# DeepMind's Gemini Robotics 2.5 Hit 96% on a Mercedes-Benz Kitting Pilot. The Last 4% Is the Hard Part

Times of AI Desk · 2026-11-11 · Research

[https://timesof.ai/2026/11/deepmind-gemini-robotics-2-5-mercedes-apollo-96-percent-kitting](https://timesof.ai/2026/11/deepmind-gemini-robotics-2-5-mercedes-apollo-96-percent-kitting)

> Google DeepMind introduced Gemini Robotics 2.5 on November 10, a vision-language-action model it says achieves 30% to 100% higher task success than previous baselines and runs 20% to 30% faster, and that let Apptronik's Apollo 2 humanoid sort automotive brackets on Mercedes-Benz production lines with a 96% success rate in a proof of concept. It also released ASIMOV-Jailbreak, an open benchmark for robot jailbreaks. The model is rolling out to partners only, and all figures are DeepMind's.

DeepMind's new robotics model has been tested where general-purpose robots usually fall short: an active car factory. Its own figure, a 96% success rate, is high enough to make the pilot newsworthy and low enough to show why factories still prefer single-purpose machines.

## What DeepMind released

**Gemini Robotics 2.5**, introduced on **November 10**, is DeepMind's latest vision-language-action model, the layer that turns plans into arm and hand movements. DeepMind says it evaluated the model on more than **200 distinct tasks**, from serving food with tongs to tying a bow and pipetting a liquid, and that it achieves a **30% to 100% higher success rate** than previous baselines depending on the task, while completing tasks **20% to 30% faster**. The model is **rolling out to DeepMind's robotics partners**; Ars Technica reported these include Agile Robots, Apptronik and Boston Dynamics, and that it is not generally available.

It extends the [Gemini Robotics 2 line DeepMind launched in July](https://timesof.ai/2026/07/google-gemini-robotics-2), which also ran on Apptronik's Apollo humanoid.

## The factory pilot

With Apptronik and Mercedes-Benz, DeepMind ran Apollo 2 humanoids on active production lines in **Kecskemét, Hungary** and **Sindelfingen, Germany**. The robots performed automotive kitting, picking metal brackets from bulk bins and placing them into sequencing trays in the order each vehicle needs, and reached a **96% success rate**, according to DeepMind.

Ars Technica put that number in context: a person doing the same repetitive job gets it right virtually every time, and DeepMind director Kanishka Rao told it that closing the last 4% will be a "steep road." DeepMind robotics head Carolina Parada described the trade-off between specialising for one application and pushing generalisation "and maybe not get to 99.9 in every application." Ars also reported that the 2.5 model still runs on cloud servers rather than on the robot.

## Safety claims and a new benchmark

DeepMind says inference-time reasoning guidance raised safe task completion by **33%** around obstacles and **43%** around nearby people. It released **ASIMOV-Jailbreak**, an open-source benchmark of prompts that disguise unsafe requests for robots, and says Gemini 4 Argon reaches **99.5%** jailbreak resistance on it, against **82%** for its previous baseline, Gemini Robotics ER 2. Ars reported different figures from DeepMind for the same comparison, 98.8% with Argon in the loop against 73.5% for ER 2 alone, and described the benchmark as about 1,500 prompts.

DeepMind also says Gemini 4 Argon acted as a simulation engine during development, generating thousands of simulated practice runs from real camera scenes. The release came a day after Nvidia's own [GR00T 2 robot model claimed 71% success from one demonstration](https://timesof.ai/2026/11/nvidia-gr00t-2-world-action-model-71-percent-one-demo-corl); the two use different tasks and setups and are not directly comparable.

## Limits

- Every performance and safety figure is DeepMind's, from its own evaluations; we have not seen an independent test of the model.
- DeepMind's post and Ars Technica's report give different jailbreak-resistance figures for the same comparison; neither explains the gap.
- The factory work was a proof of concept. Neither DeepMind nor Mercedes-Benz has announced a production deployment.

## Sources

- [Google DeepMind: Gemini Robotics 2.5 accelerates real-world general dexterity (November 10, 2026)](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/)
- [Ars Technica: DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak" (November 10, 2026)](https://arstechnica.com/google/2026/11/deepmind-reveals-gemini-robotics-2-5-and-open-safety-benchmark-asimov-jailbreak/)