RT-2

Brohan (DeepMind) 2023 · Paper

One-line summary — RT-2 represents robot actions as text tokens and co-fine-tunes a large vision-language model on both web-scale vision-language data and robot trajectories, transferring internet knowledge into robotic control and coining the term “vision-language-action” (VLA) model.

Problem

Traditional robot learning is data-hungry and brittle: the most capable language and vision-language models are trained on billions of tokens and images from the web — an amount unlikely to be matched with robot data in the near future. Prior attempts to bring LLMs/VLMs into robotics mostly used them as high-level planners parsing commands into primitives executed by separate low-level controllers, which never benefit from web knowledge. RT-2 asks: can large pretrained vision-language models be integrated directly into low-level robotic control to boost generalization and enable emergent semantic reasoning?

Method & architecture

“terminateΔposxΔposyΔposzΔrotxΔrotyΔrotzgripper-extension”\text{“terminate} \enspace \Delta\text{pos}_{x} \enspace \Delta\text{pos}_{y} \enspace \Delta\text{pos}_{z} \enspace \Delta\text{rot}_{x} \enspace \Delta\text{rot}_{y} \enspace \Delta\text{rot}_{z} \enspace \text{gripper-extension”}

e.g. "1 128 91 241 5 101 127". For PaLI-X, integers up to 1000 already have unique tokens; for PaLM-E, the 256 least-frequently-used tokens are overwritten as action tokens (a form of symbol tuning). Robot data is cast in VQA format: “Q: what action should the robot take to [instruction]? A:” followed by the action string.

Results

Evaluated over ~6,000 real-robot trials on a 7-DoF mobile manipulator:

Why it matters for SLAM

RT-2 established the actions-as-tokens VLA paradigm that OpenVLA, NaVILA, WorldVLA, and most subsequent robot foundation models adopted, and it demonstrated that web-scale pretraining genuinely transfers to embodied control. For SLAM researchers, it frames the key open question of the Spatial AI era: if web knowledge lets a policy generalize to novel objects, can SLAM-style spatial knowledge (metric maps, persistent geometry) be injected the same way to generalize across novel environments and long-horizon navigation?