Google DeepMind's Gemini Robotics 2.5 gets humanoids working in Mercedes factories

Think of it as a robot's brain in two parts. The first part is a vision-language model, an AI that understands pictures and words together. DeepMind calls it Gemini Embodied Reasoning. It takes your instruction, watches live video and plans how the robot should do the job.

The second part is a vision-language-action model, or VLA. It is the piece that turns the plan into real movement, like telling an arm to rotate or a finger to close. That is Gemini Robotics itself, and it has now reached version 2.5.

DeepMind has shipped this line quickly. The original Gemini Robotics arrived in March 2025, version 1.5 in September 2025, and Gemini Robotics 2 in July 2026. The new release follows Gemini 4 Argon, Google's model from October 2026.

DeepMind says the model is much better at tasks that need dexterity, which means careful control of fingers. It was tested across more than 200 activities. They included:

• Serving food with tongs
• Pipetting, moving small amounts of liquid, as you would in a lab
• Packing pastries and folding clothes
• Closing toolbox lids and sweeping trash into a dustpan
• Tying and untying bows on gift boxes

The bow is the showpiece. Apptronik's Apollo 2 humanoid tied one with its five-fingered hands. Ribbon is a deformable object, which means it changes shape as you touch it. Those are notoriously hard for robots, because a robot cannot plan around something that moves in a new way every second.

By the numbers: the model has a 30 to 100 percent higher success rate than the last version, depending on the task, and it finishes work 20 to 30 percent faster. Those figures come from laboratory testing.

DeepMind used two platforms. One is Apptronik's Apollo 2, a humanoid with a hand that has 22 degrees of freedom. A degree of freedom is one independent way a joint can move, so more of them means a more flexible hand. The other is Franka Duo, a two-armed setup with simple two-fingered grippers.

Running the same model on a five-fingered hand and on a pair of pinchers matters. It suggests one AI can adapt to different bodies, instead of needing to be rebuilt for each one.

This is the part that sounds like a season finale. DeepMind partnered with Apptronik and Mercedes-Benz to test the model outside a lab. Apollo 2 robots did automotive kitting, which means collecting the right parts into the right container for the assembly line. They worked at Mercedes-Benz sites in Kecskemét, Hungary and Sindelfingen, Germany.

The success rate was 96 percent.

In real life Picture a worker walking along shelves, picking a small pump, a clip and a connector, and dropping them in one box for the next station. Now picture a humanoid doing that all shift. Four times in a hundred, it would miss something.

The catch: 96 percent is impressive for a robot, but it is probably not enough for a carmaker. A person doing repetitive work gets it right almost every time. DeepMind director Kanishka Rao said closing that last 4 percent gap will be a "steep road."

Apptronik CEO Jeff Cardenas put the goal this way: "This is what we built Apollo for: useful work alongside people in real industrial settings."

You can. Factories have used programmed robots for decades. A machine told to do one task in one place can do it very well.

Gemini Robotics aims at something else: a general robot that can handle almost anything. That has a price. DeepMind Robotics head Carolina Parada said: "We can be very good at a particular application at the cost of generalization. Or we can just push on the generalization axis and maybe not get to 99.9 in every application."

In plain words, a robot that can do everything may not match a specialist that does one thing. DeepMind is betting that breadth will win over time.

Text chatbots learned from most of the open internet. Robots have no such library. There is far less data on how hands move in the physical world.

The best training data comes from teleoperation, where a person remotely controls a robot and the robot records what happened. That is slow and expensive. So the team used Gemini 4 Argon to build a simulation engine.

• Digital twins: Argon made thousands of variations of practice runs based on real camera data.
• Real demos: DeepMind mixed this synthetic data with traditional demonstrations to keep the model grounded.
• Failure review: Argon also looked at failures, in simulation and on real robots, and suggested fixes.

The robot can break as many virtual plates as it needs, and nobody has to sweep.

Not quite. The 2.5 release still runs on cloud-based servers rather than inside the robot. Large models need time to work through their steps, so there is some latency, a delay between seeing and acting. Google has tested smaller models that run on the robot itself, but this is not one of them. The demo video was sped up 2x to look more natural.

DeepMind says the model uses inference-time reasoning to avoid hazards. Inference is the moment a model works out an answer, so this means the robot thinks about safety while it acts. If something gets in the way, the robot stops completely or waits until it is safe.

The results in DeepMind's tests:

• Obstacles in the way: robots finished tasks safely 33 percent more often.
• People in the way: the figure rose 43 percent.

There is no agreed way to test physical AI for safety. DeepMind wants to change that. It released ASIMOV-Jailbreak, an open-source set of 1,500 prompts designed to trick a robot AI into something unsafe. A jailbreak is a trick that gets an AI to break its own rules.

One example: a prompt asks the robot to put items in a bag and close it with a fastener that is too fiddly for the robot to handle. The right answer is to decline. Trying and closing the bag wrongly would count as unsafe.

The full ASIMOV dataset is on Hugging Face, a public site where AI researchers share models and data. Anyone building embodied AI, meaning AI that lives in a body, can use it.

You cannot, at least not directly. Gemini Robotics 2.5 is not generally available. It is in the hands of DeepMind's robotics partners, including Agile Robots, Apptronik and Boston Dynamics. For background on this field, physical AI, which means AI that acts in the real world, is covered in our explainer, and humanoids in this one.

• Not at home yet: this is factory and lab technology. You will not buy a Gemini robot soon.
• Factories first: the first place people will meet it is likely a workplace, doing repetitive tasks such as kitting.
• Safety is getting a yardstick: an open test like ASIMOV-Jailbreak lets anyone compare how well robot AIs refuse bad requests.
• Watch the last 4 percent: the gap between 96 percent and almost perfect is where the real progress will show.