Most of the AI systems that move robots today began life doing something else. They started as models trained to generate video, or to caption images, or to answer questions about pictures, and were then fine tuned to output actions for an arm or a gripper. Robbyant, an embodied AI company inside Ant Group, argues that this borrowed lineage is the problem. This week it released LingBot-VA 2.0, which it describes as the industry’s first embodied native video action world model for robot control, meaning a model designed from the ground up to steer machines through the physical world rather than one built for digital content and later pointed at robots.
It helps to be plain about what a world model for robot control actually is. A robot that acts usefully has to carry around some internal sense of how a scene will change when it does something. Reach for a cup and the cup should move, the shadow should shift, the arm should occupy new space. A world model is the part of the system that predicts that unfolding, and a video action version predicts both the way the scene changes and the action to take in the same breath. The distinction Robbyant is pressing is between a model that has learned to make convincing looking video and one that has learned how physical dynamics behave. A pretty clip is not the same as an accurate forecast of what a real arm will do, and the company says building for the latter from the start is why it uses the label embodied native.
According to Robbyant and the accompanying paper, LingBot-VA 2.0 rests on a handful of design choices. At its center is what the company calls a semantic visual action tokenizer, which places world states and latent actions into one shared semantic space anchored to a language aligned visual foundation model, an arrangement meant to tighten the alignment between what the robot sees, the words that describe it, and the moves it makes. On top of that sits a strictly causal pretraining scheme using an autoregressive architecture, so the model predicts the future in order rather than peeking at it, along with a mixture of experts design that routes different situations to different specialized parts of the network, and an enhanced asynchronous inference setup aimed at keeping control fast. Robbyant also says the model absorbs general knowledge from web scale video pretraining, which is the source it credits for the system’s ability to generalize, and that it operates in a foresight reasoning style, thinking a step ahead about what will happen while it acts on live observations.
The numbers Robbyant puts forward are the headline. The company reports real time inference at up to 150 Hz on a single GPU, a rate it frames as answering a persistent complaint that embodied world models are too slow to run a robot in the moment. It also says the model can pick up new tasks with as few as roughly 20 demonstrations, learning them in context without retraining or updating its weights. A writeup by MarkTechPost lays out the same framing, presenting LingBot-VA 2.0 as a causal video action model built natively for physical AI. These figures are worth reading as claims from the maker rather than settled facts. Robotics has a long history of impressive laboratory and benchmark results that thin out once a machine faces the clutter, lighting, and friction of a real room, and generalizing to genuinely new tasks remains one of the hardest problems in the field.
LingBot-VA 2.0 does not arrive alone. Robbyant presents it as the piece that completes a six model full stack for embodied AI, sitting alongside LingBot-Depth 2.0, LingBot-Vision, LingBot-VLA 2.0, LingBot-World 2.0, and LingBot-Video, a spread that reaches from perception through vision language action control to longer form world generation. The company has published a code repository and the paper, titled Native Video-Action Pretraining for Generalizable Robot Control, which will let outside researchers probe whether the results hold up. That openness matters, because the gap between a demo reel and a dependable robot is exactly where claims like these get tested.
What Robbyant is really betting on is an idea about where robot intelligence should come from. If the useful thing a robot needs is a grounded feel for physical cause and effect, then a model that learned to predict how the world moves may travel further than one that learned to produce good looking video and had actions added afterward. That is a reasonable bet and an unproven one. For now LingBot-VA 2.0 is a well specified argument, backed by a paper, code, and a set of company figures, that the way to control machines in the physical world is to build for that world from the first line rather than to inherit a model shaped for the screen.