Gemini Robotics 2.5: DeepMind crosses the 96% success rate threshold on a Mercedes-Benz production line
🔎 96% on the factory floor: the number humanoid robotics has been waiting for (and what it's hiding)
On November 10, 2026, Google DeepMind lifted the curtain on Gemini Robotics 2.5, the new generation of its vision-language-action (VLA) model. The showcase: a proof of concept carried out with Mercedes-Benz, where Apollo 2 humanoids from Apptronik sort automotive parts with a 96% success rate on the lines of Sindelfingen (Germany) and Kecskemét (Hungary).
The timing is no accident. Humanoid robotics is leaving viral videos behind for the workshop floor: China is industrializing at full speed while the West stacks up pilot projects, as we noted in our analysis of Automate 2026. In this context, a VLA that posts an industrial success rate — not a staged demo — becomes a geopolitical argument as much as a commercial one.
But a figure pulled from a press release deserves as much skepticism as admiration. DeepMind publishes neither the number of trials, nor the exact definition of success, nor the frequency of human intervention, as TechTimes points out. This article examines both sides: what the announcement actually delivers, and what it skillfully glosses over.
The Essentials
- A claimed generational leap: evaluated on more than 200 tasks, Gemini Robotics 2.5 posts success rates 30 to 100% higher than previous baselines — the blog's charts naming Gemini Robotics 2, the previous generation, as the reference — and tasks executed 20 to 30% faster (DeepMind).
- The industrial proof: 96% success on automotive kitting at Mercedes-Benz, versus 82% for Gemini Robotics ER 2 on the same exercise.
- The blind spots: no number of trials, no definition of success, no frequency of human intervention; the safety gains (+33% around obstacles, +43% around people) are self-assessed.
- Two maturity signals: an open benchmark of robotic jailbreaks (ASIMOV-jailbreak) and a rare admission — Kanishka Rao, director of robotics, concedes that closing the remaining 4% will be "an uphill road".
- The structural catch: the model still runs in the cloud, on-device latency remains an open question, and neither pricing nor a general availability date has been announced.
Recommended tools
The announcement's ecosystem comes down to four building blocks. Here they are, with their actual status:
| Tool / Player | Role | Price (November 2026) | Key takeaway |
|---|---|---|---|
| Gemini Robotics 2.5 | Low-level VLA: controls arms and hands | Not disclosed | 200+ tasks, +30 to 100% success rate vs previous generation |
| Gemini Robotics ER | "High-level brain": follows the assembly workflow | Not disclosed | Already exposed on Google Cloud (see our analysis) |
| Apollo 2 | Apptronik's humanoid, the physical platform of the pilot program | Not disclosed | Tested at Mercedes-Benz, not to be confused with an off-the-shelf product |
| ASIMOV-jailbreak | Open robotics safety benchmark | Free (open access) | Robotics-specific jailbreaks, published by DeepMind |
What exactly is Gemini Robotics 2.5?
Answer: a unique model that transforms what the robot sees and reads into executed gestures, without a classic programming pipeline.
A VLA (vision-language-action) merges three capabilities that robotics has treated separately for thirty years: perceiving the scene, understanding a command in natural language, producing motor commands. The promise is not to perform a task, but to learn tasks. This is the entire difference between an automaton and a learning system.
The architecture remains hierarchical, however. At Mercedes, Gemini Robotics ER acts as the supervisor: it understands the request and follows the assembly workflow specific to each vehicle. Gemini Robotics 2.5, for its part, pilots the arms and hands, gesture by gesture. This brain/hand separation is DeepMind's structural bet—and it has an obvious business consequence: the brain is rented on Google Cloud.
Regarding skills, the announcement highlights gestures that were previously out of reach: serving with tweezers, tying a ribbon, pipetting. It is not a coincidence that these three examples are cited. Fine manipulation—delicate grasps, deformable objects, gestural precision—remained the historical weak point of humanoids. Twelve months ago, these tasks still separated the laboratory from the factory.
Simulation as an Accelerator
Another detail from the announcement: Gemini 4 Argon served as the simulation engine to accelerate the model's development. The logic is simple—generating years of synthetic experience rather than weeks of costly teleoperation. All major laboratories take this gamble; DeepMind has the unique characteristic of possessing the simulator, the model, and the cloud that connects them.
From my point of view, the real novelty of this version is not the 96%, it is the expansion of the catalog of learnable skills. A single model capable of serving with tweezers, tying a ribbon, and pipetting is a platform. An automaton does only one thing; a VLA claims to learn all the others.
The numbers: +30 to 100% success rates, but compared to whom?
Answer: to DeepMind's own previous generation — real progress, but an in-house comparison.
| Announced metric | Figure | Critical reading |
|---|---|---|
| Evaluation volume | More than 200 tasks | Credible volume for a generalist VLA |
| Success rate | +30 to 100% vs baseline | Baseline = Gemini Robotics 2, the previous generation |
| Execution speed | 20 to 30% faster | In-house measurement, protocol not published |
| Mercedes-Benz kitting | 96% (vs 82% for Gemini Robotics ER 2) | One task, two sites, a proof of concept |
| Safety | +33% (obstacles), +43% (people) | Developer figures, no external evaluator |
The range of the gain — from +30 to +100% — is paradoxically the most honest piece of information in the press release. It suggests that progress is massive on new fine-manipulation tasks (where the previous generation often failed) and more modest on tasks already mastered. That's coherent. But it also rules out summarizing the model with an average "+X%": the figure depends entirely on the task in question.
The underlying problem: in robotics, there is not yet an equivalent of the public leaderboards that pit GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.7 against each other in the LLM world (see our 2026 comparison). Each lab measures with its own tape measure, on its own tasks, with its own definition of success. A figure without a shared benchmark is a statement, not a measurement.
Mercedes-Benz: what really happened on the line?
Answer: a single, bounded, repetitive task — kitting — carried out at two sites, not an end-to-end AI-driven factory.
Let's recall what kitting is, because it's the heart of the matter: the robot picks parts from bins and arranges them on trays, in the assembly order specific to each vehicle. Every order generates a variant, so every tray differs. It's a tiring logistics task, costly in labor, and one where human error is statistically commonplace.
This choice of task is clever — and revealing. The scope is bounded, coactivity with operators is limited, errors remain recoverable downstream of the line, and the return on investment is immediate for a manufacturer that pays for these gestures three times a day. If I had to choose THE task to demonstrate a VLA in a factory, it would be exactly this one. DeepMind didn't take the hardest case; it took the most marketable one.
The 14-point gap with Gemini Robotics ER 2 — 96% versus 82% — is perhaps the most useful piece of data in the announcement. In the automotive industry, these points separate "project worth funding" from "deployable process": at 96% over thousands of daily cycles, several errors per hour remain, each one costing operator time. Interesting Engineering confirms the figure and the Apollo 2 platform, but no one, at this stage, has published the trial volume behind it.
What DeepMind Doesn't Say (and Why It Matters)
Answer: three omissions make it impossible to assess the system's real robustness — and one piece of information does, in fact, exist.
First omission: the number of trials. 96% over 25 cycles and 96% over 25,000 cycles do not tell the same story. Without a sample size, there is no confidence interval; therefore no science, just a shop window.
Second omission: the definition of success. Does a misoriented part then corrected by the robot count as a success? A cycle stopped by an operator? This boundary between "succeeded" and "recovered" is precisely what a factory director cares about — and it is nowhere defined.
Third omission, the heaviest of all: the frequency of human intervention. This is THE number that decides a deployment: how many times per hour must an operator clear, reset, or correct the robot? As long as it isn't published, the 96% remains a PR number.
On the safety side, the +33% of safe tasks around obstacles and +43% around nearby people are developer-reported figures, with no external evaluator. In pharma, aerospace, or automotive, you don't certify a system on the manufacturer's word alone. And the same TechTimes article relays the most honest admission of the announcement: Kanishka Rao, DeepMind's robotics director, admits to Ars Technica that closing the remaining 4% will be "an uphill climb".
My take: the last four points of a physical system are almost always the most expensive — deformed parts, changing lighting, poorly positioned bins, a long tail of unforeseen cases. Saying it will be hard is healthy. Publishing the methodology would be healthier.
ASIMOV-jailbreak: robotic safety moves to open benchmarks
Answer: DeepMind is releasing an open benchmark dedicated to robotics-specific jailbreaks — a rare move, and perhaps the most strategic part of the entire announcement.
A reminder of how this differs from language: a chatbot jailbreak produces inappropriate text. A robot jailbreak produces an inappropriate action — with motors, force, and mass behind it. As VLAs come to understand language and act in the physical world, the attack surface changes in nature. It's no longer about filtering responses, but about ensuring that an adversarial instruction doesn't turn a robotic arm into a hazard for its surroundings.
ASIMOV-jailbreak, picked up by Superpower Daily citing Ars Technica, gives the community a shared testing ground for this risk. It's an open safety benchmark — so anyone can test it, critique it, extend it.
My take: releasing a safety evaluation tool before selling a product is both an act of openness and a bold bet to claim the ground. If ASIMOV-jailbreak becomes the industry reference, DeepMind is writing the rules of the very game it will play itself. Competitors — and regulators — should seize it quickly, and extend it to performance, not just safety.
The cloud, the VLA's silent Achilles' heel
Answer: the model still runs on remote servers, and on-board latency remains an open question for real-world production.
A VLA controlling hands runs a continuous loop — perception, decision, motor — dozens of times per second. When that loop goes through the cloud, every network millisecond becomes a safety parameter in its own right, just like a sensor. Yet DeepMind publishes neither measured latency nor a fallback architecture for when the connection drops. In a factory, losing the link isn't an inconvenience: it's a workstation at a standstill.
For Mercedes-Benz and its network engineering, the problem is solvable. For a mid-sized subcontractor, it's a dealbreaker as things stand. Scaling industrial VLAs will therefore require on-board or edge versions — a challenge no one has yet settled, and one that Monday's announcement did not address.
The business reality also needs to be named. Google sells the brain and the cloud that brain runs on: we flagged this back when Gemini Robotics ER 2 arrived on Google Cloud. Every deployed robot is also an infrastructure contract. That's not objectionable in itself — it's the model — but an industrial company must cost it as such: a hardware dependency (Apptronik) compounded by a dependency on intelligence and compute (Google).
The global race: where does Google actually stand?
Answer: at the top of control software, with Europe's most premium industrial proof — in a West that remains, overall, at the pilot stage.
The landscape in three moves. Xiaomi put its humanoids to the test with 8 hours of sorting auto parts in a factory. Panasonic and Berkeley presented T-Rex, a bimanual robot capable of correcting its own movements. And at Automate 2026, the verdict was clear: China is moving to mass production while the West is still piling up demonstrators (see our analysis). Against this backdrop, DeepMind brings the argument the West was missing: a VLA proven on a real automotive production line, with a premium carmaker as its showcase.
Let's also map out Google's real competitors. On general-purpose models, OpenAI (GPT-5.5), Anthropic (Claude Opus 4.7) and Google are vying for the honors — and the ChatGPT vs Gemini duel is playing out mainly in browsers and desktops. But physical robotics is a field where most of the LLM giants are not yet present with a product proven in a factory. That is precisely where Google has chosen to dig in.
My take: the strategy is clear. Sell the infrastructure of physical intelligence — model, orchestration, cloud — rather than the metal, left to partners like Apptronik or Agile Robots, both thanked in the press release. If the model delivers on its promises beyond the Mercedes lines, Google becomes the brain supplier for an entire industry. If latency or the remaining 4% holds out, that lead will melt away fast.
How to prepare for this, concretely
Answer: three workstreams you can start on right now, without access to the model.
Developers: modern robotics is first and foremost a software problem — simulation, integration, model orchestration. The highest-ROI lever today is still to code better and faster; our selection of the best LLMs for coding (Claude, GPT, Gemini, DeepSeek) is made for exactly that.
The curious and students: before speculating about VLAs, calibrate your intuition about what current models can — and cannot — do. Our best free LLMs (ChatGPT free, Gemini, Groq) let you test the language and reasoning layer that underpins these systems without spending a dime.
Industrial players: map out your "kitting-compatible" tasks — bounded, repetitive, with recoverable errors, and low co-activity. These are the ones that will be automated first by this type of model. And plan ahead for the networking question: a cloud VLA assumes flawless local infrastructure.
Finally, if you want to experiment with model orchestration or host your own agents, an entry-level VPS is enough to get started — Hostinger offers plans starting at a few euros per month (November 2026, check pricing on hostinger.com).
❌ Common Mistakes
Mistake 1: Reading 96% as "production-ready"
A success rate measured in a proof of concept, on a well-defined, supervised task, is not a production rate. The remaining 4% are precisely where the costly cases are concentrated — Rao himself speaks of a "steep uphill road." The safeguard: require the number of trials and the frequency of human intervention before making any investment.
Mistake 2: Believing the intelligence is in the robot
The architecture is hierarchical and distributed: Gemini Robotics ER supervises, Robotics 2.5 executes, all via the cloud. Believing the humanoid decides on its own means underestimating the network infrastructure and vendor dependence. The safeguard: evaluate a VLA deployment as an IT infrastructure project, not as the purchase of a machine.
Mistake 3: Mistaking safety gains for certifications
+33% around obstacles, +43% around people: these are developer-reported measurements, with no external evaluator and no published protocol. The safeguard: wait for independent audits and open benchmarks — ASIMOV-jailbreak points in that direction, but it only covers jailbreaks, not overall operational safety.
Mistake 4: Interpreting "no price" as "affordable"
As of November 10, 2026, there is neither a price tag nor a general availability date. Historically, deployments of this kind are negotiated case by case, at price levels with no relation to a software subscription. The safeguard: treat any public pricing figure as premature, and follow the news on deepmind.google.
❓ Frequently Asked Questions
Is Gemini Robotics 2.5 available?
No. On November 10, 2026, DeepMind announced a deployment to its partners—Agile Robots, Apptronik, and Mercedes-Benz—without a general availability date or pricing. Previous iterations went through partner access and then an exposure on Google Cloud; this is the most likely scenario, but no official timeline has been published.
What exactly is a VLA?
A vision-language-action model simultaneously learns to perceive (cameras), understand (natural language instructions), and act (motor commands). Unlike a robot programmed task by task, a VLA can acquire new skills—serving with pincers, tying a ribbon, pipetting—without a complete rewrite of the control, which changes the economics of automation.
What exactly does the robot do at Mercedes-Benz?
Kitting: it picks parts from bins and places them on trays in the assembly order specific to each vehicle, with a claimed 96% success rate. Apptronik's Apollo 2 humanoids execute the gesture in the factories in Sindelfingen (Germany) and Kecskemét (Hungary), while Gemini Robotics ER follows the assembly workflow.
Why are the remaining 4% so difficult?
Because they concentrate the long tail: deformed parts, unusual lighting, unforeseen positions, cases never encountered during training. Every point gained costs more than the previous one, and in production, an unrecovered failure can immobilize a workstation. Kanishka Rao himself speaks of a "steep road" to close this final gap.
Should we believe the +33% and +43% safety figures?
They are plausible but self-evaluated: DeepMind is both judge and jury, with no external evaluator or published protocol. They should be treated as a statement of intent, not as a certified measure. The open benchmark ASIMOV-jailbreak is a first step toward independent evaluation—for safety, not yet for performance.
Can the robot function without a cloud connection?
This is precisely an open question. The model currently runs on remote servers, and DeepMind publishes neither measured latency nor a fallback architecture. For a well-equipped industrial site, it is manageable; for mass deployment, embedded systems or edge computing will likely be essential.
✅ Conclusion
Gemini Robotics 2.5 delivers the most serious industrial proof ever published by a major AI lab on a European automotive production line — and the most methodically incomplete. The 96% deserves to be taken seriously; the remaining 4% will decide everything. To put this announcement in the context of the global race, continue with our analysis of Automate 2026 and the reality check of humanoid robotics.