Gemini Robotics 2.5 hits 96% kitting on Mercedes-Benz lines, with Gemini 4 Argon building its practice runs

Google DeepMind introduced Gemini Robotics 2.5, its most capable vision-language-action model so far, rolling out to robotics partners. It claims 30% to 100% higher success than prior baselines depending on task, 20% to 30% faster completion, and evaluation across 200 tasks from serving food with tongs to tying a bow on a gift box to pipetting liquid. Models work across bodies, including Apptronik’s Apollo 2 with a five-finger, 22-degree-of-freedom hand and the Franka Duo with parallel grippers.

The number that will travel is 96% success in a proof-of-concept with Apollo 2 doing automotive kitting on active Mercedes-Benz lines in Kecskemét, Hungary, and Sindelfingen, Germany. The Gemini Robotics ER embodied-reasoning model tracked the workflow while the 2.5 actions model controlled hands and arms. Apptronik CEO Jeff Cardenas said the models improved dramatically within weeks of post-training and carried over to production with little drop-off.

Kitting, sorting parts from bulk bins into trays in the exact order each vehicle needs, is repetitive and physically taxing. Handing it to a robot so technicians focus on complex builds is the kind of automation worth arguing for. But it is a pilot, and on a production line the remaining 4% is stoppages, human correction and cost. The question for Mercedes-Benz isn’t whether the robot can do the task but whether failure modes are cheap enough at scale. DeepMind calls it promising progress, the right register.

The most consequential detail is Gemini 4 Argon. DeepMind says the frontier model accelerated this work as a simulation engine, generating thousands of varied practice runs from a real camera scene, building digital twins and helping researchers hypothesize why failures happen from video, extending the video-to-sim-to-real work it showed earlier (Argon for robotics). That is AI speeding AI research, now in the physical world, and it compounds: better models produce better robot training data, which yields more real-world data. Rivals are chasing the same loop with world models (Genie 4, Cosmos).

DeepMind reports inference-time guidance that improved safe completion around obstacles by 33% and around nearby people by 43%. It is releasing ASIMOV-Jailbreak, an open benchmark of disguised unsafe requests, and says Argon reached 99.5% jailbreak resistance versus 82% for its earlier baseline. An open benchmark lets outsiders check the work.

Robots that can be talked into things are a different risk category than chatbots, and a 0.5% gap matters when the system has arms. The labor shift won’t stop at kitting either; partners should say as much about redeploying displaced work as about success rates. More in our Google coverage and robotics coverage.