Google DeepMind has been quietly turning the factory floor into a test lab, and this week it showed the receipts. Gemini Robotics 2.5, the company's newest vision-language-action model, is not a demo reel. It is a robot brain measured across more than 200 distinct tasks and then dropped onto live Mercedes-Benz production lines, where it hit a 96% success rate.
That last number is the one worth pausing on. The pilot ran at Mercedes-Benz plants in Kecskemét, Hungary and Sindelfingen, Germany, with Apptronik's Apollo 2 humanoid doing automotive kitting: pulling metal brackets out of bulk bins and placing them into sequencing trays in the exact order each vehicle on the line needs. It is repetitive, physically taxing prep work. It is also exactly the kind of job that has quietly resisted automation for years.
From Tongs to Pipettes: 200 Tasks, One Model
Breadth is where 2.5 makes its case. DeepMind tested serving food with tongs, tying a bow on a gift box, and pipetting liquid, three skills that share almost nothing except the need for real hand intelligence. Against previous baselines the model delivers a 30% to 100% higher success rate depending on task type, while finishing 20% to 30% faster.
Hardware support stays deliberately wide. The same model drives Apollo 2 with its five-finger, 22 degree-of-freedom SharpaWave hand and the Franka Duo platform with simple two-finger parallel grippers. No retraining per robot, no bespoke pipeline for each vendor.
The Mercedes Test: Kitting at 96%
The architecture splits the job in two. Gemini Robotics ER tracks the workflow and reasons over multiple steps, while the 2.5 actions model handles hands and arms, correcting for gaps in the bin as it goes. Apptronik CEO Jeff Cardenas credits the pairing for models that improved dramatically within weeks of post-training and carried over to production with very little drop-off.
Safety got equal billing, which is the right instinct when a 22 degree-of-freedom hand works next to people.
- 96% kitting success rate on active Mercedes-Benz lines
- 30% to 100% higher task success versus previous baselines, and 20% to 30% faster
- 33% better safe completion around obstacles, 43% better around nearby people
- ASIMOV-Jailbreak, a new open-source adversarial benchmark, with Gemini 4 Argon at 99.5% resistance versus 82% for the old baseline
- Rolling out now to robotics partners
There is a nice loop underneath all of this. Gemini 4 Argon helped build 2.5, generating thousands of simulated practice runs from a single real camera scene, constructing digital twins, and hypothesizing why a task failed before testing the fix both in simulation and on actual hardware. The frontier model is now the research assistant for the robot model.
DeepMind frames 2.5 as a milestone on the path to physical AGI. Softer framing aside, the practical read is simpler: the gap between a lab demo and a shipping dock just narrowed, and the robot doing the packing has a 96% hit rate to prove it.
Comments