Google’s AI research division, DeepMind, has introduced a new set of artificial intelligence models, dubbed Gemini Robotics, designed to help robots interact with objects, navigate environments, and carry out practical tasks.

In a series of demonstration videos released on Wednesday, robots powered by Gemini Robotics were shown performing actions like folding paper and placing glasses into a case — all in response to voice commands.

DeepMind says the models are built to generalize behaviors across different types of robotic hardware. This allows the AI to link what the robot “sees” with appropriate actions, enhancing flexibility and adaptability.

The lab claims that Gemini Robotics performed effectively in unfamiliar environments not included in its training data.

Additionally, DeepMind has released a streamlined version called Gemini Robotics-ER for researchers to develop their own robotic control systems. A new benchmarking tool, Asimov, was also introduced to help assess potential risks associated with AI-driven robots.