Google DeepMind released Gemini Robotics 2 on Wednesday, a three-model suite that extends the company’s foundation model into full-body humanoid control and multi-robot coordination. The launch pushes Alphabet’s robotics stack from the tabletop into the world of legs, balance, and shared workspaces, and it arrives with an unusually candid concession from its authors: the hardest parts remain unsolved.

The suite splits along familiar lines. Gemini Robotics 2 is the vision-language-action model that runs the body. Gemini Robotics ER 2 is the embodied-reasoning layer, capable of supervising more than one machine at once and now available to developers through the Gemini API and Google AI Studio. Gemini Robotics On-Device 2 runs locally and, according to DeepMind, can adapt to unfamiliar hardware with fewer than 200 training examples in some cases. Alongside the models, the company released ASIMOV-Agentic, a safety benchmark designed to test whether the reasoning layer refuses unsafe motor-control instructions.

In a demonstration recorded for reporters on July 28, the model directed Apollo, the humanoid built by Apptronik, to walk across a room, pick up a watering can, navigate obstacles, and place the can on a lower shelf. It’s the sort of scripted errand that reads as mundane until you remember that whole-body coordination on legs has been the field’s persistent bottleneck.

The numbers around the demo are where the story gets interesting. Bloomberg reported a 92 percent success rate on unscrewing a light bulb. Figures relayed via The Chosun Daily and summarized by The Next Web put tying a trash bag at 44 percent and sealing a ziplock bag at 40 percent. Kanishka Rao, director of robotics at DeepMind, called true dexterity “a distant goal” and noted that robots still learn far less efficiently than humans.

Carolina Parada, DeepMind’s VP of robotics, framed the stakes in different terms. “The safety question is even more pressing because you’re putting them in a lot of other situations. There’s a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply.”

There’s a structural detail worth naming. Google is shipping software; it isn’t shipping bodies. Most of the humanoid hardware this stack will animate is manufactured in China, and Washington has already moved to restrict that supply chain on security grounds, echoing the 2022 CHIPS Act logic now being extended into embodied systems. The interesting question isn’t whether Gemini Robotics 2 can unscrew a light bulb. It’s whose chassis it’ll be legally allowed to run on.

Sources