Google DeepMind has launched Gemini Robotics On-Device, a new language model that enables robots to perform tasks entirely offline, without relying on an internet connection. An evolution of the Gemini Robotics model released in March, the on-device version allows developers to fine-tune robotic behavior through natural language prompts.
In demonstrations, robots powered by the model performed physical tasks like unzipping bags and folding clothes. DeepMind says the model was initially trained for ALOHA robots but now runs successfully on the Franka FR3 and Apptronik’s Apollo humanoid. A new Gemini Robotics SDK accompanies the launch, letting developers train robots on new tasks with just 50–100 demonstrations using the MuJoCo physics simulator.
