Google DeepMind is launching a new autonomous version of artificial intelligence, Gemini Robotics, which can function without an Internet connection. This model provides robots with high-level motor skills and command understanding, similar to the flagship version introduced in March. About this informs The Verge.
The main advantage of the new model is its compactness and sufficient performance for direct use at work without the need for cloud computing. The flagship version allows robots to act flexibly, adapt to new conditions and make precise movements.
The head of robotics at Google DeepMind, Carolina Parada, explained that the previous model functioned in a mixed format — partly locally, partly through the cloud. The new version is completely offline-oriented and shows almost the same level of efficiency.
“The hybrid version of Gemini Robotics is still more powerful, but we are genuinely surprised at how strong this local model is. I would consider this a starter model or a model for applications with limited network access.” – said the official.
The stand-alone version of the model is capable of performing a variety of tasks without prior configuration and can adapt to new conditions after only 50-100 demonstrations. Although training was conducted exclusively on ALOHA, the model has been successfully reconfigured to work with other platforms, including Apptronik’s Apollo humanoid robot and the Franka FR3 bi-arm.
Google also presented the SDK for the standalone version of Gemini Robotics, which supports entering commands in natural language and testing in the MuJoCo physical simulator. This is the first time developers have had access to tools for testing and fine-tuning DeepMind’s VLA model. While internal security checks are ongoing, access to the model and SDK will be restricted to verified partners.
