Gemini Robotics

New AI system from Google can control robots

(Bild: Google)

The US internet company Google has presented two new versions of its Gemini AI system, which are intended to lay the foundations for a whole new generation of helpful robots. With the help of AI, machines should be able to think spatially and interact flexibly with their environment.

The first AI model is called Gemini Robotics. It ensures that the system can not only produce texts, images or videos, but can also control the actions of a robot. The second model is called "ER" ("embodied reasoning"). It extends the spatial thinking of Google Gemini.

Skilled robot
According to a Google blog post, Gemini Robotics-ER can develop completely new skills in the blink of an eye. "If the AI model is shown a coffee cup, for example, it can intuitively choose a suitable two-finger grip to pick up the cup by the handle and calculate a sequence of movements to approach it safely."

The Internet company's AI specialists at Google DeepMind are working together with the Texan robotics start-up Apptronik to build AI-controlled humanoid robots that can operate efficiently and safely in dynamic environments. Apptronik has already developed several robot models, including the humanoid robot "Apollo", which is designed for tasks such as lifting, moving and stacking boxes in logistics and manufacturing.

Three elementary properties
Google's blog post states that in order to be useful and helpful to humans, AI models for robotics need three essential properties: firstly, they need to be general and able to adapt to different situations. Secondly, they need to be interactive in order to understand instructions or changes in their environment and be able to react quickly. And thirdly, they need to be dexterous and be able to do things that humans can normally do with their hands and fingers, such as carefully manipulating objects. "Gemini Robotics represents a significant performance boost in all three areas and brings us closer to true general-purpose robots."

In the blog entry, Google illustrates the performance of the new AI system in several videos. For example, Gemini Robotics can sort objects into containers on command as they are moved through the room. Another example shows the robot folding an origami.

Google will also make its robot AI available to testers. These included robotics specialists Agile Robots, Agility Robots, Boston Dynamics and Enchanted Tools.

Built-in safety
The research project also covers the topic of safety. The physical safety of robots and the people around them has long been a fundamental concern of robotics. The initial aim is to avoid collisions, for example, using conventional safety measures. Gemini Robotics-ER can be connected to these safety-critical control systems. The AI model can also be used to understand whether a potential measure is safe or not in a certain context and trigger appropriate reactions.