AI robots just passed their first real factory test
Google DeepMind's Gemini Robotics 2.5 hit 96% success on a live Mercedes-Benz production line, showing that AI-controlled robots are moving from impressive demos to actual production work.
- 96%
- success rate in a live Mercedes-Benz factory
- 200+
- distinct tasks tested
- 20–30%
- faster task completion vs. previous version
- 3
- robot hardware partners in first rollout
For years, robots looked impressive in labs but kept failing on real factory floors — things move, parts vary, lights change. That just shifted. Google DeepMind's Gemini Robotics 2.5 was tested on a live Mercedes-Benz production line in Hungary and Germany, where Apptronik's Apollo 2 humanoid robot picked and packed car parts with a 96% success rate. This was not a staged demo. It was a working factory.
The model acts as an AI brain for the robot's whole body. It takes in what cameras see, understands plain-language instructions, and translates them into precise hand and arm movements. It was evaluated across more than 200 distinct tasks — handling tools, deformable objects, and delicate parts — and completed them 20 to 30% faster than the previous version.
The bigger story is not just speed. Traditional industrial robots need months of custom engineering every time a task changes. With a general-purpose AI brain, a robot can be retrained for a new task much faster using the same model. That means the cost per new task drops sharply — which is what makes this commercially interesting, not just technically impressive.
The model is rolling out to hardware partners first — Agile Robots, Apptronik, and Boston Dynamics — not to the general public yet. But the direction is clear. If your business involves repetitive physical work at scale, start mapping which tasks are routine and rule-based. AI-controlled robots are not ready for everything, but their task list is growing fast, and the ones worth deploying cost less to reach than they did a year ago.