"Hey Duo, kit all tools in the bin, close the kit and put the kit back into the bin." Duo answers back, "On it." No human typed that instruction. One robot said it to another, and a second robot understood and acted on it. That's the actual demo Google DeepMind used to show off Gemini Robotics 2, its newest robotics model family, and it's a stranger detail than the specs sheet lets on.
Three Things Gemini Robotics 2 Is Built For
Google DeepMind's team frames the release around a simple problem. Humans move through cluttered, unpredictable spaces without thinking about it. Robots have never been able to do that. In DeepMind's own demo material, Gemini Robotics 2 is described as the intelligence layer that controls a humanoid's whole body, its hands, and the grippers on other robot platforms, not just one arm reaching for one object.
Three areas got the focus. Whole-body control means coordinating movement across an entire humanoid rather than just the upper body, something DeepMind's own release notes confirm was a limitation of earlier versions. Dexterous manipulation covers tasks well past pick-and-place, like screwing in a lightbulb or tying off a trash bag, tasks the team reportedly doubted were solvable at all before they worked. And multi-robot collaboration lets more than one machine work the same job at once.
Why Each Robot Runs Its Own Copy
Here's the part that's easy to miss. When two robots work together in this system, it isn't one neural network steering both bodies like a puppeteer with two sets of strings. Each robot runs its own copy of the same model stack, doing its own individual reasoning, and the two coordinate by talking, not by sharing a single set of weights.
That's what the tool-kit demo actually shows. Apollo tells Duo to gather the tools, close the kit, and set it back in the bin. Duo confirms it's on the task. Apollo checks in mid-task: "Awesome precision, Duo. Keep going." It's a small moment, but it's the clearest illustration of the design choice: intelligence distributed across machines that reason independently and negotiate through language, not a single brain wired into two bodies.
Google DeepMind's own blog post on the release frames this differently, describing it as robots sharing a semantic understanding to hand off subtasks. Both descriptions point at the same underlying idea: no single model is puppeting every robot in the room. The system is built for robots that can improvise around each other.
The Hand Problem: 22 Joints, One Model
A human hand has close to 27 degrees of freedom, something we covered in our piece on the Wuji Hand and why most robot hands stop at 6. The team behind Gemini Robotics 2 puts the number at 22 separate joints for the hand it's targeting, and says nobody thinks about driving each one individually when they pick something up. Getting a model to make that many coordinated decisions, instead of a human's unconscious motor control, is the actual engineering problem behind tasks that sound trivial, like screwing in a lightbulb.
Worth checking separately: DeepMind's own published figures put dexterity success rates on individual multi-finger tasks anywhere from roughly 32% to 92% depending on the task, according to external reporting on the model card. That range wasn't part of the demo footage itself, so treat it as a data point to verify against DeepMind's model card directly if it matters for your use case.
- Gemini Robotics 2 controls whole-body humanoid movement, not just upper-body pick-and-place.
- Multiple robots coordinate by each running their own copy of the model and reasoning independently, not by sharing one network.
- The dexterity target is roughly 22 joints in a single hand, matched against tasks like tying a trash bag or screwing in a lightbulb.
- The team has said some of these tasks, notably the trash bag, were doubted internally before they worked.
My Take
The tool-kit exchange between Apollo and Duo is more telling than the marketing demo Google leads with. A robot walking over and picking up a watering can is impressive, but it's still one machine following one instruction. Two robots negotiating a shared job out loud is the part that actually points toward something new. I don't think that's an accident of the demo. It's the point.
FAQ
Is Gemini Robotics 2 one model or several?
It's a family of models: a vision-language-action model for physical control, an embodied reasoning model that plans and coordinates, and an on-device version built to run without a network connection.
Do the two demo robots share the same AI?
They run the same underlying model stack, but each robot holds its own copy and does its own reasoning. They aren't controlled by one shared network split across two bodies.
What robots was this demonstrated on?
The source footage shows the model controlling a humanoid (referred to as Apollo) and a second robot (referred to as Duo) with a gripper-style hand, working together on a shared task.
Can I use Gemini Robotics 2 today?
That depends on which model in the family and what access tier Google has assigned it, which is outside what this specific demo material confirms. Check Google DeepMind's own release page for current access details before assuming availability.
The specs and success-rate charts will get argued over for weeks. The detail worth remembering is smaller: one robot gave another robot an instruction in plain language, and it worked. For more on where robot hardware is still catching up to that kind of intelligence, our Robotics coverage has the rest of the picture.
0 Comments