A vision-language-action model starts from a vision-language model that already carries web-scale knowledge about objects, scenes, and instructions, and teaches it to act: RT-2 coined the term by emitting discretized robot actions as text tokens, and later systems attach a small generative action expert that produces continuous action chunks at control rates.
The appeal is transfer: semantics learned from the internet (what a mug is, what “the extinct animal” refers to) show up in the robot’s behavior without robot data teaching them. Modern VLAs typically pair a slow, semantic backbone with a fast motor head (the System 2 / System 1 split) and are trained with the LLM-style ladder of pretraining, supervised fine-tuning on demonstrations, and increasingly reinforcement learning post-training.
