Google Gemini Robotics 2.5 Hits 96% in a Mercedes Factory Pilot — Partners Only for Now
Google DeepMind says its newest robot model helped a humanoid pick and pack car parts with a 96 percent success rate inside a Mercedes-Benz factory pilot. The model is Gemini Robotics 2.5, announced on 10 November 2026, and the pilot ran with Apptronik's Apollo 2 humanoid on active production lines in Hungary and Germany. The task was automotive kitting, sorting parts from bulk bins into trays and sequencing containers in vehicle order, and the 96 percent figure is Google-reported, with no trial count published. This is a proof of concept on one task, not a production rollout and not a fleet purchase: Mercedes has not announced buying robots, and the model is available to robotics partners only, with no public API, no price, and no general availability date.
What is Gemini Robotics 2.5
Gemini Robotics 2.5 is a vision-language-action model, often shortened to VLA. A vision-language-action model is an AI model that takes in what a robot sees, understands instructions in ordinary language, and turns that understanding into physical actions. The vision part reads the scene through cameras, the language part understands the instruction, and the action part decides how the robot's arms and hands should move. That combination lets a robot handle a task described in words, such as sorting parts into trays in vehicle order, rather than only repeating a fixed motion it was programmed to perform.
DeepMind calls Gemini Robotics 2.5 its most capable vision-language-action model to date. The claim is DeepMind's own, and the evidence it offers is an evaluation across more than 200 distinct tasks. Across those tasks, DeepMind reports a success rate 30 percent to 100 percent higher than previous baselines, depending on the task type, and tasks completed 20 percent to 30 percent faster. Both ranges are wide, which means the improvement varies a great deal from one kind of task to another. All of these figures are DeepMind-reported, with no independent check, and the announcement does not publish the underlying task list or the baseline figures.
The Mercedes factory pilot: what actually happened
The headline result comes from a proof of concept with Mercedes-Benz and Apptronik, using Apptronik's Apollo 2 humanoid robot. DeepMind says the model helped Apollo 2 pick and pack automotive parts with a 96 percent success rate. That sentence needs careful reading: the 96 percent figure is a success rate, and no trial count was published, so it must not be converted into a count of successful attempts. The figure is also Google-reported, and it describes one task in a pilot, not the robot's performance across all factory work.
The task itself is specific and genuinely hard for a robot. Automotive kitting means sorting parts from bulk bins, where parts arrive jumbled together, into trays or sequencing containers arranged in vehicle order, so that the right parts reach the line in the right sequence for each car. A robot doing this work must recognise different parts in a messy bin, pick the right one without damaging it, and place it in the correct position for the correct vehicle. A 96 percent success rate on that task would be a meaningful result for a factory, but it is a pilot result on one task.
The pilot ran on active production lines in two places: Kecskemét in Hungary, and Sindelfingen in Germany. Active lines means the pilot took place in real factories where cars are actually built, not in a laboratory or a demonstration hall. That matters, because a robot that works in a lab can still fail on a real line, where lighting, noise, timing, and the flow of parts are all less controlled. Running on active lines is a stronger test than a lab demonstration, but it is still a proof of concept, and the announcement does not say how long the pilot ran or what happened to the attempts that did not succeed.
 |
| A humanoid robot arm picking up metal parts in an automotive factory |
The robots: Apollo 2 and Franka Duo
Gemini Robotics 2.5 runs on two robot platforms named in the announcement: Apptronik's Apollo 2 and the Franka Duo. Apollo 2 is the humanoid used in the Mercedes pilot, and its hand is part of the story. The robot uses a five-finger hand called the SharpaWave hand, with 22 degrees of freedom, a measure of how many independent ways the hand can move. A hand with 22 degrees of freedom can shape itself around different objects in ways a simple gripper cannot, which makes it plausible for picking varied car parts out of a bulk bin. The Franka Duo is the other platform named, a two-armed system.
| Result | Figure, per DeepMind |
|---|---|
| Apollo 2 kitting success rate, Mercedes pilot | 96% |
| Distinct tasks in evaluation | 200-plus |
| Success rate vs previous baselines | 30% to 100% higher, by task type |
| Task completion speed vs previous baselines | 20% to 30% faster |
| Gemini 4 Argon jailbreak resistance, ASIMOV-Jailbreak | 99.5% vs 82% prior baseline |
Partners only: who can use it, and who cannot
Gemini Robotics 2.5 is rolling out to robotics partners only. The partners named in the announcement are Agile, Apptronik, and Boston Dynamics, and secondary reporting puts the number of trusted testers at more than 100. There is no public API, no published price, and no general availability date. That means a developer, a small factory, or a researcher cannot sign up and start using the model today. Access runs through DeepMind's partner programme, and the announcement gives no date by which that will change.
The partner-only route reflects the reality that a robot model is not the same as a chatbot. A language model that gives a wrong answer is an annoyance; a robot model that moves a heavy arm in the wrong direction can damage equipment or hurt someone. Rolling out to partners first lets DeepMind test the model on real hardware with teams that know how to run robots safely. Readers who want the wider context of Google's model family can read our flagship AI model comparison, our coverage of Gemini 4 Carbon, and our report on Genie 4.
Safety testing: the ASIMOV-Jailbreak benchmark
Alongside the model, DeepMind has open-sourced a new safety benchmark called ASIMOV-Jailbreak. A jailbreak, in this context, is an attempt to get an AI model to do something it is supposed to refuse, by phrasing the request in a misleading way. For a robot, that risk is physical: a jailbroken robot model might be talked into moving in a way that is unsafe. The benchmark tests how well a model resists such attempts, and DeepMind reports that Gemini 4 Argon achieves 99.5 percent jailbreak resistance on it, against an 82 percent prior baseline. That figure is DeepMind-reported, and it describes Argon, not Gemini Robotics 2.5 itself.
 |
| An automotive assembly line with robotic arms and car body parts |
What the pilot does not mean
Three misunderstandings are worth correcting directly. First, the pilot is not a production rollout: a proof of concept on one task is not the same as robots working across a factory. Second, it is not a fleet purchase. Mercedes has not announced buying robots, and no order, quantity, or price has been published. Third, the 96 percent figure is not a guarantee for other tasks. It describes kitting on Apollo 2 in two factories, and a robot that sorts parts well may still struggle with a different task or on different hardware.
 |
| An engineer in a hard hat and high-visibility vest observing a humanoid robot in a factory |
Key Takeaways
- Gemini Robotics 2.5, DeepMind's most capable vision-language-action model to date, helped Apptronik's Apollo 2 pick and pack automotive parts with a 96 percent success rate in a Mercedes factory pilot, per DeepMind.
- The pilot ran on active lines in Kecskemét, Hungary, and Sindelfingen, Germany, on automotive kitting: sorting parts from bulk bins into trays or sequencing containers in vehicle order.
- DeepMind reports the model was evaluated across 200-plus tasks, with success rates 30 percent to 100 percent higher than previous baselines and tasks completed 20 percent to 30 percent faster.
- The model runs on Apollo 2, with a five-finger, 22-DoF SharpaWave hand, and Franka Duo, and is rolling out to robotics partners only, with no public API, price, or general availability date.
- This is a proof of concept on one task, not a production rollout or fleet purchase, and Mercedes has not announced buying robots; DeepMind also open-sourced the ASIMOV-Jailbreak benchmark, reporting 99.5 percent jailbreak resistance for Gemini 4 Argon versus an 82 percent prior baseline.
Conclusion
Gemini Robotics 2.5 has cleared a real hurdle: a 96 percent success rate on automotive kitting, on active Mercedes lines, with a humanoid robot. The broader evaluation suggests the model is capable across many tasks, and the open-sourced safety benchmark is a welcome step for a field where a wrong move has physical consequences. But the pilot covers one task, the model is partners only with no public access or price, and Mercedes has not bought a fleet.
Comments
Post a Comment