WangDou logo
WangDou

Google's Gemini Robotics 2.5 sorts car parts on Mercedes lines with a 96 percent success rate

2026-11-11·WangDou AI Express·Google / Robotics / Embodied AI

The robot finally stopped dropping the cup. It can now pick the cup off the tray, set it in the right spot, and get it right about nine and a half times out of ten.

Three key facts

Google DeepMind released Gemini Robotics 2.5 on November 10, calling it its most capable vision-language-action model to date. It was evaluated across 200 distinct tasks, including serving food with tongs, tying a bow on a gift box and pipetting a liquid, all of which require combining tool use, manipulation of deformable objects and delicate grasping. Depending on task type, success rates rose 30 to 100 percent over previous baselines, and tasks were completed 20 to 30 percent faster.

The hardest evidence comes from real production lines. DeepMind worked with Apptronik and Mercedes-Benz to run Apollo 2 humanoids on active assembly lines in Kecskemét, Hungary and Sindelfingen, Germany, doing automotive kitting: sorting individual parts out of bins and assembling them per order. On that repetitive, physically taxing job the robot reached a 96 percent success rate autonomously.

The model works natively across robot types, including Apollo 2 with its five-fingered, 22-degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered parallel grippers. On safety, DeepMind improved physical safety steerability and introduced a benchmark for adversarial jailbreaks specific to robotics. The model is rolling out to its robotics partners.

Wang Dou's Take

Put 96 percent on a factory floor and it becomes a different number.

A line does not care about success rate. It cares about what happens in the other 4 percent. One mis-sorted part means rework at best, a stalled line at worst. But flip it around: human workers on repetitive sorting jobs miss things too. The difference is that a person notices, corrects, and raises a hand for the shift lead. The robot's 96 percent is its autonomous success rate. Who catches the remaining 4 percent is the line between a demo and a deployment.

Look at the tasks it picked: tying bows, serving with tongs, pipetting. Those are the party tricks. Mercedes does not need a bow tied on a gift box. It needs the right part grabbed out of several hundred. Putting both kinds of task on one scorecard looks great and actually says something simpler: generalization is good enough, and the rest is per-site tuning.

The real signal is the partner list. Apptronik brings the body, Mercedes brings the floor, Google brings the brain. Each takes one piece and none claims to do it all. That is more convincing than any pitch about humanoids entering homes next year.

Source: Google DeepMind Blog

Comments

Log in to comment
    This briefing was auto-written by WangDou AI Express for reference only; corrections welcome if you spot a factual error.
    指挥舱👽