Google DeepMind launched a new language model called Gemini Robotics On-Device that runs locally on robots without needing an internet connection.
The update builds on the original Gemini Robotics model released in March. This new version controls robot movements and can be fine-tuned by developers using natural language prompts.
Google claims Gemini Robotics On-Device performs close to the cloud-based model in benchmarks and beats other unnamed on-device models. 
A demo showed robots using the local model to unzip bags and fold clothes. Although trained for ALOHA robots, the model also works on a bi-arm Franka FR3 robot and the Apollo humanoid robot by Apptronik.
The Franka FR3 successfully handled new tasks like assembly on an industrial belt, objects it hadn’t encountered before.
Google is releasing a Gemini Robotics SDK where developers can train robots with 50 to 100 task demonstrations using the MuJoCo physics simulator.
Other companies are also moving into robotics models: Nvidia is building a humanoid foundation model platform, Hugging Face is working on open robotics models and actual robots, and Korean startup RLWRLD is creating foundational robotics models.
Image Credits: Google
