Google DeepMind has taken a significant step toward making robots smarter and more independent with the release of Gemini Robotics On-Device, announced on June 24, 2025.
The new model represents a major advancement in robotics AI, allowing robots to operate without cloud connectivity while maintaining impressive capabilities. Unlike its predecessor launched in March, Gemini Robotics On-Device runs entirely on the robot's local hardware, eliminating the need for constant internet access while still delivering performance comparable to cloud-based alternatives.
"Since the model operates independent of a data network, it's helpful for latency sensitive applications, and ensures robustness in environments with intermittent or zero connectivity," Google DeepMind stated in their announcement.
The technology demonstrates remarkable dexterity and adaptability, enabling robots to complete complex tasks requiring fine motor skills. In demonstrations, robots powered by the model successfully performed challenging operations like folding clothes, unzipping bags, and executing precise industrial assembly tasks.
Alongside the model, Google is releasing a Gemini Robotics SDK that allows developers to evaluate and fine-tune the system for specific applications. The SDK enables training robots on new tasks with as few as 50 to 100 demonstrations using the MuJoCo physics simulator. While initially trained for ALOHA robots, Google has successfully adapted the model to work with other platforms including the bi-arm Franka FR3 robot and Apptronik's Apollo humanoid robot.
The development comes as competition in robotics AI intensifies, with companies like NVIDIA and Hugging Face also developing foundation models for robots. Google's on-device approach addresses critical challenges in robotics deployment, including privacy concerns and operation in environments with limited connectivity.
Developers interested in working with Gemini Robotics On-Device can apply to join Google's trusted tester program to gain access to both the model and SDK.