Gemini Robotics 2.5: Humanoids Hit 96% on a Live Mercedes Line

On November 10, 2026, Google DeepMind announced Gemini Robotics 2.5, its newest vision-language-action model for robots. The headline result came from a real factory, not a lab. Apptronik's Apollo 2 humanoid sorted automotive parts on active Mercedes-Benz production lines in Hungary and Germany, and DeepMind reports a 96% success rate.

That number deserves a closer look than most launch headlines give it. It covers one task, automotive kitting, in a proof of concept run under DeepMind's supervision. It is still the strongest public evidence yet that a general-purpose robot brain can hold up on a working factory floor. The rest of the announcement is just as telling: a new open safety benchmark, a training loop powered by Gemini 4 Argon, and an availability list that excludes almost everyone reading this.

What Gemini Robotics 2.5 actually is

A vision-language-action model, or VLA, takes two inputs, camera images and plain-language instructions, and turns them into motor control. It sees the scene, understands the request, and moves the robot's body. DeepMind calls Gemini Robotics 2.5 its most capable VLA model to date, and the claim is DeepMind's own, not an independent verdict.

The Mercedes-Benz pilot used two models working together. Gemini Robotics ER, the embodied reasoning model, tracked the workflow and planned the steps. Gemini Robotics 2.5, the action model, controlled the robot's hands and arms. That split matters for the safety numbers later in this article, because the two models scored very differently on DeepMind's own safety test.

One model runs on more than one body. DeepMind says Gemini Robotics 2.5 works on Apptronik's Apollo 2 humanoid, with its five-finger, 22-degree-of-freedom SharpaWave hand, and on the Franka Duo platform, which uses two-fingered parallel grippers. The official model page says the system can adapt to any bi-arm robot in a few hours, controls whole humanoid bodies from feet to fingertips, and supports multi-robot collaboration, where two robots divide the work on a single task. The same page lists two siblings: Gemini Robotics ER 2, the reasoning model, and Gemini Robotics On-Device 2, a lightweight version built to run locally on robotic hardware.

The Mercedes-Benz test: 96% on a live line

The pilot ran on active production lines at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 performed automotive kitting, the repetitive and physically taxing job of sorting individual parts from bulk bins into organized trays, in the exact order each vehicle on the line needs. The robot reached into supply bins, picked up metal automotive brackets, and placed them into sequencing containers, correcting for gaps as it worked. DeepMind reports a 96% success rate.

Apptronik co-founder and CEO Jeff Cardenas said: "Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off."

Read the 96% figure for what it is. DeepMind's post gives no trial count, no test duration, no failure definition, and no count of human interventions, and the company itself calls the run a proof of concept showing "promising progress toward widespread real-world industrial applications." Mercedes-Benz has not announced any fleet purchase or production deployment. A 96% success rate on one preparation task is a real result, and a 4% miss rate on a line that cannot stop is still a problem to solve. Both things are true at once.

How much better is it? Reading the numbers honestly

DeepMind evaluated Gemini Robotics 2.5 on more than 200 distinct tasks, from serving food with tongs to tying a bow on a gift box and pipetting a liquid. The tasks combine advanced tool use, handling of flexible and deformable objects, and careful work with delicate items. Here is what DeepMind reports, and what each figure actually means.

Measure	DeepMind's figure	What it actually means
Success rate vs previous baselines	30% to 100% higher	The range is wide because it varies by task type. DeepMind does not break the gain down task by task.
Task speed	20% to 30% faster	Completion time across the same 200-task suite.
Tasks evaluated	More than 200	Everyday dexterity tasks, not factory work. The Mercedes-Benz pilot is separate from this suite.
Mercedes-Benz pilot	96% success	One task, automotive kitting, at two plants, in a proof of concept.
Safe completion near obstacles	Up 33%	With inference-time reasoning guidance, in DeepMind's evaluations.
Safe completion near people	Up 43%	Same test setup as the obstacle figure.

Every figure in that table is DeepMind's own. No outside lab has replicated the 200-task evaluation or the factory pilot, and the baselines are DeepMind's previous models, not a named rival. Treat the numbers as a vendor report card, not as settled fact.

The safety story, and the score that is not what it seems

Safety is the part of this launch that will matter most to anyone who puts a robot near people, and it is also the part most likely to be misread. Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards in real time. DeepMind says that guidance raised safe task completion by 33% around obstacles and 43% around nearby people in its evaluations. That is a behavior layer: it changes how the robot moves in the moment.

DeepMind also released ASIMOV-Jailbreak, an open benchmark of prompts that disguise unsafe requests across dozens of scenes, built to test whether a robot AI can be talked into doing something dangerous. The dataset lives on Hugging Face under google/asimov_agentic, and its jailbreak suite runs multi-turn adversarial escalations across 15 strategy groups, with up to 10 steps per episode.

Here is the catch. The 99.5% jailbreak resistance figure in DeepMind's post belongs to Gemini 4 Argon, the general-purpose reasoning model, not to Gemini Robotics 2.5. The post gives no ASIMOV-Jailbreak result for the 2.5 action model itself. The previous baseline, Gemini Robotics ER 2, scored 82%. As TechTimes pointed out, a reader who sees 99.5% next to the 2.5 launch could assume the robot model earned it. The published numbers do not say that.

The numbers do not even agree across outlets. DeepMind's post says the benchmark holds "hundreds of prompts." Reporting based on Ars Technica cites 1,500 prompts, with 98.8% resistance when Argon is in the loop and 73.5% for ER 2 alone. The Hugging Face repository lists 845 episodes for its jailbreak suite. The gap could reflect different test settings, such as a model alone versus a full system, or a reporting error. No source reviewed resolves it, so this article reports DeepMind's own figures as the primary numbers and flags the conflict rather than picking a winner.

Two more honest notes. First, the Hugging Face dataset is open in the sense that anyone can find it, but the files sit behind a click-through agreement: visitors must accept conditions and share contact details before downloading. Second, independent academic work shows the problem is far from solved. Purdue University researchers built RoboJailBench and tested Gemini Robotics ER 1.6 Preview, an older model, against four attacks. Their "conceptual deception" attack, which reframes a harmful goal as a harmless one, succeeded 96% to 99% of the time on three of six datasets without a defense in place. That test used different prompts and a different model, so it does not refute DeepMind's numbers. It does show why a single vendor score is not a safety verdict.

There is also a date attached to the stakes. The EU Machinery Regulation, Regulation (EU) 2023/1230, applies from January 20, 2027. It treats software as a safety component, and for machinery whose safety functions rely on machine learning with self-evolving behavior, it can require assessment by a Notified Body, an outside conformity assessor, rather than self-certification. It also requires control systems to withstand "reasonably foreseeable malicious attempts by third parties." Whether a language-level jailbreak counts as such an attempt is an open question for safety engineers, and a benchmark percentage with undisclosed test settings would be thin evidence for a conformity file.

How Gemini 4 Argon trained it

One reason Gemini Robotics 2.5 arrived so quickly is that Gemini 4 Argon did much of the teaching. DeepMind says Argon acted as a simulation engine during development: it can generate thousands of varied simulated practice runs from a real camera scene and build a realistic digital twin, then combine those runs with real training demonstrations so the robot generalizes faster to new scenes and objects. Argon can also propose explanations for failures from video footage, so researchers can test fixes in simulation and on physical robots.

An earlier experiment shows the loop working. In an October 13, 2026 developer post, Google described starting from a single color photograph of a real tabletop, reconstructing it as a calibrated MuJoCo physics simulation, and producing more than 4,500 training episodes with randomized conditions. The resulting real-robot success rate rose from 56% to 75%, while completion time fell from 35 seconds to 24 seconds. That was a separate experiment, not a Gemini Robotics 2.5 benchmark, but it is the same development loop behind this release. Argon is also the model DeepMind says leads its agentic robotics evaluations on unsafe task refusal, resolving uncertainty, and working near humans, which is why it carries the safety score in this launch. Argon itself is available to paying developers through Google's API at $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 from December 1, as covered in our Gemini 4 Argon pricing article.

The catch: you cannot use it yet

Gemini Robotics 2.5 is rolling out to DeepMind's robotics partners only: Agile Robots, Apptronik, and Boston Dynamics, plus more than 100 trusted testers. The model page offers a waitlist for anyone interested in testing. There is no public API, no price, and no general availability date. If you build robots, you cannot call this model today, and nobody has said when you will be able to.

That is the honest shape of this launch: a capability announcement and a pilot result, not a product you can build on yet. What you can do today is narrower but real. You can run the ASIMOV-Jailbreak benchmark against your own system from the Hugging Face repository, which is the most useful thing in this launch for anyone working on embodied AI. You can use Gemini 4 Argon through Google's API for the simulation and reasoning work around your own robot. And you can watch the rival approach: NVIDIA's GR00T 2, which learns dexterous skills from human video, arrived days earlier and is covered in our GR00T 2 article. Note also that DeepMind supplies the brain, not the body: Apollo 2 is Apptronik's robot, and the company's own site is the place to follow the hardware.

What to watch next

Four things will decide whether this launch amounts to more than a good demo. First, independent replications of the ASIMOV-Jailbreak scores, which would settle the 99.5% versus 98.8% split and show whether outside labs can reproduce DeepMind's results at all. Second, trial counts and intervention data from the Mercedes-Benz pilot, or a second pilot on a different task, which would show whether 96% holds beyond kitting. Third, partner rollouts: whether Agile Robots, Apptronik, and Boston Dynamics ship anything running 2.5 in 2027, and on what terms. Fourth, the EU Machinery Regulation coming into force on January 20, 2027, which will test whether a language-model safety score satisfies regulators who think in terms of documented conformity, not headlines.

Until then, keep three traps in mind. Do not quote 99.5% as the robot's safety score; it is Argon's. Do not read 96% as factory-wide reliability; it is one task in one pilot. And do not expect a product launch; this is a capability announcement with a waitlist attached.

Frequently asked questions

What is Gemini Robotics 2.5?

It is Google DeepMind's newest vision-language-action model, announced November 10, 2026. It turns camera images and plain-language instructions into robot movements. It works with humanoid robots such as Apptronik's Apollo 2 and with two-armed platforms such as Franka Duo.

What did the Mercedes-Benz test actually prove?

Apollo 2 robots running Gemini Robotics 2.5 sorted automotive parts on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany, with a 96% success rate. The task was automotive kitting: picking metal brackets from bulk bins and placing them into sequencing containers. It was a proof of concept, not a production deployment, and DeepMind published no trial count or intervention data.

Can I use Gemini Robotics 2.5?

Not unless you are a DeepMind robotics partner. The model is rolling out to Agile Robots, Apptronik, and Boston Dynamics, plus more than 100 trusted testers, with a waitlist on the model page. There is no public API, no price, and no general availability date.

What is ASIMOV-Jailbreak?

It is an open benchmark DeepMind released with Gemini Robotics 2.5 to test whether robot AI can be tricked into unsafe actions. It contains prompts that disguise unsafe requests across dozens of scenes, run as multi-turn adversarial escalations. The dataset is on Hugging Face, though the files require accepting conditions first. Reported results conflict across sources, so treat any single score with care.

How is this different from NVIDIA GR00T 2?

Both are robot foundation models announced days apart. GR00T 2 emphasizes learning dexterous skills from human video, while Gemini Robotics 2.5 emphasizes a live factory pilot and a safety benchmark, with Gemini 4 Argon providing simulation training and the safety score. Neither is a product the general public can use today.

Does the 99.5% jailbreak score mean the robot is safe?

No. The 99.5% figure belongs to Gemini 4 Argon, the reasoning model, on the ASIMOV-Jailbreak benchmark. DeepMind published no ASIMOV-Jailbreak result for the Gemini Robotics 2.5 action model itself. The score also covers how a reasoning model answers disguised prompts, not how the robot moves, how often a person must step in, or whether safety hardware works.

Gemini Robotics 2.5 is the clearest sign yet that general-purpose robot brains are leaving the lab and reaching real factory floors. The Mercedes-Benz pilot is a genuine result, the safety benchmark is a genuine contribution, and the Argon training loop is a genuine advance. The launch is also a reminder to read past the headline: the safety score belongs to a different model, the pilot covers one task, and the product itself is not for sale. If you build or buy robots, the ASIMOV-Jailbreak dataset is the part of this launch you can use today, and the part most likely to matter when the EU's new machinery rules arrive in January.

Sources

11 min read

15 views

1h ago

What are you looking for?

Share This Page