Google DeepMind Debuts On-Device Gemini Robotics Model for Offline Robot Control
Google DeepMind has unveiled a new version of its language-based artificial intelligence model, titled Gemini Robotics On-Device, designed to enable robots to function locally without the need for an internet connection. Based on the earlier Gemini Robotics model released in March, this upgraded version allows robots to perform physical tasks and control movement autonomously. The news was first reported by TechCrunch.
According to Google, the on-device model delivers performance nearly equivalent to its cloud-based counterpart and outperforms other local models. Demonstrations showcased the model successfully completing tasks such as unzipping a bag, folding clothes, and assembling parts in an industrial setting.
Initially designed for the ALOHA robot, the new model has since been successfully deployed on other platforms, including the FRANKA FR3 and Apptronik’s humanoid robot Apollo. In addition to the model itself, Google has introduced the Gemini Robotics Development Kit, which allows developers to train robots in the Mujoco simulator with just 50 to 100 demonstrations per task.
Meanwhile, other industry players such as NVIDIA, Hugging Face, and South Korea's RLWRLD are also working to develop their own foundational models in the robotics domain.
