DeepMind's Gemini Robotics 2.5 hits 96% in a Mercedes-Benz kitting pilot

Google DeepMind introduced Gemini Robotics 2.5 on 10 November 2026, calling it its most capable vision-language-action (VLA) model, the kind that turns camera input and spoken or written instructions into arm and hand movements. DeepMind says it is rolling out to its robotics partners; the post's acknowledgements name Agile, Apptronik and Mercedes-Benz.

The factory pilot

DeepMind, Apptronik and Mercedes-Benz ran Apollo 2 humanoid robots on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The job was kitting: picking metal automotive brackets from bulk bins and placing them into sequencing trays in the order each vehicle needs. Two models split the work. Gemini Robotics ER, DeepMind's embodied reasoning model, tracked the workflow; Gemini Robotics 2.5 drove the hands and arms. DeepMind reports a 96% success rate and calls the work a proof-of-concept.

Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off.

Jeff Cardenas, co-founder and CEO, Apptronik, in DeepMind's post

What DeepMind says improved

• Breadth: evaluated on 200 distinct tasks, including serving food with tongs, tying a bow and pipetting liquid.
• Success and speed: a 30% to 100% higher success rate than DeepMind's previous baselines depending on task type, and tasks completed 20% to 30% faster.
• Hardware: works on Apollo 2 with the five-finger, 22-degree-of-freedom SharpaWave hand and on the Franka Duo two-gripper platform.
• Physical safety: inference-time reasoning guidance raised safe task completion by 33% around obstacles and 43% around nearby people, in DeepMind's evaluations.

A jailbreak test for robots

DeepMind also released ASIMOV-Jailbreak, an open benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes. The Hugging Face dataset google/asimov_agentic is licensed CC BY 4.0, requires accepting access terms on Hugging Face before download, and ships evaluation scripts for physical safety, human safety monitoring and gauge reading. On it, DeepMind says Gemini 4 Argon resists 99.5% of jailbreaks against 82% for Gemini Robotics ER 2. That figure is for Argon, the general model, not for Gemini Robotics 2.5.

DeepMind credits Gemini 4 Argon with speeding up development: it generated thousands of simulated practice runs and digital twins from real camera scenes, and proposed causes of failures from video.

The catch

The 96% has no denominator. DeepMind does not say how many picks the robots made, over how many shifts, how fast they worked against a human kitter, what happened in the 4% of failures, or whether people stood by to recover them. In a sequencing task, one wrong bracket in a tray can stop a line, so the failure mode matters as much as the rate. There is no price, no general availability date and no independent evaluation; every benchmark gain is measured against DeepMind's own earlier models.

For an operations team, this is a signal about timing, not a purchase decision. The useful pieces today are the task definition, which is a template for scoping a pilot, and the jailbreak benchmark, which any team putting an agent near physical equipment can run now.