Google’s new robot brain gets a real job: sorting brackets on a Mercedes-Benz line

Kitting is about as unglamorous as factory work gets. Someone pulls parts out of bulk bins and lays them into trays in the exact order a specific car on the line will need them, so the technician doing the actual build never has to stop and hunt for the right bracket. It is repetitive and physically draining, and it changes every time the vehicle mix changes. That last part is why it has resisted ordinary automation, and why it is now the first real job Google DeepMind is pointing to for its newest robot model.

On November 10, DeepMind released Gemini Robotics 2.5, the latest version of its vision-language-action model, the kind of AI that looks at camera images and outputs arm and hand movements. The launch came with a factory result attached. Running the new model, Apptronik’s Apollo 2 humanoid sorted metal automotive brackets from bulk bins into sequencing trays on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, at a 96 percent success rate.

Two models, one robot

The setup splits the thinking from the doing. A separate model, Gemini Robotics ER, keeps track of where the robot is in the workflow, while Gemini Robotics 2.5 drives the hands and arms and corrects when a tray slot is left empty. Apollo 2 was fitted with the SharpaWave hand, a five-fingered hand with 22 degrees of freedom.

Apptronik’s co-founder and CEO Jeff Cardenas said the models “improve dramatically in just a few weeks of post-training” and carried over to Mercedes-Benz production “with very little drop-off.” That second claim is the one to watch. Lab-to-factory drop-off is where a lot of impressive robot demos quietly fall apart.

What the 96 percent does and doesn’t cover

DeepMind calls this a proof of concept, not a rollout, and the public numbers are thin in a few places that matter to anyone running a plant. The announcement doesn’t say how many robots took part, how many picks the 96 percent is based on, how long the pilot ran, or how Apollo’s pace compared with a person doing the same job. A 96 percent hit rate also means roughly one placement in 25 still needs fixing, which is why a pilot like this runs next to people rather than instead of them.

Beyond the factory, DeepMind tested the model on more than 200 tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid, on both Apollo 2 and a Franka Duo setup with two-fingered grippers. It reports a 30 to 100 percent higher success rate than its previous baseline, depending on the type of task, and 20 to 30 percent faster completion. Those are the company’s own measurements and haven’t been checked independently.

Safety gets its own numbers

DeepMind is also leaning on safety. It says reasoning at run time helps the robot steer around hazards, which raised safe task completion by 33 percent around obstacles and by 43 percent with people nearby. It also released ASIMOV-Jailbreak, an open-source benchmark of prompts that try to disguise unsafe requests to a robot. DeepMind says Gemini 4 Argon resisted 99.5 percent of those attempts, compared with 82 percent for Gemini Robotics ER 2. Argon is the same frontier model that taught Apollo to play cornhole from a video last month, and DeepMind credits it with speeding up this release by generating thousands of simulated practice runs from real camera scenes and helping researchers work out why tasks failed.

The race for the robot brain

The release lands in a crowded week for robot intelligence. At the Conference on Robot Learning in Austin, NVIDIA unveiled GR00T 2, its own model for humanoids. Startups such as Physical Intelligence, Skild AI and Field AI are all pitching one brain for many robots, and Figure builds its own Helix model in-house, most recently testing it on chores in homes it had never seen.

What sets this release apart is the paying customer in the frame. Mercedes-Benz invested in Apptronik in 2025, so this is a partner putting a partner’s robot on its own line rather than a lab staging a demo. It is the same pattern that has helped Agility Robotics, whose first customer keeps expanding Digit to more warehouses. A factory owner willing to let a robot work beside its own staff is still the most convincing evidence in this industry.

Our take

Gemini Robotics 2.5 is going only to DeepMind’s robotics partners for now, with no public API or pricing announced. So the next real signal is not another benchmark. It is whether Mercedes-Benz moves kitting from proof of concept into everyday production, and whether the model performs as well on robots other than Apollo. Until then, 96 percent on one well-chosen task is a strong start, not a finished case. Compare Apollo with other machines on our humanoids page, follow Apptronik for updates, and look for robot learning and manipulation roles on our jobs board.