Gemini Robotics On-Device 2

Gemini Robotics On-Device 2 is Google DeepMind's most efficient vision-language-action (VLA) model, introduced on July 30, 2026, optimized to run locally on robotic devices without network latency or internet connectivity. It is the on-device variant of the Gemini Robotics 2 generation, alongside the flagship Gemini Robotics 2 (VLA) and Gemini Robotics ER 2 (embodied reasoning). Fast adaptation: the model is natively multi-embodiment and inherits the motion transfer techniques of Gemini Robotics 1.5. It can adapt to completely new bi-arm robot embodiments in just a few hours of adaptation time, typically with fewer than 200 examples — even for embodiments with drastically different shapes, sensors, and degrees of freedom, as demonstrated on the Dexmate, SO101, and Trossen platforms. Availability: available to early-access partners via Google DeepMind's application process. It serves use cases that require on-device operation: no network dependency, low latency, and local data processing.