Gemini Robotics 2.5: what DeepMind's new robot model can do and what is unproven

Gemini Robotics 2.5 is Google DeepMind's newest robot brain, and the company says it is the most capable vision-language-action model it has built. DeepMind published the announcement on November 10, 2026, and trade press covered it the next day. The headline numbers are real but narrow: a 96% success rate on one automotive packing job, and big jumps on lab tasks the previous model could not do at all. What the evidence does not yet show is robots doing this work unsupervised at scale.

What DeepMind actually released

A vision-language-action model, or VLA, takes camera images and a written instruction and outputs motor commands. Gemini Robotics 2.5 is the latest in a line that DeepMind began in March 2025, when its Gemini Robotics team described a VLA built on Gemini 2.0 that could be fine-tuned to learn new short tasks from as few as 100 demonstrations. This version was evaluated across more than 200 distinct tasks, from serving food with tongs to tying a bow on a gift box and pipetting liquid.

DeepMind says success rates are 30% to 100% higher than earlier baselines depending on the task type, and that tasks finish 20% to 30% faster. SiliconANGLE's write-up rephrased the first figure as a "30% to 100% task success rate", which is a different claim. The DeepMind post is the primary source, and we use its wording: the improvement is relative to previous baselines, not an absolute success range.

The model works across robot bodies. DeepMind names Apptronik's Apollo 2 humanoid, which has a five-finger, 22-degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-finger parallel grippers. It is rolling out to DeepMind's robotics partners, not to the public. DeepMind's post names Apptronik, Mercedes-Benz and Agile Robots; SiliconANGLE also lists Boston Dynamics among the partners, which the DeepMind page we read does not.

The numbers that matter

The task-level table in DeepMind's post is the most honest part of the announcement, because it shows the starting point. Packing pastries with tongs went from 5% to 100%. Closing a black toolbox lid went from 0% to 93%. Hanging a cardigan went from 0% to 62%, sweeping trash into a dustpan from 0% to 52%, and unhanging a mug onto a table from 29% to 75%. Folding pants reached 40% from zero, and so did packing a rice cooker.

Read those as a map of where the frontier sits. Several tasks went from impossible to a coin flip or worse. A 40% success rate on folding pants is a research result, not a product. Even the 62% on the cardigan means roughly four failures in ten attempts.

The industrial figure is different in kind. In a proof of concept on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 sorted and packed automotive parts with a 96% success rate. That is the number a factory manager would care about, and it is also the number with the most caveats. It comes from a proof of concept, it was run by the model's developer and its partners, and DeepMind did not publish the sample size, the cycle time against a human worker, or what happened in the other 4%.

The safety claims

DeepMind reports two safety results. Inference-time reasoning guidance raised safe task completion by 33% around obstacles and 43% around nearby people. And Google is releasing ASIMOV-Jailbreak, an open-source benchmark with hundreds of prompts that disguise unsafe requests across dozens of scenes.

Be careful with the jailbreak figure. DeepMind says Gemini 4 Argon achieved 99.5% jailbreak resistance on it, against 82% for the previous baseline, Gemini Robotics ER 2. That is the resistance of the reasoning model that can sit above the robot, not a score for Gemini Robotics 2.5 itself. It is still useful, because an open benchmark lets outsiders test the claim, which is rarer in robotics than in chatbots. A robot that follows a disguised instruction can break something or hurt someone, so the cost of a failed refusal is physical.

The "safe completion" percentages have the same limit: they are DeepMind's own measurements on its own scenarios, and the post does not say how many trials sit behind each one.

What the partners say

The one on-record quote from a partner comes from Apptronik chief executive Jeff Cardenas, who said: "This is what we built Apollo for: useful work alongside people in real industrial settings". That is a statement of intent from a company that sells the humanoid, and it should be weighed that way. Mercedes-Benz is named as the site of the proof of concept, but the DeepMind post we read carries no Mercedes quote and no commitment to deploy.

How this fits the race

Frontier labs now treat embodied AI as part of the same contest as chat and coding models. DeepMind credits its flagship Gemini 4 Argon with helping speed the robotics work, which is the pattern to watch: general reasoning gains flowing into physical control. We covered how Argon's spatial reasoning scores were reported in our earlier look at Gemini 4 Argon's robotics claims. NVIDIA announced its own open robot model, GR00T 2, a day earlier, so two of the biggest players in the field shipped within 24 hours of each other.

None of this changes the central constraint. A language model that is wrong produces a bad paragraph. A robot that is wrong drops a part, jams a line or swings an arm near a person. That is why the gap between a 96% proof of concept and a dependable factory cell is where the money and the risk both sit.

What we would watch

First, an independent replication. Nothing in the announcement lets an outside lab run Gemini Robotics 2.5, so every performance figure is the vendor's. Second, the failure data: how many trials, how many interventions, how many near misses. Third, the commercial terms with Mercedes-Benz and the other partners, which would show whether anyone is paying for this yet. Fourth, whether ASIMOV-Jailbreak gets adopted by other robotics labs, since a shared test is how a field starts holding itself to account.

Our read: this is a genuine step in capability, with clear improvement on hard dexterous tasks, and a good example of a lab publishing a task table that shows its failures. It is not evidence that humanoid workers are ready for general factory use. Treat the 96% as a promising pilot result and the rest as lab progress.