Google DeepMind just shipped Gemini Robotics 2. It’s the closest thing yet to one AI brain that runs any robot. The new model family controls humanoids from feet to fingertips and handles fiddly multi-finger tasks. It can adapt to an entirely new robot body in a few hours.
That’s the part that matters. Most robots today are pre-programmed for narrow, repetitive jobs. Gemini Robotics 2 is built around the opposite idea. It’s a general intelligence layer that any robot can drop in and use.
Here’s what I found in the DeepMind announcement and what it actually means.
What Gemini Robotics 2 actually does
The release is three models, not one. A vision-language-action model turns what a robot sees into movement. Gemini Robotics ER 2 handles the reasoning: planning and orchestrating tasks. Gemini Robotics On-Device 2 runs locally for robots that can’t count on a cloud connection.
The headline feature is whole-body control. Instead of the old arm-focused approach, the model coordinates the entire humanoid. That lets it bend, crouch, stretch, reach, and balance through full-range movement. In the demo, a robot walks into a cluttered room, cleans it up, and teams with another robot to finish faster.
The same checkpoint runs on different robot bodies. DeepMind tested Gemini Robotics 2 on the Apptronik Apollo 2 with two different hand setups. It also ran on the Franka Duo with a standard gripper.
All from one model. That cross-hardware trick is the hardest part of robotics. It’s also the reason Google calls this an intelligence layer, not another robot demo.
The numbers tell a mixed story
DeepMind published success rates across task categories, and they’re worth reading carefully.
Whole-body manipulation lands in the 45 to 76 percent range depending on the task. Picking up from a table hit 68.4 percent, from the floor 45.7 percent, from a shelf 76.3 percent. Gripper work is stronger, with precise insertion at 89.6 percent and general pick and place at 74.2 percent.
Multi-finger dexterity is the honest weak spot. Unscrewing a bulb hit 92 percent, but screwing one in dropped to 36 percent. Tying a trash bag was 44 percent, dustpan 32 percent, ziplock 40 percent. DeepMind says multi-finger manipulation remains challenging, and the numbers back that up.
These are lab benchmarks, not field reliability. I read them as capability signals, not product promises. And the pattern matters more than the exact numbers. Whole-body and gripper work is getting genuinely good. Fine finger control is still the bottleneck.
Where you can actually get it
Gemini Robotics ER 2, the reasoning model, is live in Google AI Studio right now and in private preview on the Gemini Enterprise Agent Platform. VLA and on-device models are locked to early-access partners for now.
The partner list matters more than the availability. Apptronik, Boston Dynamics, and Agile Robots are all in. The model is already being pointed at some of the most advanced humanoids in the world.
For robot developers, the pitch is simple. You bring the hardware, Google brings the brain. The same Gemini Robotics 2 checkpoint that runs on an Apollo 2 today is supposed to run on your custom rig tomorrow. Adaptation takes a few hours. That’s a very different procurement model from buying a robot with a locked-in operating system.
Why this is a bigger deal than another robot demo
I’ve watched the robotics space for years, and the pattern here looks familiar. First came the LLM moment, where one model replaced a thousand narrow NLP tools. This is that same play for physical AI. DeepMind isn’t selling robots, it’s selling the intelligence layer, and it wants that layer in every robot body on the market.
That’s a smart position. If one model genuinely controls most robot hardware, Google gets its Android moment. The same checkpoint powers everything from warehouse arms to home humanoids. Robot makers get a better brain without building their own, and Google gets the platform economics.
It also connects to the broader Gemini push. Google has been moving fast on the model side all month, shipping Gemini 3.6 Flash with lower prices. Now the same smarts extend into the physical world.
The honest caveats
The benchmark numbers are lab results, not real-world reliability. A 68 percent pick-from-table rate is promising, but it’s not ready to run your warehouse. Multi-finger work is clearly years from human-level.
And there’s no consumer timeline here. The on-device models are going to early partners, not to the public. If you’re waiting for a Gemini-powered home robot, this announcement doesn’t change your calendar.
Still, the direction is unmistakable. One AI model that adapts to any robot body, in hours, is the foundation the whole industry has been waiting for. The robot soccer league I covered earlier is starting to look like an early warning, not a joke.
What’s next: Gemini Robotics 2 is the intelligence layer. The real test comes when partner robots start shipping with it. I’ll be watching whether the lab numbers survive contact with real floors.
