In the field of artificial intelligence, Google DeepMind has today launched an exciting new product — Gemini Robotics On-Device. This is a new member of the Gemini family, specifically designed for robots, capable of running locally on the device without relying on continuous internet connectivity. This means that robots can not only adapt more quickly to new tasks and environments but also maintain stable performance without a network connection.
Gemini Robotics On-Device is based on the multimodal reasoning capabilities of the Gemini 2.0 model, showcasing strong flexibility and task generalization ability. It has been carefully optimized for various intelligent operations, such as folding clothes and opening bag zippers, all of which can be performed directly on the robot itself.
Notably, Gemini Robotics On-Device is especially suitable for applications sensitive to latency, ensuring normal operation even in environments with poor network connectivity. To help developers make the most of this new technology, Google will also release the Gemini Robotics SDK, making it easier for developers to evaluate the model's performance on specific tasks. With this SDK, developers can test the model in DeepMind's MuJoCo physics simulator and quickly adapt it to new domains, requiring only 50 to 100 demonstrations.
In terms of performance, Gemini Robotics On-Device demonstrates astonishing adaptability across multiple tasks. The model performs exceptionally well in seven different difficulty levels of dexterous manipulation tasks, capable of handling objects and scenes it has never encountered before. This not only demonstrates its adaptability across different robots but also proves its versatility.
DeepMind's breakthrough marks a new advancement in building powerful robot models, taking an important step toward the era of true embodied intelligence.