Gemini Robotics 2.5 kits car parts at Mercedes with 96%
Apptronik's Apollo 2 picked brackets from bulk bins on active Mercedes-Benz lines in Hungary and Germany, and DeepMind calls it a proof of concept.
Google DeepMind released Gemini Robotics 2.5, its most capable vision-language-action model so far, and ran it on Apptronik's Apollo 2 humanoids inside Mercedes-Benz plants. Lukas Ziegler broke the news, noting it arrives just four months after Gemini Robotics 2. The robots worked on active production lines in Kecskemét, Hungary and Sindelfingen, Germany.
The task is automotive kitting. The robot picks metal brackets out of bulk bins and places them into sequencing containers in the exact order each vehicle needs. Carolina Parada of Google DeepMind Robotics says Gemini Robotics 2.5 helped Apollo 2 pick and organize car parts for assembly with 96% accuracy. Ziegler reports Apptronik CEO Jeff Cardenas said the models improved dramatically after just a few weeks of post-training, and that the gains carried over to Mercedes-Benz production with very little drop-off.
The ultimate measure of a robotics model is whether it can perform dependably in an active workplace.
Shown versus claimed
Parada frames this as a proof of concept with Apptronik and Mercedes-Benz, not a production rollout, and says there is lots more work to be done. The 96% number comes from Google and its partners. No outside tester has checked it.
One model, several bodies
Ziegler says the model runs natively across robot types, including Apollo 2 with the five-finger, 22-DoF SharpaWave hand and the Franka Robotics Duo with parallel grippers. That matters because most VLAs (models that turn camera images and text into robot actions) are tied to the hardware they trained on.
Safety got the bigger numbers in the release. Inference-time reasoning guidance steers the robot around physical hazards in real time, and Ziegler reports safe task completion is up 33% around obstacles and 43% around nearby people. DeepMind also open-sourced ASIMOV-Jailbreak, a benchmark for adversarial attacks on robots, where Gemini 4 Argon reaches 99.5% jailbreak resistance against 82% for the previous baseline.
Models building models
Gemini 4 Argon helped build Gemini Robotics 2.5. Ziegler says it acted as a simulation engine, generating thousands of varied practice runs from a single real-world camera scene and building a digital twin. It also helps researchers work out why tasks failed from video footage, then test fixes in simulation and on real robots.
Keerthana Gopalakrishnan posted factory footage of the robots sorting car parts in Sindelfingen and called the pace surreal. The model is rolling out to robotics partners now.
As a researcher it is quite surreal to see robotics go from not working at all, to inventing VLAs, to now real-world usage across the world
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.