DeepMind's Gemini Robotics 2.5 hits 96% task success in Mercedes-Benz factory test
Google DeepMind presented Gemini Robotics 2.5 at the Conference on Robot Learning in Austin. It is a vision-language-action (VLA) model, meaning it takes camera input and language instructions and produces robot actions. DeepMind evaluated it on more than 200 physical tasks. The headline result comes from an industrial deployment: in automotive parts handling on Apptronik's Apollo 2 humanoid at Mercedes-Benz facilities, the model reached a 96% success rate and worked 20% to 30% faster than earlier baselines.
The architecture splits planning from control. A Gemini Embodied Reasoning planner decomposes a task and reasons about the scene, while the VLA executes manipulation. DeepMind says the combination makes physical tasks more accurate, faster and safer, including in industrial settings where errors damage parts or equipment. This mirrors how language agents separate a planning model from tool execution, applied to motors and grippers.
The Mercedes result matters because factory tasks are repetitive, high-volume and unforgiving, so success rates must be very high before automation pays off. A 96% rate is impressive for a general-purpose model but still means some failures that need human recovery on a production line. Competition in physical AI is intensifying. Runway released the open 27B Praxis-1 world action model on November 12, NVIDIA is pushing its robotics and autonomy stack, and Tesla, Figure and Physical Intelligence are all pursuing humanoid deployments. DeepMind's advantage is pairing frontier reasoning models with hardware partners like Apptronik.
Caveats: the 96% figure covers a specific parts-handling trial, not all 200 tasks, and the coverage does not detail the failure modes or how long the trial ran. Results from vendor-controlled tests often fall in less structured settings. Watch for whether Mercedes moves from trials to production use, whether Apptronik announces volume orders, and whether DeepMind makes the model available to other robotics partners or keeps it limited to selected customers.