-Fei Li, the creator of ImageNet and now all-in on spatial intelligence; Jiajun Wu, the Stanford assistant professor who is a leading figure in physical reasoning and a fresh winner of the IJCAI 2026 Computers and Thought Award; Lvmin Zhang, author of ControlNet; Gordon Wetzstein, who heads Stanford’s computational imaging lab; and Wenlong Huang, a core driver of robot foundation models, among more than a dozen top researchers.
The bigger surprise: Yilun Du, a Harvard assistant professor, is also on it. Li and Du are the figureheads of two rival embodied-AI camps, one arguing that AI should think inside visuals it can understand, the other that everything should be compressed into abstract numbers first. They last co-authored five years ago. Seeing them on the same paper is itself the headline.
In academia, certain names together are a signal. Two routes that normally refuse to yield appeared in one paper. The work may have found the crossing point both camps accept.
Here is the uncomfortable consensus it proposes: robots may not need their own dedicated large model at all. MVA suggests a robot does not even need a retrained video model. Find the right interface and you can use a mature video-generation model directly.
Video models today understand physics but cannot issue commands. The gap is translation. A robot speaks a dry language of joint angles and centimetres; a video model speaks in pixels. Past methods forced the video model to learn a foreign tongue. MVA instead speaks the video model’s native language.
The method is disarmingly simple, three steps.
First, turn action into a coloured video. Using SAM (Segment Anything Model, Meta’s segmentation tool), the robot and the object are separated into masks that move frame by frame, forming two motion trajectories. That is MVA’s entire definition of action: no angles, no coordinates, just colour blocks travelling through the frame.
Second, mask one trajectory and let the model fill it in. Hide the object, keep the arm, and ask what happens next: the model acts as a world simulator. Hide the arm, show a cup sliding, and it must infer how the arm moved: the model acts as an action generator. Same model, same weights, no switch. Which role it plays depends only on which colour you hide.
Third, no training from scratch. MVA stands on Alibaba’s open-source video model Wan2.2 (14 billion parameters, a diffusion transformer) and teaches it to read the coloured blocks with LoRA, a light fine-tuning method, using just 15 hours of data. Rivals spend hundreds of millions training a base model from zero plus thousands of hours of robot data. MVA borrows a free open model and 15 hours of data for zero-shot generalisation.
The numbers: on LPIPS, which measures how far generated frames drift from real ones (lower is better), MVA scored 0.0945 against 0.362 for the next method, Ctrl-World, nearly four times closer. Trained on a single-arm robot, MVA was then tested on a dual-arm robot it had never seen. Ctrl-World failed on the spot. MVA still produced sound predictions.
Industry readers care less about the clever cloze design and more about generalisation. Li Linxin, CTO and co-founder of Shenzhen’s Vitoka Robotics, which deploys world models and 3D reconstruction in logistics, puts it plainly: most vision-language-action policies output low-level actions bound to a specific body, so they break on a different machine. MVA adds a layer. It understands the environment first, then solves the joints through inverse kinematics, which is why its generalisation is so strong.
The key detail is URDF, the robot’s universal “spec sheet” listing every joint and link. Most models output joint angles directly and brittlely. MVA decides where the end-effector should go, then uses the machine’s own URDF and inverse kinematics to compute the joints. The instruction stays generic; execution adapts per machine. URDF also reins in the video model’s wild imagination, keeping every pose physically possible.
The industry trend this year is to split seeing from acting. Seeing is trained on near-infinite internet video; acting is filled in with far less robot data plus kinematics. MVA sits exactly at that interface.
Two diverging routes, it turns out, both bottom out at the same URDF and the same forward and inverse kinematics. Go deep enough and the answer is not either-or.
MVA argues robots need no dedicated action language, just an interface that translates action into the pixel language video models already speak. The deeper bet: embodied AI’s path forward may not be making robot-specific models bigger, but borrowing the strong vision models that already exist. Internet video is nearly endless. Robot real-world data is always scarce and expensive.
The best translation is not forcing the other side to learn your language. It is saying what you mean in the words they already understand. When the heaviest names in academia and a quiet shift inside industry point the same way, the direction is worth watching.
Li Linxin is CTO and co-founder of Vitoka Robotics; his work covers world models, AI vision and generalisable 3D reconstruction.
Source: Leiphone (leiphone.com). Translated and adapted for The Robot Belt.