Google Introduces Gemini Robotics On-Device Model for Local Robot Control
Google DeepMind has unveiled a new language model called Gemini Robotics On-Device, designed to enable robots to perform tasks locally without relying on an internet connection. Building upon their earlier Gemini Robotics model released in March, this new version can control robot movements and be fine-tuned through natural language prompts. In benchmark tests, Google reports the model achieves performance levels close to the cloud-based Gemini model and surpasses existing on-device models in general benchmarks.
During demonstrations, robots equipped with this local model successfully performed activities such as unzipping bags and folding clothes. Although initially trained on ALOHA robots, the model was adapted to work with a bi-arm Franka FR3 robot and the Apollo humanoid robot by Apptronik.
Google claims the Franka FR3 was able to handle unforeseen scenarios like industrial belt assembly and objects it had not previously encountered. Additionally, Google is releasing a Gemini Robotics SDK, allowing developers to train robots on new tasks using demonstrations in the MuJoCo physics simulation environment.
The move indicates a push toward more autonomous, offline-capable robots. Other industry players, including Nvidia, Hugging Face, and RLWRLD, are also investing in foundation models for robotics, signaling a growing interest in edge AI for automation and robotics applications.
Carolina Parada
