Gemini Robotics 2.5 hits 96% on a Mercedes-Benz kitting line pilot

Google DeepMind has introduced Gemini Robotics 2.5, which it calls its most capable vision-language-action model to date. In a Mercedes-Benz proof of concept, a humanoid running it reached a 96% success rate in automotive kitting.

A 96% kitting result on a Mercedes-Benz line

According to DeepMind, the model was tested on more than 200 distinct tasks involving tools, deformable objects and delicate handling. Examples include serving food with tongs, tying a bow on a gift box and pipetting liquid.

The headline industrial result comes from a Mercedes-Benz production line, where Apptronik’s Apollo 2 humanoid performed kitting, meaning the collecting of parts into sets. In that pilot, Gemini Robotics ER, the company’s embodied reasoning model, tracked the workflow while Gemini Robotics 2.5 controlled the hands and arms. DeepMind reports the 96% figure as coming from a single proof of concept, not a long production run.

Against its predecessor, Gemini Robotics 2, DeepMind reports success rates 30% to 100% higher depending on the task, and tasks finished 20% to 30% faster. Its own charts show packing pastries with tongs rising from 5% to 100%, untying a ribbon bow from 0% to 70%, hanging a cardigan from 0% to 62% and folding pants from 0% to 40%. These are the company’s numbers; the page does not point to independent replication.

The model works across several robot types, according to the page, including Apollo 2 with its five-finger, 22-degree-of-freedom SharpaWave hand and the Franka Duo platform with two-fingered parallel grippers.

Simulation and safety guidance do part of the work

DeepMind says Gemini 4 Argon sped up development by acting as a simulation engine. It can produce thousands of simulated practice runs from a real camera scene, build digital twins and help analyze failures from video.

On safety, the model uses guidance at inference time to avoid hazards. DeepMind says this raised safe task completion by 33% around obstacles and by 43% around nearby people. It also released ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes. On it, Gemini Robotics ER 2 scored 82% and Gemini 4 Argon scores 99.5%, the company says. Those scores belong to the planning and reasoning models, not to the new action model itself.

Partners first, with no public access yet

The model is rolling out to DeepMind’s robotics partners. The page names Agile, Apptronik and Mercedes-Benz. It mentions no public API, pricing or general availability, so developers outside the partner group cannot use it yet. The page also lists no explicit limitations and gives no failure rates for the harder tasks beyond the improvement figures.

Closed partner access against Nvidia’s open approach

The release lands as rivals push in a different direction on openness. Nvidia’s GR00T 2 runs on-board and ships with open weights and 100K hours of data, as our coverage of Nvidia’s 397 MB browser-run humanoid walking policy also shows for the company’s wider push to make robot control small and accessible. Gemini Robotics 2.5, by contrast, is offered only through selected partners, and DeepMind has not said whether weights will be released.

Scale of training data is the other point of comparison. Our explainer on vision-language-action models notes that the open OpenVLA model was trained on about 970,000 real-world robot demonstrations and beat the 55-billion-parameter RT-2-X by 16.5 percentage points on 29 tasks, according to its authors. DeepMind does not disclose the size of the dataset behind Gemini Robotics 2.5, which makes a like-for-like comparison impossible. The same explainer cautions that labs use different robots, tasks and success criteria, so cross-lab results are hard to compare.

What Google can claim is breadth. The task list spans food, fabric, liquids and car parts, and the model runs on both a humanoid hand and a two-gripper platform. Rivals have shown laundry folding and box assembly, but a named automaker pilot with a stated success rate is a stronger commercial signal than a demo video, even if it is one pilot.

What remains unclear is cost, speed under real factory conditions and how often the model fails when it does. Until DeepMind or its partners publish those figures, the 96% number is best read as a promising result from the company’s own test.