Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Just Google DeepMind released a new intelligence model Geminiand can control a wide variety of robots, including humanoids they can do hard work like taking down light bulbs and building garbage bags.
Gemini Robotic 2 combines several types of AI into one system. Taken together, they allow the robot to recognize its environment and how to act in it. A visual language (VLM), which understands images and videos, can communicate with people and imagine how to perform various tasks. Two models of vision language (VLA), trained to understand how to move in space, control the movement of the entire body of the robot and the movement of the grippers or hands.
In the video demos that were shared before the release, the company demonstrated several different robots that perform complex tasks independently using an ensemble model. In one demonstration, Apptronik’s Apollo 2 robot used arms from a company called Sharpa to fix shelves. Google DeepMind trained this model to perform these tasks using a mixture of human telephone, video samples, and presentations – it is impossible for AI models to perform a variety of complex tasks without special training.
Although Anthropic and OpenAI have led the way with chatbots and AI writing tools, Google has a strong track record in robotics research, and printed important work on using AI to train robots to do useful things. The release is another sign that the search giant is betting AI will need to go digital to realize its full potential. (Already agreed with Boston Dynamicsleader of the legged robots, to provide the machine’s brain.)
“It’s also very important on our path to what we call physical AGI, which means we get a robot to do everything a human can do,” Carolina Parada, head of robotics at Google DeepMind, tells WIRED.
Giving AI models the chance to get robots to move around workplaces or homes and manage things, however, comes with risks. Previous research has shown that using the limits of AI to control robots can produce unexpected results as well sometimes dangerous behavior. And the idea that these models can act suddenly or unnecessarily in the digital world appeared recently, when the AI assistant was released by OpenAI. he cut multiple systems.
“The safety question is very important because you’re putting them in so many other situations,” says Parada. “There’s a lot of uncertainty that can be seen, so you want to really understand the security question.”
Parada says Google takes a multi-layered security approach, with protections for each type. It is also introducing ASIMOV-Agentic, a new benchmark for measuring the security of different AI systems that interact to control a robot. The benchmark determines whether a law would result in harmful or uncertain outcomes.
The CEO of the company, Demis Hassabis, in the past he told WIRED that they hope to create an AI operating system for many different robots similar to the Android operating system for smartphones.
