RT-2 (robotics model)

RT-2 (robotics model)
Robotic Transformer 2, a research model introduced by Google DeepMind in 2023 for mapping perception and language to robot actions.
Developer	Google DeepMind
Model type	Vision–language–action research model
Official resource	Research description

RT-2 (robotics model) robotic Transformer 2, a research model introduced by Google DeepMind in 2023 for mapping perception and language to robot actions.

Research context

Google DeepMind described RT-2 as a vision-language-action model trained using both web-scale and robot-related material. It illustrates an approach in which representations used for understanding text and images are also adapted to produce action tokens.

Action representation

In the published demonstration, action variables such as pose changes and gripper state are encoded into sequences that can be handled by a model originally designed to produce tokens. The model therefore needs an interface that translates outputs into physical commands.

Reported evaluation

The authors tested familiar and unfamiliar manipulation tasks and described improvements in generalization on their selected benchmarks. These are research findings under defined experimental conditions, not claims that RT-2 works reliably in every robotic environment.

Interpretation

RT-2 is a useful historical example of VLA research. Operational deployment must additionally address perception failure, safe motion, authentication, supervision and the limitations of training data.

See also

• Vision-Language-Action Models
• Robot Foundation Models
• Open X-Embodiment

Further reading

• RT-2: New model translates vision and language into action — Google DeepMind
• AI Risk Management Framework (AI RMF 1.0) — NIST