Gemini Robotics 2.5: Google's Humanoids Hit 96% on a Real Mercedes-Benz Line
Google DeepMind has launched Gemini Robotics 2.5, its most capable vision-language-action (VLA) model so far — the kind of AI that watches, understands an instruction, and moves a robot's body. The headline proof is not a lab video: Apptronik's Apollo 2 humanoids running the model worked on active Mercedes-Benz production lines in Kecskemét, Hungary and Sindelfingen, Germany, and picked and packed car parts with a 96% success rate.
What actually improved
- More tasks: DeepMind tested the model across 200+ distinct tasks — from serving food with tongs to tying a bow and pipetting liquids.
- More reliable: success rates rose 30% to 100% over previous models, depending on the type of task.
- Faster: tasks were completed 20% to 30% quicker.
- One brain, many bodies: the same model runs on Apollo 2 (with a five-finger SharpaWave hand, 22 degrees of freedom) and on the two-armed Franka Duo platform with simple grippers.
The Mercedes-Benz test, in plain words
The factory job is called kitting: taking individual parts out of bulk bins and laying them into trays in the exact order each car on the line needs them. It is repetitive and physically tiring — exactly the work factories struggle to staff. Apollo 2 did it on live lines at 96% success. Two honest qualifiers matter: DeepMind calls this a proof of concept, not a permanent deployment, and these are DeepMind's own results — no independent evaluation has been published yet. The lab itself says closing the last 4% will be a steep road, because a factory line cannot shrug off one mistake in twenty-five.
Safety work released alongside
DeepMind also published ASIMOV-Jailbreak, an open safety benchmark for robot AI, on Hugging Face — free for any robotics team to test against. If you build or study embodied AI, that benchmark is the part of this launch you can actually download and use today.
Can you get Gemini Robotics 2.5?
Not yet, unless you are inside the circle. The model is rolling out to DeepMind's robotics partners only — names like Agile Robots, Apptronik and Boston Dynamics, plus 100+ trusted testers. There is no public API, no price and no general release date. For everyone else, this launch is a progress report: humanoid robots have crossed from staged demos to supervised real work, and the remaining gap is reliability, not spectacle.
Source: Google DeepMind — Gemini Robotics 2.5: Accelerating Physical Intelligence (10 November 2026).
Originally published on Ai Game Changer Pro: https://aigamechangerpro.blogspot.com/2026/11/gemini-robotics-25-googles-humanoids.html