Dek: DeepMind's newest robot model handles a gift box and a pipette better than the last one. The more interesting release is the benchmark that tries to talk robots into doing something unsafe.
What Gemini Robotics 2.5 actually is
Google DeepMind published Gemini Robotics 2.5 on November 10 and called it its most capable vision-language-action model to date. A vision-language-action model takes camera frames and a written instruction and outputs motor commands, which is the difference between a chatbot that describes how to fold a towel and a machine that folds it.
The model isn't a robot, and it isn't an app either. It's the part that decides where the arm goes.
DeepMind measured it on more than 200 tasks: serving food with tongs, tying a bow on a gift box, pipetting liquid into wells. Those examples are chosen to be awkward. Tongs slip, ribbon is floppy, and a pipette punishes any tremor.
The numbers DeepMind published, and who ran them
According to DeepMind, the new model completes tasks with a success rate 30 to 100 percent higher than previous baselines, depending on task type, and finishes 20 to 30 percent faster. Those figures come from DeepMind's own evaluations, not from an outside lab, and nobody has replicated them yet.
Keep the range in mind. A 30 percent gain on one task family and a 100 percent gain on another is a wide spread, which usually means the model fixed something specific rather than everything at once.
The Mercedes-Benz kitting pilot in Hungary and Germany
The demo that will get quoted most is factory work. In a proof of concept with Apptronik and Mercedes-Benz, Apollo 2 humanoids performed automotive kitting on production lines in Kecskemét, Hungary and Sindelfingen, Germany, with a 96 percent success rate.
Kitting means picking the right parts and putting them in the right container for an assembly step. It's repetitive, it's easy to get wrong, and it's exactly the job factories have been trying to automate for forty years.
Not a fleet.
DeepMind calls the work a pilot, the runs happened on production lines rather than across them, and 96 percent on a defined task set doesn't tell you what happens when a part arrives out of order at 3 a.m.
Apollo 2 uses a five-finger SharpaWave hand with 22 degrees of freedom. The same model also runs on the Franka Duo platform and on two-fingered grippers, which matters for anyone who doesn't own a humanoid.
How Gemini Robotics 2.5 handles safety at inference time
The safety work is the part I'd read twice. DeepMind added inference-time reasoning guidance, a step where the model reasons about the scene before it moves instead of reacting to whatever the camera shows.
In DeepMind's evaluations, that guidance improved safe task completion by 33 percent around obstacles and by 43 percent around nearby people. Those are the two situations where a robot arm hurts someone: it hits a fixed object and something falls, or it swings into a person who walked in.
Read the fine print on both numbers. They measure whether the robot finished the task safely in a test setup, not whether it would refuse an instruction that a reasonable person would call dangerous.
ASIMOV-Jailbreak is a robot safety benchmark you can run
DeepMind also released ASIMOV-Jailbreak, an open-source benchmark for exactly that question. It holds hundreds of prompts that disguise unsafe requests across dozens of scenes, so a model gets asked to do something harmful in language that looks ordinary.
Every robot company will claim its model is safe. A benchmark with published prompts and scenes gives buyers something to compare, and it gives outside researchers something to attack.
That's the useful part. The goal isn't a certificate. It's a target that survives contact with people who want to break it.
Why the 99.5 percent figure belongs to Gemini 4 Argon
Here is where headlines go wrong. On ASIMOV-Jailbreak, Gemini 4 Argon reaches 99.5 percent jailbreak resistance, against 82 percent for Gemini Robotics ER 2. That number is about Gemini 4 Argon, a different model, and it isn't a score for Gemini Robotics 2.5.
The two are connected. DeepMind used Gemini 4 Argon as a simulation engine during development, generating practice runs, building digital twins of workcells, and analysing why tasks failed. The reasoning model helped train the robot model, and its safety result got attached to the same announcement.
So the honest reading is narrower than the press release suggests: one model was tested for refusal, another model does the lifting, and the second one hasn't published an equivalent number.
What Gemini Robotics 2.5 doesn't tell you
No price, no general availability date, and no download link. The model is rolling out to DeepMind's robotics partners, which means the people who can use it this month already have a contract and an arm.
There's also a gap the announcement doesn't close. Refusing a bad instruction is a language problem. Being physically safe next to a person is a control problem, and a model that scores well on prompt resistance can still move an arm through a space where a hand happens to be.
For anyone watching this from a factory floor in Europe or a warehouse in the Midlands, the practical question is boring and unanswered: which tasks will a vendor actually sell you in 2027, and what happens to the line when the model gets one wrong? The pilot says the technology is close. It doesn't say close to what.
