Google DeepMind has successfully integrated its advanced large language model, Gemini, into robots. The company claims that these machines can now perform physical tasks without prior human supervision or programming. By connecting to Gemini’s robotic models, developers can enhance their robots with improved spatial awareness and understanding of the physical world.
A team at Google DeepMind started by training a specialized version of Gemini 2.0 on patterns in large volumes of data. They then further trained it on thousands of hours of real robot demonstrations. This allowed the model to implement real actions, similar to how LLMs generate words in a sentence.
The team tested Gemini Robotics on humanoid robots and robotic arms, achieving consistent outperformance against state-of-the-art rivals. The new technology holds promise for creating intuitive machines that can tackle a range of physical tasks without relying on human supervision.
Source: https://www.nature.com/articles/d41586-025-00777-x