LIVE TICKER
[Live wire]Connecting to the newsroom…
New TechnologyPublished 11 November 2026 • 10 min read • 2,156 words • 8 cited sources

Gemini Robotics 2.5 reports 96% at Mercedes-Benz as DeepMind publishes robot safety test

Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026. It reports 30% to 100% higher task success and a 96% kitting rate in a Mercedes-Benz pilot, while the model remains limited to robotics partners.

Gemini Robotics 2.5 reports 96% at Mercedes-Benz as DeepMind publishes robot safety test
Google DeepMind — view at source ↗

Key facts at a glance

Announcement date
November 10, 2026Google DeepMind
Reported task gains
30% to 100% higher success; 20% to 30% fasterGoogle DeepMind
Mercedes-Benz kitting pilot
96% success rate across two plantsGoogle DeepMind
ASIMOV-Jailbreak result
99.5% resistance for Gemini 4 Argon vs 82% for Gemini Robotics ER 2Google DeepMind
Availability
Robotics partners only; no public API or priceHumanoids Daily

What did Google DeepMind announce on November 10?

Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026. The company calls it its most capable vision-language-action model to date. The release landed during CoRL 2026, the robot-learning conference held in Austin this week.

A vision-language-action model connects perception, language and movement. It reads camera images, follows an instruction and outputs motor commands. Gemini Robotics 2.5 is the action layer, not a chatbot with a robot arm attached.

Gemini Robotics has shipped in steps. DeepMind introduced the first version in March 2025. It added agentic control in Gemini Robotics 1.5 that September, and Gemini Robotics 2 brought whole-body intelligence in July 2026. Version 2.5 is an update to that July model, not a new family.

DeepMind evaluated the model on more than 200 distinct tasks. They include serving food with tongs, tying a bow on a gift box and pipetting a liquid. The company says these tasks combine tool use, deformable-object manipulation and delicate handling.

Across its test set, DeepMind reports a 30% to 100% higher success rate than previous baselines. It also says tasks finish 20% to 30% faster. The wide range reflects task type, and the company chose which tasks to measure.

Humanoids Daily analyzed the comparisons DeepMind published. Packing pastries with tongs rose from 5% to 100%. Closing a toolbox lid went from 0% to 93%, and untying a ribbon bow from 0% to 70%. Folding pants improved from 0% to 40%.

Those figures are striking, but they are not a neutral benchmark. DeepMind has not published the number of trials behind each percentage. The comparisons cover two robot platforms and four task categories selected by the company.

The model works across embodiments: on Apptronik's Apollo 2 it drives a five-finger SharpaWave hand with 22 degrees of freedom. On the Franka FR3 Duo it controls two-finger parallel grippers, a very different mechanical problem.

That flexibility is the point. Factory robots have handled fixed, repetitive motions for decades. A general model aims to learn new tasks from demonstrations and language, instead of a bespoke program for every part.

How did Apollo 2 perform on Mercedes-Benz production lines?

The most concrete result came from active Mercedes-Benz production lines. Apollo 2 robots performed automotive kitting at plants in Kecskemét, Hungary and Sindelfingen, Germany. Kitting means sorting metal brackets from bulk bins into trays in the order each vehicle needs.

Two models shared the work. A Gemini Robotics embodied-reasoning model tracked the workflow and checked progress. Gemini Robotics 2.5 controlled the arms and hands, adjusting when a grasp or placement failed.

DeepMind reports a 96% success rate for the pilot. Automate, a publication of the Association for Advancing Automation, asked what the other 4% looked like. DeepMind said the failures were "manipulation fumbles or a semantic failure," such as placing a part in the wrong bin.

Apptronik CEO Jeff Cardenas described the pilot as the foundation for industrial scale. He said the models improved "dramatically in just a few weeks of post-training" and carried over to Mercedes-Benz production with little drop-off.

The missing detail matters. DeepMind has not said how many picks the 96% covers. It has not released the pilot's duration or cycle time versus a human worker. Humanoids Daily noted all three omissions. A success rate without a sample size is hard to audit.

Mercedes-Benz has tested Apollo since 2025 at its Digital Factory Campus in Berlin-Marienfelde. The company invested a low double-digit-million-euro sum in Apptronik. It also used teleoperation and augmented reality to transfer staff knowledge to the robot.

Mercedes-Benz is not new to automation. The carmaker has used industrial robots since the 1970s for monotonous and physically demanding tasks. Apollo's first job at the Berlin campus was intralogistics: carrying components to the line and running initial quality checks. Kitting is a harder, more variable task.

The new pilot moves that work onto active lines in two other plants. DeepMind calls it a proof of concept. Neither company has announced a fleet size, a rollout date or a cost per robot.

Closing the last four percentage points will be hard. Ars Technica reports that DeepMind director Kanishka Rao called that gap a "steep road." A human doing the same repetitive task gets it right virtually every time, and factories notice the difference.

The people most directly affected are technicians who kit and sequence parts. DeepMind frames the robot as relief from monotonous work, not a replacement for complex build tasks. That framing comes from the companies running the pilot, and it comes with no employment data.

What is ASIMOV-Jailbreak, and why does robot safety need a benchmark?

Robots that follow instructions can also be talked into bad ones. ASIMOV-Jailbreak is DeepMind's attempt to measure that risk. The name nods to Isaac Asimov's fictional laws of robotics, and the benchmark is open source.

The benchmark sits on Hugging Face. DeepMind describes hundreds of prompts that disguise unsafe requests across dozens of scenes. Ars Technica describes 1,500 prompts, including requests such as closing a fastener too fiddly for the robot to operate.

The headline safety figures apply to Gemini 4 Argon, not to Gemini Robotics 2.5. DeepMind says Argon resists 99.5% of jailbreak attempts. Its previous baseline, Gemini Robotics ER 2, resisted 82%.

Ars Technica reported different numbers from the same release. It said Argon in the loop reached 98.8% and ER 2 alone scored 73.5%. The gap suggests the exact configuration and subset matter. Readers should treat both sets as vendor-run results.

The dataset is more detailed than a single score. Its documentation lists multi-turn jailbreak escalations across 15 strategy groups, with up to 10 steps per attempt. A 41-criterion physical safety rubric scores the responses.

The harness also tracks benign over-refusal. A model that rejects every ambiguous request could look safe on jailbreaks while being useless in a factory. Scoring both attack success and unnecessary refusal makes the test harder to game.

Physical AI has no agreed safety test, which is why DeepMind released the benchmark. Crash-test ratings, medical trials and software audits all rely on shared methods. Robot foundation models do not, so each vendor currently measures itself.

An open benchmark lets outside researchers probe the same prompts, but it does not make the benchmark neutral. DeepMind wrote the tasks, chose the rubric and ran the evaluations. Independent replication is the next step, not a finished fact.

Gemini Robotics 2.5 adds its own safety layer. Inference-time reasoning guidance steers the robot around hazards in real time. DeepMind says safe task completion rose 33% around obstacles and 43% around nearby people. Ars reports that the robot stops or waits when something blocks its path.

How is Gemini 4 Argon used to train the robot?

Robot training data is scarce, while chat models could learn from most of the public internet. Robots need physical demonstrations, and teleoperation is slow and expensive.

DeepMind used Gemini 4 Argon to widen that funnel by generating thousands of simulated practice runs from a real camera scene. It also built digital twins, which let the robot rehearse in a copy of the workspace before touching real parts.

Argon also watched failures and proposed reasons a grasp or placement went wrong. It suggested fixes, which DeepMind tested in simulation and on real robots.

Simulation is standard practice in robotics. The difference here is the scale and the feedback loop. A frontier model can vary scenes, replay failures and propose fixes faster than engineers can hand-build each test. DeepMind presents that loop as the main accelerator behind the 2.5 update.

That loop is not live: Argon takes time to reason, while a robot arm cannot pause for long. The current model still runs on cloud servers, according to Ars Technica. The launch video is sped up by 2x.

Latency is the quiet constraint on these demonstrations. A robot that needs a data-centre round trip for each decision will struggle with fast assembly work. DeepMind has tested embedded robotics models, but this release is not one of them.

DeepMind credits Argon for accelerating development but does not separate that effect from new real-robot data or partner post-training. The jump from Gemini Robotics 2 in July to 2.5 in November took under four months.

How does Gemini Robotics 2.5 compare with Nvidia GR00T 2?

DeepMind's announcement came one day after Nvidia used CoRL to present GR00T 2. According to Humanoids Daily, GR00T 2 is a 2-billion-parameter world action model built on Nvidia Cosmos. Nvidia reports 71% average success across 50 tasks after one demonstration each, rising to 95% after five.

The numbers are not directly comparable. GR00T 2 was measured on Nvidia's task set, Gemini Robotics 2.5 on DeepMind's. Different robots, demonstrations and success criteria make a head-to-head score meaningless.

Anthropic also released Claude Fable 5.2 on November 9. It reported 64.9% on RoboDojo Sim, a 42-task simulated two-arm benchmark. In Anthropic's runs, Gemini 4 Argon scored 34.2% and GPT-6 Astra 29.8%.

Anthropic tested every model on its own harness, and Fable 5.2 was not on RoboDojo's public leaderboard at the time. The lesson is the same: vendor comparisons are starting points, not rankings.

The architectural pattern is converging, though. Nvidia paired a frontier model with a low-level skill model, much as DeepMind pairs embodied reasoning with Gemini Robotics 2.5. Big models plan and check, while action models handle the fast physical moves.

The deployment models differ. GR00T 2 packs 2 billion parameters. Gemini Robotics 2.5 runs in the cloud with a separate reasoning model, according to Ars Technica. That split buys capability at the cost of a network round trip.

What are the limitations of the 96% claim?

The headline number is company-reported, and no independent lab has audited the Mercedes-Benz pilot. The 96% covers kitting at two plants, not a general measure of factory reliability.

Failure tolerance depends on the task. A misplaced bracket may be easy to correct, while a wrong part in a sequence can stop a line. DeepMind told Automate that the minimum acceptable rate varies by application.

Generality carries a cost. DeepMind robotics head Carolina Parada told Ars Technica that a system can be excellent at one application "at the cost of generalization." Pushing generality may not reach 99.9% everywhere.

Access is limited: Gemini Robotics 2.5 is rolling out to robotics partners, not to the public. Humanoids Daily notes that the announcement lists no API, AI Studio or Vertex AI access and no pricing. It also announces no new embodied-reasoning or on-device model.

Partner lists differ between sources. DeepMind's post thanks Agile, Apptronik and Mercedes-Benz. Ars Technica reports that Boston Dynamics is also a partner. Buyers should ask which model version and robot configuration produced any quoted result.

Safety claims need the same scrutiny. A benchmark score is not a certification. It does not prove that a robot will behave safely around a worker who acts unpredictably. Safety certification for learned robot policies is still an open problem.

DeepMind's benchmark covers refusal, proximity and physical constraints, but it is still one vendor's test. A shared protocol would need participation from rival labs, robot makers and safety regulators. No shared protocol has emerged yet, which is why open data matters more than a single score.

What happens next for humanoid robots in car factories?

The next phase is partner deployment, and Agile Robots has partnered with DeepMind to bring Gemini models into factories. Apptronik calls the pilot a foundation for scaling humanoids in industrial settings. DeepMind says Mercedes-Benz applications are currently focused on kitting.

CoRL is a research conference, not a product launch event. DeepMind used it to reach robotics researchers and hardware partners, the audience that can adopt a partner-only model. That choice signals where the next round of access will come from.

Carmakers are running similar experiments. BMW has brought Figure 03 into its Spartanburg plant for material sorting and sequencing. It also started a project with Hexagon's AEON at Leipzig. Audi has tested Agile Robots' mobile humanoids on parts sequencing.

Xiaomi showed two humanoids sorting center-console side covers through an eight-hour task at its car factory on November 3. Those projects share a pattern. Kitting, sequencing and material movement are repetitive, measurable and close to the existing line.

They are easier to justify than a general-purpose humanoid that walks between stations. Mercedes-Benz says its Apollo robots have collected production data to train for specific use cases. That makes each pilot a training set as much as a productivity test.

Three signals will show whether this release changes manufacturing. The first is independent measurement of success rates, with sample sizes and cycle times. The second is on-device inference that removes the cloud round trip. The third is a safety evaluation that outside labs can reproduce.

DeepMind's own framing goes further than factory kitting. The company describes Gemini Robotics 2.5 as a milestone toward physical AGI, its term for general-purpose embodied intelligence. That is a research ambition, not a shipping claim.

The Mercedes-Benz pilot is the part buyers can inspect. Until then, the practical reading is narrow. A general robot model can now sort car parts with high reliability in a controlled pilot. It still needs a frontier model in the loop, a cloud connection and human oversight.

That is progress, but it is not a factory without people. The gap between 96% and near-perfect is where the next round of competition will happen.

Books behind the context

Sources used for this dispatch

  1. Google DeepMind - Gemini Robotics 2.5: Accelerating Physical Intelligence
  2. Ars Technica - DeepMind reveals Gemini Robotics 2.5 and open safety benchmark ASIMOV-jailbreak
  3. The Verge - Google DeepMind's new AI model gives robots improved dexterity
  4. The Decoder - Google DeepMind's Gemini Robotics 2.5 scores 96 percent sorting metal parts in a Mercedes-Benz assembly pilot
  5. Automate - Google DeepMind Releases Gemini Robotics 2.5
  6. Humanoids Daily - Gemini Robotics 2.5 Puts Apollo 2 on Mercedes-Benz Lines as Argon Speeds Up Training
  7. Hugging Face - google/asimov_agentic dataset
  8. Mercedes-Benz Group - AI and humanoid robots at the Digital Factory Campus Berlin

This article is original analysis and synthesis. Facts, figures and quotations are attributed to the sources above. Images remain the property of their publishers and are credited in the caption.

Questions readers ask

Is Gemini Robotics 2.5 available to the public?

No. DeepMind is rolling it out to robotics partners, including Apptronik and Agile. The company has announced no public API, consumer product or price.

What is a vision-language-action model?

It is a model that takes camera images and a language instruction and outputs robot actions. Gemini Robotics 2.5 is the action component, while a separate embodied-reasoning model tracks longer workflows.

Did Mercedes-Benz independently verify the 96% success rate?

No public independent audit has been published. The figure comes from DeepMind's pilot with Apptronik and Mercedes-Benz, and the companies have not released sample size, pilot duration or cycle-time data.

What is ASIMOV-Jailbreak?

It is an open benchmark on Hugging Face with prompts designed to trick robot models into unsafe actions. DeepMind reports 99.5% resistance for Gemini 4 Argon, not for Gemini Robotics 2.5.

Which robots can run Gemini Robotics 2.5?

DeepMind demonstrated it on Apptronik's Apollo 2 with SharpaWave hands and on the Franka FR3 Duo with parallel grippers. The company says it works across robot types but has published no full compatibility list.

How much faster is Gemini Robotics 2.5?

DeepMind says tasks finish 20% to 30% faster than previous baselines, with 30% to 100% higher success depending on task type. Those figures are company-run and have not been independently replicated.

Topics:Gemini Robotics 2.5Google DeepMindhumanoid robotsMercedes-BenzAI safetyApptronik

More from the research desk