Google DeepMind has taken a major step toward bringing advanced AI capabilities to physical robots with the release of Gemini Robotics On-Device, a model designed to run entirely on local robotic hardware.
The new system, announced in late June 2025, builds upon the Gemini Robotics platform introduced in March that first brought Gemini 2.0's multimodal reasoning into the physical world. What makes this latest release groundbreaking is its ability to operate independently of cloud connectivity while maintaining impressive performance levels.
"Gemini Robotics On-Device shows strong general-purpose dexterity and task generalization, and it's optimized to run efficiently on the robot itself," according to Google DeepMind's official announcement. This independence from network connectivity makes it particularly valuable for latency-sensitive applications and environments with intermittent or zero connectivity.
In benchmark testing, Google claims the on-device model performs at a level close to its cloud-based counterpart while outperforming other on-device alternatives, especially on challenging out-of-distribution tasks and complex multi-step instructions.
The model demonstrates remarkable adaptability, requiring only 50-100 demonstrations to learn new tasks. While initially trained for ALOHA robots, Google has successfully adapted it to work with bi-arm Franka FR3 robots and Apptronik's Apollo humanoid robot, showcasing its versatility across different robotic platforms.
Alongside the model, Google is releasing a Gemini Robotics SDK to help developers evaluate and customize the technology for their specific applications. The SDK allows testing in Google's MuJoCo physics simulator and provides tools for rapid adaptation to new domains.
This development represents a significant advancement in practical robotics by bringing sophisticated AI directly to robotic devices. While consumer applications may still be years away, Carolina Parada, head of robotics at Google DeepMind, sees broad potential: "They could be more useful in industries where setups are complex, precision is important and the spaces aren't human-friendly. And they could be helpful in human-centric spaces, like the home."