Google DeepMind has officially unveiled an updated version of the Gemini Robotics 2 model, designed for integrated control of the entire robot body. This VLA (Vision-Language-Action) model converts text commands and visual data into motor actions, supporting robots of various form factors, from desktop manipulators to full-size humanoids. Unlike the previous version, the new system controls not only the upper part of the body, but also provides the ability to move, squat, bend and work in a confined space. During the demonstration tests, the Apptronik Apollo 2 humanoid, controlled by this model, successfully performed the tasks of searching for objects, removing objects from shelves and manipulating various objects.

The key improvement was improved motor skills, which allows you to work with complex brushes with twenty-two degrees of freedom. The model is capable of performing delicate operations such as tying knots, closing packages, or unscrewing light bulbs. At the same time, the company introduced an update to the ER 2 system, which functions as a high-level task scheduler. This solution coordinates multi-stage processes, monitors the execution of actions and distributes work between several robots. In the demo scenario, Apollo 2 supervised Google's two-armed robot while cleaning the garage, instructing it to put tools in a container.

Additionally, a local Gemini Robotics On-Device 2 model was presented, which works without an Internet connection. The system adapts to new types of robots in less than a few hours of training based on two hundred examples. Currently, Gemini Robotics 2 is available in private viewing mode, ER 2 is provided through Gemini API and Google AI Studio, while On-Device 2 is being tested by trusted partners of the company. This development marks significant progress in the field of robotics, bringing closer the creation of universal autonomous systems capable of performing complex household and industrial tasks without constant human intervention.