Gemini Robotics On-Device

Google DeepMind's lightweight Gemini Robotics VLA that runs directly on the robot itself, with no internet connection required.

Gemini Robotics On-Device is a vision-language-action model (VLA, a model that takes in images and spoken instructions and outputs robot actions directly) released by Google DeepMind on June 24, 2025, a stripped-down version of Gemini Robotics specifically optimized to run on a robot's own onboard compute with no cloud connection needed — suited to poor-connectivity or latency-sensitive settings. It targets bimanual manipulation mainly, capable of dexterous tasks like unzipping a zipper or folding clothes, and supports fine-tuning to a new task with just 50 to 100 demonstrations. The model was trained on the ALOHA bimanual platform, and official demos showed it transferred to a dual-arm Franka FR3 setup and to Apptronik's Apollo humanoid. It was made available, together with the Gemini Robotics SDK, through a trusted-tester program. DeepMind later released Gemini Robotics On-Device 2, which it says can adapt to a new robot embodiment with fewer than 200 examples, also currently available through the trusted-tester program.

ExampleA developer collects 50–100 demonstrations for a new task, such as folding a new type of garment, and can fine-tune the On-Device model to that task; inference then runs entirely on the robot itself, so it keeps working even with no network connection.

Also called

Gemini Robotics On-Device 2

Related

Gemini Robotics · Gemini Robotics 2 · On-device Model · Gemini Robotics SDK · ALOHA · Apptronik Apollo

Sources

Gemini Robotics On-Device brings AI to local robotic devices (Google DeepMind 博客, 2025-06-24) (Chinese)
Gemini Robotics On-Device 2 模型页 (Chinese)
Gemini Robotics 模型总览 (Chinese)

As of

2026-09