Google DeepMind has launched a new version of its artificial intelligence (AI) system called ‘Gemini Robotics On-Device’, built to run directly on robots without needing any internet connection. This on-device model allows robots to process vision, language, and actions using only their built-in hardware. This makes them faster, more reliable, and able to operate in places where internet connections are weak or not allowed.
The on-device model is a smaller, more efficient version of the Gemini Robotics system that was first introduced in March 2025. While the original model relied on cloud computing, this new version works entirely on the robot’s hardware, which reduces response times and improves privacy and security. Despite its smaller size, the model still performs most of the same tasks as the cloud version.
The company claims that robots using Gemini Robotics On-Device were able to carry out a wide range of tasks, like folding clothes, unzipping bags, assembling parts, and picking up objects. According to the tech giant, these tests were conducted on various robot platforms, including the ALOHA robot, the Franka FR3 dual-arm system, and the Apptronik Apollo humanoid. Even on robots the model had not been trained on before, it was reportedly able to adapt and perform tasks after seeing only a few examples.
Meanwhile, to make the system more flexible, the company is providing developers with tools like the Gemini Robotics SDK and the MuJoCo simulator. These tools allow engineers to train and fine-tune the model for new tasks by showing it just 50 to 100 demonstrations. This helps reduce development time and lets the same model be used across many different types of robots and environments. According to Google, ‘Gemini Robotics On-Device’ outperformed other on-device AI models when tested on tasks that were complex, multi-step, or outside the model’s original training.
In terms of safety and responsible use, the company stated that the model is designed to work alongside safety-critical control systems that manage the robot’s movements. The company also claimed it is limiting access to the model for now, allowing only trusted partners to use it while additional testing and safety evaluations are carried out.
This comes at a time when the robotics industry is increasingly moving toward on-device AI. Many other big tech companies, like Nvidia, are also focusing on bringing advanced AI capabilities directly to robot hardware. For example, Nvidia’s Jetson platform is designed to support powerful, low-latency AI performance without relying on the cloud.