RT-2 (Robotic Transformer 2) is a vision-language-action model from Google DeepMind, released in 2023. It builds on large vision-language models and co-trains them on robot demonstration data, representing robot actions as text tokens so the same model that understands web images and text can also output motor commands. It is widely credited with establishing the VLA category.

Google DeepMind · 2023

RT-2: the model that started the VLA era

Q: What was novel about RT-2?

RT-2 showed that knowledge from internet-scale vision-language pretraining transfers into physical robot control. By encoding actions as tokens, it could be trained jointly on web data and robot data, and it displayed emergent semantic generalization — reasoning about objects and instructions it had never seen in robot training.

RT-2 (Robotic Transformer 2) is a vision-language-action model from Google DeepMind. Its breakthrough idea was simple and radical: represent robot actions as text tokens, so a model trained on the web can also drive a robot arm.

What it is

Actions as language

RT-2 takes a large vision-language model and co-trains it on robot demonstration data. Because actions are encoded as tokens — just another kind of "text" the model can emit — a single network learns from both web images-and-text and robot trajectories at once.

The consequence: the semantic knowledge a model absorbs from the internet flows into physical control. RT-2 could act on concepts it never saw in robot training — reasoning about which object is which, and why, from web knowledge alone.

Why it mattered

The proof of concept for the field

Before RT-2, robot learning and internet-scale AI were largely separate worlds. RT-2 demonstrated that they could be one: that the recipe behind large language models — scale, web pretraining, transfer — applies to physical action too. Nearly every VLA that followed, from π0 to GR00T N1, builds on that premise.

RT-2 is the action layer at its most influential — a VLA supplies the skill. What it does not supply is context about a specific home or business: that is the knowledge layer benned builds.

FAQ

Common questions

What is RT-2?

RT-2 (Robotic Transformer 2) is a vision-language-action model from Google DeepMind (2023). It co-trains a vision-language model on robot data, encoding actions as text tokens, and is credited with establishing the VLA category.

What was novel about RT-2?

It showed that internet-scale vision-language knowledge transfers into physical robot control, with emergent semantic generalization to objects and instructions never seen in robot training.

VLA Models π0 GR00T N1 Helix

Last updated: July 2026