New Delhi: Google DeepMind has introduced Gemini Robotics on-device, a version of its robotics model that runs directly on robot hardware, removing the need for a cloud connection. By processing everything locally, robots can operate in environments with unreliable or no internet access. The Vision Language Action model is built on a variant of Gemini Robotics ER. Its architecture features a VLA backbone that acts as the brain, interpreting what the robot sees and determining the appropriate actions. In contrast, a local action decoder translates those decisions into real-world movements. The entire perception-to-action cycle takes just 250 milliseconds, fast enough for smooth, responsive controls.
We’re bringing powerful AI directly onto robots with Gemini Robotics On-Device. 🤖
It’s our first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments – without needing a constant internet connection. 🧵 pic.twitter.com/1Y21D3cF5t
— Google DeepMind (@GoogleDeepMind) June 24, 2025
In the test, Gemini Robotics’ on-device capabilities handled tasks such as unzipping bags, folding clothes, and pouring salad dressing, all without requiring a connection to external servers. Google says it outperformed other locally run systems on seven different manipulation tasks.
Running the model locally does require some trade-offs. For especially complex reasoning tasks, the cloud-based version achieves higher success rates. However, Google says the on-device model delivers strong enough performance for many practical scenarios.
Google DeepMind is providing a developer kit to make adaptation easier. Instead of using millions of training examples, the robot can learn new tasks from just 50 to 100 demonstrations. Developers can also run tests in a simulator without meeting physical hardware.
Although the base model was initially trained on ALOHA robots, it can be adapted to work with a wide range of systems. For example, a Franka industrial robot achieved a 63% success rate on familiar tasks. The model can also control humanoid robots, such as Apollo, which features a human-like body.
Multiple safety layers are built in. The system checks commands for potential hazards and works with hardware safeguards to prevent collisions. Even so, Google DeepMind recommends testing and deploying the system in real-world settings.
Access to Gemini Robotics On-Device is currently available through a closed testing program. Developers can apply for the Trusted Tester Program as Google DeepMind gathers feedback and gradually improves the system.
