Google DeepMind Launches Offline AI Model for Robots, Enhancing Dexterity and Task Adaptation

Google DeepMind has introduced a new on-device version of its Gemini Robotics AI model, enabling robots to operate without an internet connection. This latest iteration, called Gemini Robotics On-Device, builds upon the original Gemini VLA model that Google launched in March. Designed to help robots understand and perform complex tasks more effectively, the updated model is optimized to run locally on robotic devices, enhancing its capabilities in general-purpose dexterity and rapid task adaptation. By operating offline, the new model broadens the potential use cases for robots in environments with limited or no connectivity. According to Google, Gemini Robotics On-Device can follow natural language instructions to complete a variety of intricate tasks, such as unzipping bags or folding clothes, all while working directly on the device itself. Google tested the model on seven dexterous manipulation tasks, ranging from zipping a lunch box to drawing a card and pouring salad dressing. The results showed that the model could quickly adapt to new tasks, requiring as few as 50 to 100 demonstrations to learn and execute them effectively. Initially trained on the ALOHA robot, the model was further refined for compatibility with the bi-arm Franka FR3 robot and Apptronik’s Apollo humanoid robot. To complement the launch, Google is also releasing a software development kit (SDK) that developers can use to evaluate and fine-tune the new model. This SDK is available through Google’s “trusted tester” program, making it the first vision-language-action (VLA) model Google has made available for customization. In a blog post, Google emphasized that this release represents a significant milestone in its mission to make robotics AI more accessible and to bridge the gap between digital AI and the physical world. “Our on-device solution will help the robotics community address key challenges related to latency and connectivity,” the company stated. “We are eager to see what innovative applications the robotics community will develop using these new tools.” This development underscores Google’s commitment to advancing AI technology and providing robust solutions that can operate in diverse and challenging environments.