2QR / JOURNAL✳

From language to physical action

On November 10, Google DeepMind announced Gemini Robotics 2.5, a vision-language-action model intended to help robots interpret their surroundings and carry out physical tasks. The company says it evaluated the system across more than 200 tasks, including using tools and handling flexible or delicate objects.

DeepMind reports a proof of concept on Mercedes-Benz production lines in which an Apptronik Apollo 2 humanoid picked and packed automotive parts with a 96% success rate. That figure is specific to the company’s described test and should not be read as a general reliability rate for every factory task.

Why ordinary objects are difficult

A robot manipulating a garment, a container, or a cable has to deal with variable shapes, imperfect positioning, friction, and unexpected contact. A useful system must also know when it has failed, avoid damaging objects, and operate safely around people.

DeepMind says the model can work with different robot types, including humanoid hands and parallel grippers. It is rolling out to robotics partners, so public experience outside the company’s selected evaluations will be important.

What to watch

The next step is less about a single polished demonstration and more about repeatable performance over long shifts, new environments, maintenance demands, and clear safety boundaries. Independent evaluations will help determine where the system can be useful in practice.

Source

Google DeepMind’s robotics announcement

Cover image: AI-generated conceptual illustration.

✳
← Back to all stories