VLA model

A VLA model, short for vision-language-action model, is a neural network that maps camera images and a natural language instruction to robot actions. It pairs a vision-language backbone, often adapted from a model trained on image and text pairs, with an action head. That head converts the backbone's output into motor commands, joint positions, or gripper poses instead of text.

Google's RT-2 and Physical Intelligence's pi-0 are both built this way. Published RT-2 numbers put its largest model at 1 to 3 Hz on robot hardware, since the vision-language backbone runs before every action step. That backbone is typically a multi-billion parameter model, so inference needs a dedicated GPU either onboard the robot or reachable over a low latency link, not a CPU.

Specing hardware for a VLA workload means budgeting for real-time inference latency on the vision-language backbone, not just training throughput. A robot that needs 10 Hz control cannot wait 300 ms for a forward pass.

Sources

Last verified August 29, 2026.