Google spent the last two years showing you robot arms that could fold laundry on a table, and the whole time the robot was basically a torso bolted to a bench. Today that changed. DeepMind announced Gemini Robotics 2, and the headline is not smarter chat. It is feet. The model now drives an entire humanoid body, walking included, and it can hand work off to a second robot standing next to it.

I want to be careful here, because robot demos are the most edited video on the internet. So let me separate what Google published from what Google implied.

What Google actually shipped

Gemini Robotics 2 is three models, not one. DeepMind's own post, published July 30, 2026 and credited to Carolina Parada, lays them out plainly. Gemini Robotics 2 is the vision-language-action model, the thing that turns "put the watering can in the green bin on the bottom shelf" into motor commands, and it can now control a full humanoid from feet to fingertips instead of just an upper body. Gemini Robotics ER 2 is the embodied reasoning model, a vision-language model that plans multi-step jobs lasting several minutes and talks to humans while it does it. Gemini Robotics On-Device 2 is the small one that runs locally on the robot and, per Google, adapts to a completely new robot body with a few hours of data.

Only one of those is something you can touch today. Ars Technica's Ryan Whitwam reported the same day that ER 2 is available now on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, while the action model and the on-device model went to early-access partners only. Three models, one door open.

The verdict on availability: this is a research release with a developer sample attached, not a product launch.

The numbers Google chose to publish

This is where it gets interesting, and where most coverage will skip.

Ars reported two accuracy figures from Google's own materials. ER 2 classifies video frame completeness with almost 60 percent accuracy, and it identifies key moments in a video feed with almost 90 percent accuracy. That second number matters more than it sounds. Knowing when to stop pouring the coffee is the difference between a robot that helps and a robot that floods your counter.

But look at the first one again. Sixty percent on whether a frame shows a finished task. Google says that beats the 1.6 release and beats competing models on the same test. Fine. It is still a coin flip with a lean. Ars put it in a caption without flinching: better than other models, "but not by much."

Google's own chart is honest in a way I appreciate. It shows Gemini Robotics 2 running three different robot bodies off one model checkpoint — Apptronik's Apollo 2 with SharpaWave hands, the same Apollo 2 with Inspire hands, and a Franka Duo with a Robotiq gripper — and it reports medium to high success on whole-body and gripper tasks while stating outright that multi-finger dexterous manipulation "remains challenging." That is DeepMind telling you the fingers are still the hard part. Believe that sentence more than the video.

Why the multi-robot thing is the real story

The demos show Apollo 2 and a Franka F3 Duo working the same task without colliding. Two robots, one plan, coordinated by the reasoning model.

Everybody is going to read that as a productivity story. I read it as an access story. The moment one model orchestrates several bodies, you have a coordinator with authority over multiple physical machines, and every question I keep asking about software agents gets a sharper edge. Who can issue that instruction. What is the scope of what it may touch. How do you revoke it in the middle of a task. If you have followed my writing on auditing what your AI agents are allowed to do, you already know the software version of this problem is unsolved for most teams. Now it has arms.

There is also a failure-recovery change worth naming. Ars describes ER 2 catching a failed step in real time and retrying just that step instead of restarting the whole task — readjusting its grip when a ball rolls away. That is the boring engineering win that actually decides whether a robot is useful in a warehouse, and it does more for reliability than any leap in raw model quality.

What is unproven

Speed. Google's own post says the robots "have more to advance in movement speed." Ars agrees the test robots still cannot match human speed or grace, though it credits the demos for including real-time footage rather than the speed-ramped clips this industry loves. Real-time footage is a small honesty signal and I will take it.

Generalization. One checkpoint across three embodiments is genuinely impressive. Three embodiments is also three, all of them lab hardware from partners. Google says it is also testing on Boston Dynamics hardware. Testing is not shipping.

Safety. The word appears in every headline about this release, including Ars's. What I can confirm from the published material is that DeepMind is framing safety around the reasoning layer understanding its environment and its failures. What I cannot confirm is any independent evaluation. There is no third-party audit in this announcement. Treat the safety claim as the vendor's claim, because that is exactly what it is.

What this means if you are not building robots

Most of you are not buying a humanoid this year. Here is the part that touches you anyway.

The same company that wants a model driving a physical body is the company deciding how much compute to build and where. The AI buildout has a real-world footprint, and I wrote about the regulatory side of that buildout earlier this week. Physical AI raises the compute bill, not lowers it.

And the trust question follows the hardware. When I covered the US restriction on connected robots, the whole fight was about who controls a networked machine sitting inside your building. Gemini Robotics 2 is a cloud-adjacent brain for exactly that class of machine, with an on-device variant as the hedge. That hedge exists for a reason, and the reason is that nobody wants their warehouse to stop when the link drops.

My take

This is the most credible whole-body robotics release I have seen from a general model vendor, and it is still a preview. The honest read is: reasoning got noticeably better, hands did not get solved, and one company now has a plausible path from "chat model" to "thing that walks across a room and does a chore."

If you build with this, start with ER 2 in AI Studio, because that is the only piece you can actually get. If you are just watching, watch the multi-finger numbers in the next release. That single line on Google's chart is the whole game.

Sources: Google DeepMind's announcement, Ars Technica's report on the three models and availability, DeepMind's ER 2 technical post, Engadget's coverage of the demos.