A Vision-Language-Action (VLA) model is a single neural network that takes a camera image of a scene and a natural-language instruction as input, and outputs a sequence of low-level robot actions directly — no hand-written rules or explicit intermediate program. It fuses three real components end to end:
- Vision encoder — turns the camera image into features that locate objects and their positions in the scene.
- Language encoder — turns the text command into features that identify which object is the target and what action to perform on it.
- Action decoder — combines both feature streams and generates the trajectory (grasp point, motion path, release point) the arm executes.
The glowing ring above an object is a simplified stand-in for attention — the weight the model places on that region of the scene once it has matched the command's target phrase to something it can see.