Google DeepMind announced Gemini Robotics 2.5 at CoRL in Austin this week, describing it as its most capable vision-language-action model. Architecturally it is two layers: a Gemini embodied-reasoning planner that interprets the scene and sequences steps, and a low-level policy that turns those steps into stable motion, with better finger-level dexterity than earlier versions.
The headline number: across more than 200 physical tasks, it reached 96% success on automotive parts-handling tests using Apptronik's Apollo 2 humanoid at a Mercedes-Benz facility, running 20% to 30% faster than prior baselines.
What 96% does and does not prove
Manufacturing has always wanted repetition, which is why the line is full of fixed-purpose arms welding, gluing and moving at a fixed beat. Humanoids never got onto the line not because they were weak, but because they could not cope with variation: a part two centimeters off position, a changed package color, a mixed lot. A fixed arm stops. A vision-language-action model re-plans.
A 96% task success rate says the understand-then-plan approach works inside a controlled cell. The value it creates is a collapse in changeover cost: a new line does not need new programs, just a few demonstrations.
The limits are equally concrete. 96% is not unattended operation. Production asks for shift-level metrics: success rate over eight continuous hours, mean time between failures, human takeover frequency, cost per good unit. Miss any of those and 96% is a launch-event number.
Three competing approaches
Google's route is a general foundation model plus embodied reasoning. Anthropic's is more aggressive: its new Fable 5.2 was not trained specifically for robots, yet pushed success on the RoboDojo dual-arm benchmark from 27.6% in the previous version to 64.9%, and demonstrated drawing, opening a bottle and solving a Rubik's cube.
NVIDIA's GR00T 2 takes the data route, pretraining on more than 400,000 hours of first-person human video, about 10,000 hours of gripper data and 23,500 hours of teleoperation. Adding human video lifted average completion across 72 real-robot tasks from 50% to 70%.
The common thread: none of them write a program per task. They perceive, attempt, and consolidate skills.
The Chinese production line is moving in parallel
Xiaomi's CyberOne 02 completed an 8-hour-5-minute continuous shift trial at an automotive assembly workshop, including two autonomous recharges and coordination with wheeled robots. UniX AI reports shipping more than 100 units per month with repeatable deployments in hotels, security and retail. Joyson says it has delivered over 2,000 high-compute robot controllers.
Individually these are modest. Together they mark a shift in the competitive question from "can it demo" to "can it ship."
A four-step evaluation frame
Pilot a station with high variation and contained consequences, such as incoming-goods sorting, not the main assembly line.
Quantify the labor you are replacing, including hiring difficulty, turnover and training time. If the station already cannot be staffed, the economics clear faster.
Demand shift-level data, not videos: continuous runtime, takeover counts, recovery time.
Write interfaces and exit terms into the contract. Who owns the data, whether the model can run on-premises, whether history exports if you change vendors.
Gemini Robotics 2.5 Reaches 96% Success in a Mercedes-Benz Plant
← Previous: Elvis and Eminem Rights Holders Sue Suno for More Than $202M
Next: →Alibaba's Qwen4-Max Preview: 1M Context, 393K Output, No Spec Sheet →