Until a few years ago, foundation models were mostly confined to generating text and images. Lately, though, they're being folded into robot control at a fast pace. Look at a camera feed, understand the situation, and output the next action directly — that's the core idea behind Vision-Language-Action, or VLA, models.
The idea of "turning what you see directly into motion" isn't new. But researchers broadly agree that the arrival of large foundation models marked a turning point, dramatically improving how well systems handle unfamiliar objects and situations.
A trade-off between generality and reliability
The biggest strength of VLA models, unlike traditional control programs built for a single task, is their generality — some ability to handle objects and environments they weren't explicitly trained on. But that comes with a persistent complaint from the floor: "sometimes it does something you can't predict."
The kind of "creativity" you get from a generative model can be flat-out frightening on a factory floor. — an automation consultant, speaking anonymously
As a result, most production lines don't execute a VLA model's output directly. Instead, a hybrid setup has become standard, with a rule-based safety layer sitting between the model and the machine.
Where the next race is headed
The next battleground is a "lightweighting race" — shrinking model size while preserving accuracy. Many real-world settings can't tolerate the latency of cloud-based inference, so compressing models down to a size that fits on the robot itself is becoming the key to real deployment.