What Is a VLA Model, and Why Does Every Humanoid Company Suddenly Have One?

Vision-language-action models are the software layer letting robots turn a camera feed and a text instruction into motor commands. Here's how they work.

If you've followed humanoid robotics announcements over the last two years, you've probably run into the acronym VLA — vision-language-action. It shows up in nearly every major platform's software stack now, from open-source research models to the systems running on commercial factory floors.

What a VLA model actually does

A VLA model unifies three things that used to be separate systems: a vision backbone that understands what a camera is looking at, a language model that understands an instruction like "pick up the blue box," and an action head that outputs actual motor commands.

Older robotics stacks handled this as a pipeline — a perception system identifies objects, a planner decides what to do, and a controller executes it. VLA models collapse most of that pipeline into a single learned model trained end to end on demonstration data.

The models actually shaping the field

Google's RT-2 was an early, influential demonstration: fine-tune a large vision-language model to output discretized actions as text tokens, showing that web-scale pretraining transfers to robot control. It stayed closed-source and required TPU-class compute, though, which put real-time use out of reach for most teams.

OpenVLA, an open-source 7-billion-parameter model trained on hundreds of thousands of real robot demonstrations, showed that a smaller, openly available model could match or beat much larger closed systems on manipulation benchmarks — a strong signal that data curation matters more than raw parameter count.

Physical Intelligence's pi0 takes a different architectural approach: a smaller pretrained vision-language backbone paired with a separate, fast "action expert" module, aimed at getting high-frequency control output without running the entire language model at control-loop speed.

Why it matters, and where it breaks down

The appeal is generalization. A model trained on a wide enough set of tasks and objects can, in principle, handle something new without being explicitly reprogrammed — the same way a language model can answer a question it never saw verbatim during training.

In practice, generalization is still limited and highly dependent on how much and how varied the training data is. A lot of what looks like open-ended understanding in a demo is closer to strong interpolation within a well-covered training distribution than genuine novel reasoning.

That gap between demo and reliable deployment is the main reason you'll still see teleoperation quietly propping up a lot of "autonomous" manipulation footage — the VLA model handles the easy, well-rehearsed cases, and a human handles the rest.