Gemini Robotics 2.5: what it can do and how to get access
2026-11-11 · 4 min read
Gemini Robotics 2.5 is Google DeepMind's newest model for controlling robots, announced November 10, 2026. It takes camera input and an instruction and turns them into arm and hand movements. You can't download it or call it from an API: it is rolling out to DeepMind's robotics partners, and everyone else joins a waitlist.
What Google DeepMind posted
@GoogleDeepMind opened its thread with the claim that the model "is able to learn hundreds of everyday skills," and listed folding laundry, serving pastries and sorting car parts on a factory floor.
The announcement post calls it the lab's most capable vision-language-action model so far. DeepMind says it tested more than 200 distinct tasks. Against earlier baselines, it reports a success rate 30% to 100% higher depending on the task, and completion 20% to 30% faster.
We covered the previous version in August, in our Gemini Robotics 2 post. The story then was a robot that could tie a knot and failed a dustpan chore two times out of three. So the first thing we checked this time was the per-task numbers.
What it can do
The figures below come from the charts in DeepMind's post, comparing Gemini Robotics 2 with 2.5 on the same tasks. They are DeepMind's own measurements.
| Task | Gemini Robotics 2 | Gemini Robotics 2.5 |
|---|---|---|
| Pack pastries with tongs | 5% | 100% |
| Close the lid of a toolbox | 0% | 93% |
| Move a hanging mug to the table | 29% | 75% |
| Untie a ribbon bow | 0% | 70% |
| Hang a cardigan | 0% | 62% |
| Sweep trash into a dustpan | 0% | 52% |
| Fold pants | 0% | 40% |
| Pack a rice cooker | 0% | 40% |
The jump is large. Most of these tasks went from never working to working at least some of the time, and for half the table "some of the time" is still as far as it goes. Folding pants fails six times in ten.
A note from us on the baseline: the July launch listed a dustpan task at 32%, and this chart starts the sweeping task at 0%. Those are probably different setups, so we would not compare numbers across the two launches.
The Mercedes-Benz pilot
The headline number is 96%. In pilots at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 humanoid sorted parts from bulk bins into trays in the order each vehicle needed. Gemini Robotics 2.5 moved the hands and arms. A second model, Gemini Robotics ER, kept track of the workflow.
DeepMind's post does not say how many trials sit behind the 96%, or how a success was counted. Heise reports that DeepMind robotics director Kanishka Rao called the path from 96% to near-human reliability a "steep road," and that the robot is still slower than a person. No outside group has measured any of this yet.
Safety numbers
DeepMind says the model steers around hazards as it moves. In its own tests, safe task completion improved 33% around obstacles and 43% around nearby people.
It also published a test set called ASIMOV-Jailbreak: prompts that dress up an unsafe request to see if the robot's planner goes along with it. DeepMind reports 99.5% resistance when Gemini 4 Argon does the reasoning, against 82% for Gemini Robotics ER 2.
How to get access
DeepMind's pages describe three situations today:
- Gemini Robotics 2.5 itself is rolling out to robotics partners. The post names Agile, Apptronik and Mercedes-Benz. For everyone else, the model page has a "Join waitlist for early access" button that opens a Google form.
- Gemini Robotics ER 2, the planning model, has a "Try" button on DeepMind's models page that opens it in Google AI Studio as a preview. It is the only piece a developer can use right now.
- DeepMind has not published pricing for 2.5.
DeepMind describes its testers as a group of 100-plus, so expect a wait on the form.
Our read
We think the gains are real, and they came fast. Gemini Robotics 2 launched on July 30. A little over three months later, a pastry task that worked 5% of the time works every time in DeepMind's tests. Reliability like that matters because cost is what keeps robots out of most workplaces. An Anthropic study we covered found robots can already do most physical tasks but beat a person on cost for only 0.3% of them.
The caution is the same one we gave in August. Every number here was measured by the company selling the result, in a setup it chose. A factory pilot at 96% means one part in 25 goes wrong, and somebody on the line has to catch it.
What this means for your business
We don't expect many Fox Valley shops to price out a humanoid this quarter. What carries over is the habit DeepMind's own charts teach: ask for the per-task success rate, and ask who measured it. That applies to a phone agent or a quoting tool as much as to a robot.
If you want help working out which parts of your week software can already handle reliably, New Face Design offers a free process audit.