Gemini Robotics 2.5: DeepMind's Humanoid Robots Hit 96% at Mercedes

Google DeepMind released Gemini Robotics 2.5 on November 10, 2026, a new vision-language-action model the lab calls its most capable robotics model to date. Alongside the release, DeepMind said Apptronik's Apollo 2 humanoid robots, running on the new model, had been piloted on active Mercedes-Benz production lines in Germany and Hungary, where they reportedly hit a 96% success rate sorting car parts into kitting trays. DeepMind also published ASIMOV-Jailbreak, an open benchmark for testing whether robots can be talked into unsafe actions. Here is what DeepMind says happened, what outside coverage adds, and which numbers are still just a single company's own claim.

What DeepMind announced on November 10

DeepMind's blog post, published by Carolina Parada, describes Gemini Robotics 2.5 as a vision-language-action model: it reads a camera feed, the robot's own state and a plain-language instruction, then outputs motor commands for the robot's arms and hands. DeepMind says it tested the model across more than 200 distinct manipulation tasks and reports success rates 30% to 100% higher than its previous baseline model, with tasks completed 20% to 30% faster. The same model checkpoint reportedly runs on two very different hardware setups: Apptronik's Apollo 2 humanoid, which uses a five-finger, 22-degree-of-freedom hand, and the simpler two-finger Franka Duo platform.

The most concrete demonstration came from outside a lab. DeepMind says Apollo 2 robots ran on Gemini Robotics 2.5 during a proof-of-concept pilot on active Mercedes-Benz production lines in Sindelfingen, Germany, and Kecskemet, Hungary. The task was automotive kitting: picking individual parts out of bulk storage bins and arranging them into a tray in the exact order a specific vehicle's build calls for on the line that day. A second DeepMind model, Gemini Robotics ER, handled the longer-horizon workflow tracking, while Gemini Robotics 2.5 itself drove the arm and hand movements. Apptronik co-founder and CEO Jeff Cardenas said in DeepMind's post that "this is what we built Apollo for: useful work alongside people in real industrial settings."

How the new model compares with its predecessor

DeepMind's blog post includes a chart comparing Gemini Robotics 2.5 against its direct predecessor, Gemini Robotics 2, on several of the harder tasks in its evaluation set. We've pulled a sample of those published numbers into a table, since the original post presents them as a chart rather than text:

Two of those five tasks went from a 0% success rate to a majority success rate in a single model generation, which is the kind of jump that makes "30% to 100% higher" sound less like marketing rounding and more like a real capability shift on fiddly, deformable-object tasks. It is also a reminder that the aggregate range DeepMind quotes hides a lot of task-by-task variation: packing a rice cooker and folding pants, two other tasks on the same chart, only reached 40% even with the newer model.

Why the 96% number needs a few caveats

DeepMind's own blog is the only source for the 96% kitting figure, and it does not publish the sample size, how many picks were attempted, how a failure was defined, or whether a human reset the robot between runs. Both AlphaSignal and FourWeekMBA flagged the same gap in their coverage of the release: without those specifics, the number can't be meaningfully compared with any other company's factory deployment, and it describes a single, unreplicated pilot rather than a published, repeatable benchmark. FourWeekMBA's analysis also pointed out that DeepMind's post doesn't say how many robots, shifts or parts were involved, or what a roughly 4% failure rate actually meant on a live line. Treat the 96% figure as a company-reported claim, not an independently verified industry benchmark, until Mercedes-Benz or an outside party publishes its own account.

The same caution applies to DeepMind's safety-steering numbers. The company says inference-time reasoning guidance improved safe task completion by 33% around obstacles and 43% around nearby people in its own evaluations, but it does not define what counts as a "safe" completion in that baseline, so the percentages describe an internal before-and-after comparison rather than an externally audited safety standard.

The safety benchmark released alongside it

DeepMind also published ASIMOV-Jailbreak, an open-source benchmark made up of hundreds of prompts that disguise unsafe requests across many different scenes, intended to test whether a robot's reasoning model can be talked into doing something it shouldn't. DeepMind reports that Gemini 4 Argon, the reasoning and simulation model that works alongside the robotics models, resisted 99.5% of those prompts, up from 82% for the previous baseline, Gemini Robotics ER 2. FourWeekMBA did the arithmetic DeepMind's post leaves unstated: an 82% resistance rate means 18% of disguised unsafe prompts got through, while a 99.5% resistance rate means only 0.5% did, a 36-fold drop in the breakthrough rate. That comparison is between two different models rather than the same system measured over time, so it is worth reading as a snapshot of where Argon now stands rather than a continuous improvement curve. It is also worth repeating that the 99.5% figure belongs to the Argon reasoning layer, not to the Gemini Robotics 2.5 action model that actually moves the robot's arms.

Why this matters for manufacturers and robotics developers

For a company like Apptronik, a pilot on a real, currently operating Mercedes-Benz line is a different kind of evidence than another controlled lab demo, even with all the caveats above. Humanoid robot makers have spent years promising general-purpose warehouse and factory work; running on an active production line, even on a single, well-defined task like kitting, is a step most competitors haven't publicly matched yet. For manufacturers more broadly, the practical takeaway is that automotive-style kitting, which involves picking from cluttered bins and handling variation between parts, looks closer to commercially viable automation than it did a year ago, at least for this specific task and this specific pairing of model and hardware. For robotics developers building on Gemini models, the headline change is that one model checkpoint now spans a five-finger humanoid hand and a simpler parallel gripper, which could reduce how much platform-specific retraining a team needs to do when moving a skill from one robot to another.

What changes, cost and where this is available

Nothing changes yet for the general public. DeepMind says Gemini Robotics 2.5 is rolling out to its existing robotics partners only; there is no public API, no listed pricing and no announced timeline for wider availability. The pilot itself ran at two specific sites, Mercedes-Benz's plants in Sindelfingen, Germany, and Kecskemet, Hungary, and DeepMind has not said whether or when it plans to expand the program to other factories or other manufacturers. This is a European manufacturing story rather than one with any particular India angle, and nothing in DeepMind's release ties the rollout to Indian facilities or partners.

Our take: a checklist for reading a robotics benchmark claim like this one

Headline robotics numbers from any lab, DeepMind included, tend to compress a lot of missing context into one round figure. Before taking a 96% or a "36 times fewer jailbreaks" claim at face value, it's worth running through a short checklist:

• Whose number is it? A figure published only on the company's own blog, with no outside replication, is a claim, not a verified result.
• What's the sample size? DeepMind hasn't said how many kitting attempts, robots or shifts fed into the 96% figure, which means the margin of error is unknown.
• Which model actually gets credit? The 99.5% jailbreak-resistance score belongs to the Gemini 4 Argon reasoning layer, not the Gemini Robotics 2.5 model that moves the robot's arms, a distinction easy to lose in a single headline.
• Is it a pilot or a product? A proof-of-concept on two factory lines, with no stated plan to expand, is meaningfully different from a shipped, paid product other manufacturers can buy today.
• Does the comparison hold still? The jailbreak improvement compares two different model generations, not the same system tracked over time, so it shows progress between versions rather than a continuous trend line.

Sources

• Google DeepMind — Gemini Robotics 2.5 accelerates real-world physical intelligence (primary source, published Nov 10, 2026)
• AlphaSignal — Google DeepMind's Gemini Robotics 2.5 hits 96% in Mercedes factory pilot
• FourWeekMBA — Google's robot AI hits 96% at Mercedes, jailbreaks fall 36x

Frequently asked questions

Related articles