← All models

Gemini Robotics On-Device 2

Published Publisher Google

Overview

Gemini Robotics On-Device 2 is an efficient vision-language-action (VLA) model designed for general-purpose robotic manipulation and local inference on robotic devices. Its architecture builds on Gemini Robotics 1.5 technology and on-device Gemma models. (deepmind.google)

Capabilities and uses

  • Accepts text instructions, images and numerical robot proprioception; it outputs robot actions as numerical values. (deepmind.google)
  • Runs locally to address network latency and connectivity constraints. It is intended as a component of robotic systems that understand language instructions and act in their environments, including bi-arm manipulation. (deepmind.google)
  • Is designed for multiple robot embodiments. Google reports adapting it to new bi-arm platforms with a few hours of adaptation and typically fewer than 200 examples; results are not a guarantee for every platform or task. (deepmind.google)
  • In Google's simulation and on-robot evaluations, it learned more efficiently and achieved higher final success rates than On-Device 1 on the evaluated new platforms. (deepmind.google)

Availability and limitations

The model is distributed to trusted testers; Google provides an early-access waitlist. Its ability to generalize to out-of-distribution tasks and control high-degree-of-freedom robots is limited. Evaluations focused primarily on standing bi-arm manipulation, so risks from mobile platforms and whole-body control were outside that evaluation scope. Google recommends layered safeguards, including low-level motion and force controls and hardware-specific safety mechanisms. Do not assume that capabilities demonstrated for the separate Gemini Robotics 2 or ER 2 models apply to this model. (deepmind.google)

Pricing

As of November 16, 2026, an official price for this model could not be verified. Currency, billing unit, input charges and output charges are unverified; the absence of a published price does not establish that access is free. The official model page and model card describe trusted-tester access but do not specify pricing. (deepmind.google)

Release and official sources

Announced July 30, 2026. (deepmind.google)