Google's Robot Model Hits 96% Sorting Brackets on Mercedes Lines

On November 10, Google DeepMind said its Gemini Robotics 2.5 model ran Apollo 2 humanoids at 96% success sorting metal brackets on active Mercedes-Benz lines. The post gives no trial count, no duration and no price.

Google's Robot Model Hits 96% Sorting Brackets on Mercedes Lines

Google DeepMind put one number in front of everything else on November 10, and it sits in the Mercedes paragraph. Apptronik’s Apollo 2 humanoid, running the new Gemini Robotics 2.5, picked and packed automotive parts on active Mercedes-Benz production lines at a 96% success rate.

The task is kitting. The robot pulls metal brackets out of a bulk bin and sets them into containers in assembly sequence. DeepMind names two sites, Kecskemét in Hungary and Sindelfingen in Germany.

Then comes what the post leaves out. No trial count. No pilot length. No parts count, and no word on whether this is one workstation or several lines. A 96% with no denominator is a launch-day claim, not a measurement. Mercedes appears once, in the acknowledgements, with no statement of its own. The Decoder’s coverage matches DeepMind’s figures and adds nothing on any of those gaps.

What else is in the post

DeepMind says 2.5 was evaluated on more than 200 distinct tasks, with success rates 30% to 100% above its prior baselines depending on task type and completion 20% to 30% faster. All of it is Google’s own testing. Two models split the work: Gemini Robotics ER tracks the workflow, and 2.5 drives hands and arms. The hardware is Apollo 2 with the 22-degree-of-freedom SharpaWave five-finger hand, plus a Franka Duo bi-arm rig with two-finger parallel grippers.

On safety, Google reports that inference-time reasoning raised safe task completion by 33% near obstacles and 43% near people. Again, its own evaluation.

Availability is a single phrase: rolling out to robotics partners. No price, no API, no general-availability date.

Against the last number Google published

On July 30, Gemini Robotics 2 came with a chart that put the five-finger hand behind the plain gripper: bulb out 92%, bulb in 36%, and a two-finger gripper at 89.6% on precise insertion. The gripper rig’s diverse tool kitting scored 78.9%. Today’s 96% is a different task on a different platform, so the two cannot be subtracted. The direction is still hard to miss. Kitting-type work went from just under 80% to 96% in three and a half months, and this time it was the five-finger Apollo 2 doing it.

In August we reported that BMW, Hyundai and Mercedes had all handed their humanoids the same first job, kitting. Mercedes then had only a vague line about handling and sorting. Now the job has a score and two named plants. BMW’s Figure trial at Spartanburg targets the same workstation with a different robot maker, and Aptiv’s plan for 100 embodied robots across 46 stations is more of it.

The headcount math cuts the other way too. Suppose a station makes 1,000 picks a shift. That is our assumption, not pilot data. A 4% miss rate means 40 interventions a shift. The human does not leave the loop. The human changes from the person who picks to the person who handles the 40.

Compare what each side publishes. Beijing’s November 11 plan sets a reliability floor for its target makers by 2028: more than 2,000 hours between failures and at least 4 hours of continuous battery life. Google’s pilot gives neither a number of hours nor a failure count. One side has published a standard it wants met in two years, the other a score with no stated test conditions. A plant buyer needs both, a yardstick to hold vendors to and a result earned under conditions someone can check. This week the market received one half of each, and neither half is complete.

The job category

Line-side logistics is where the exposure sits: kitting operators, bin replenishment, parts sequencing. Carmakers have told reporters for months that they cannot keep people in these roles, and three of them picked the same roles as the first humanoid assignment. Google frames the benefit as freeing technicians for complex build tasks. The post does not mention headcount, and Mercedes has said nothing.

On timing, this is a pilot result for partners, not a purchase order. Three things decide how fast it moves: whether Google or Mercedes backfills the trial count and duration, whether a price and a date appear, and whether 96% repeats at a third carmaker on a second part.

Sources

Keep reading

Beijing's 2028 Robot Plan Sets 200,000 Units of Annual Capacity Robotics

Beijing's 2028 Robot Plan Sets 200,000 Units of Annual Capacity

On November 11, Beijing's economic and information bureau issued a 2026-2028 plan for embodied-AI robots: at least 5 world-class and 10 domestic-leading humanoid makers, 200,000 units a year of whole-machine capacity, and a 2,000-hour failure-free floor. It names no employment target.

#beijing#humanoid-robots#china-robotics