The robot demo is dead; the robot shift has begun — or so Google DeepMind would have you believe. On Tuesday it unveiled Gemini Robotics 2.5, its most capable vision-language-action model to date. The headline is not a lab benchmark: Apptronik's Apollo 2 humanoid picked and packed car parts on active Mercedes-Benz lines in Hungary and Germany, hitting a 96% success rate, according to Google DeepMind.

The job was kitting: pulling parts from bulk bins into trays in the exact order each vehicle needs. DeepMind's reasoning model tracked the workflow while 2.5 drove the hands and arms, reaching into bins for metal brackets and correcting for gaps.

The model is a generalist. DeepMind tested it on 200+ tasks — serving food with tongs, tying a bow, pipetting liquid — reporting success rates 30–100% higher than its previous models, and tasks 20–30% faster. It runs on different bodies, from Apollo 2's five-finger SharpaWave hand to the Franka Duo. "Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off," said Apptronik co-founder and CEO Jeff Cardenas. "This is the foundation for bringing humanoid robots to industrial and commercial environments at scale."

Now the fine print. DeepMind has published no trial count, no definition of a successful pick, and no word on how often a human stepped in. Mercedes has not said the robots are staying; this was a proof of concept, not a deployment. A 96% rate over 25 attempts and over 1,000 are very different claims, and only DeepMind knows which it is.

DeepMind also shipped real-time safety steering — safe completion around obstacles up 33%, around people up 43% — and open-sourced ASIMOV-Jailbreak, a benchmark that tries to trick robot AI into unsafe actions. Gemini 4 Argon resisted 99.5% of attempts, against 82% before.

The timing is the other story: one day after Nvidia's GR00T 2, which learned robot skills by watching humans and claimed 95% on its own tasks. The robot-brain race now moves weekly — Nvidia answered with a benchmark, Google with a factory floor. Watch whether Mercedes keeps the robots past the pilot, and whether any independent lab reproduces 96%. Until then, the honest reading is also the exciting one: a humanoid did real work on a real line, and it mostly worked.

— via Google DeepMind