Google DeepMind has launched a powerful on-device version of its Gemini Robotics AI model. The new system can control physical robots without relying on cloud connectivity.

It marks a major step in deploying fast, adaptive, and general-purpose robotics in real-world environments.

The model, known as ‘Gemini Robotics On-Device,’ brings Gemini 2.0’s multimodal reasoning into robots with no internet required. It’s designed for latency-sensitive use cases and environments with poor or no connectivity.

With strong task generalization, natural language understanding, and fine motor control, the model allows robots to complete complex tasks directly on the device.

Developers can now sign up for access through Google’s trusted tester program. Google is also releasing a full software development kit (SDK) to support experimentation and customization.

Optimized For Real-World Speed And Autonomy

Unlike its cloud-connected predecessor, Gemini Robotics On-Device runs entirely on the robot itself. That allows for faster reactions and better reliability, especially in offline or restricted settings. It is already capable of completing out-of-the-box tasks and can adapt to new ones with minimal data, with just 50 to 100 demonstrations.

“It’s small and efficient enough to run directly on a robot,” Carolina Parada, head of robotics at Google DeepMind, told The Verge. She added, “I would think about it as a starter model or as a model for applications that just have poor connectivity.”

Although the flagship hybrid model remains more powerful, the on-device variant holds its own. “We’re actually quite surprised at how strong this on-device model is,” Parada said.

Adapts Across Platforms And Tasks

The model was trained on Google’s ALOHA robot but has been adapted to others, including Apptronik’s Apollo humanoid and the bi-arm Franka FR3. It handles detailed actions like unzipping bags and folding clothes with smooth, low-latency inference.

This is also the first version of a DeepMind robotics model that developers can fine-tune.

Parada says, “When we play with the robots, we see that they’re surprisingly capable of understanding a new situation.”

Fine-tuning involves tele-operating the robot to complete a task a few times. This gives the model enough grounding to perform that task autonomously. Developers can test the model in Google’s MuJoCo simulator or in physical environments.

Designed For Privacy And Offline Performance

Gemini Robotics On-Device is ideal for security-sensitive or remote environments. It processes all data locally, helping protect user privacy in use cases like health care. Its offline capability ensures continuous operation even with unstable connectivity.

Parada noted that traditional reinforcement learning made training slow and brittle. Generative AI, by contrast, lets robots generalize from minimal input. “It’s drawing from Gemini’s multimodal world understanding in order to do a completely new task,” she explained.

Safety Systems Not Included

Unlike the hybrid model, the on-device version does not include built-in semantic safety reasoning. Google urges developers to replicate the safety stack used internally. This includes integrating the Gemini Live API and connecting the VLA to low-level safety-critical controllers.

“With the full Gemini Robotics, you are connecting to a model that is reasoning about what is safe to do, period,” said Parada.

Google is currently limiting access to selected developers to study real-world safety risks. The company hopes this release will expand the robotics community’s ability to develop adaptive, secure, and capable robots across diverse environments.