Google DeepMind's new robot model is its clearest attempt yet to show general-purpose robots doing paid work rather than lab demos. Gemini Robotics 2.5 is a vision-language-action model, meaning it takes camera input and instructions and outputs movements, and the headline evidence comes from a car factory floor rather than a benchmark table.
DeepMind says it evaluated the model on more than 200 distinct tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid. Against its earlier baselines, success rates rose by 30% to 100% depending on the type of task, and tasks finished 20% to 30% faster. The model runs on different bodies, including Apptronik's Apollo 2 humanoid with a five-finger hand and the two-armed Franka Duo with simple grippers.
The more concrete result is a pilot with Apptronik and Mercedes-Benz at plants in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 robots did kitting: picking metal brackets from bulk bins and placing them in trays in the order each car on the line needs. They reached a 96% success rate, with a separate reasoning model, Gemini Robotics ER, tracking the workflow while 2.5 controlled the arms and hands. On safety, reasoning guidance at run time raised safe task completion around obstacles by 33% and around people by 43%, and DeepMind released ASIMOV-Jailbreak, an open benchmark of hundreds of prompts that hide unsafe requests. On it, Gemini 4 Argon resisted 99.5% of attempts against 82% for the older Gemini Robotics ER 2.
The link to our Gemini 4 Argon review is direct. DeepMind credits Argon with speeding up development by generating thousands of simulated practice runs from a single real camera scene and helping researchers diagnose failures from video. If that holds, the bottleneck in robotics shifts from collecting physical data to how well a general model can imagine it.
Most of this cannot be checked from outside. Every figure comes from DeepMind's own evaluations against its own earlier models, and the 30% to 100% range is too wide to say much about any single task. The model is rolling out only to robotics partners, so independent developers cannot test it. A 96% success rate is a strong pilot, but a production line usually needs higher reliability than that before a robot replaces a person at a station, and the post gives no figures for speed against human workers or for cost.