Subscribe
0%
Illustration for: DeepMind's Gemini Robotics 2.5 ran a humanoid on live Mercedes lines at 96% success. What the numbers do and don't show

DeepMind's Gemini Robotics 2.5 ran a humanoid on live Mercedes lines at 96% success. What the numbers do and don't show

Google DeepMind's new vision-language-action model drove Apptronik's Apollo 2 humanoid through parts kitting at two Mercedes-Benz plants and posts big jumps over its predecessor on 200 lab tasks. It's going to partners only, and almost every figure is DeepMind's own.

Listen to this article0:00 / 4:06

Google DeepMind on November 10 released Gemini Robotics 2.5, which it calls its most capable vision-language-action (VLA) model yet, and said it had already run the model on working car production lines. In a pilot with Apptronik and Mercedes-Benz, an Apollo 2 humanoid controlled by the model completed parts kitting with a 96% success rate on active lines in Kecskemét, Hungary, and Sindelfingen, Germany, according to DeepMind.

That matters because a general-purpose model doing a real logistics job on a live line, with a named carmaker, is a tougher test than a lab demo, and it's the kind of evidence factory buyers will ask for. But the details DeepMind left out matter as much as the headline number.

What the robot actually did

Kitting means pulling individual parts from bulk bins and placing them into trays in the order each car on the line needs them. In this pilot the parts included metal automotive brackets, DeepMind says. The setup split the work between two models:

  • Gemini Robotics ER tracked the overall workflow.
  • Gemini Robotics 2.5 controlled the robot's arms and hands.

The hardware was Apptronik's Apollo 2 humanoid with a five-finger, 22-degree-of-freedom SharpaWave hand. Apptronik CEO Jeff Cardenas said performance improved significantly within a few weeks of post-training and carried over to Mercedes production with "minimal performance loss," as quoted by The Decoder.

The lab numbers

DeepMind says it tested the model on 200 distinct tasks and reports success rates 30% to 100% higher than its previous baselines, depending on task type, with tasks finished 20% to 30% faster. Selected results from DeepMind's charts:

Task Previous baseline Gemini Robotics 2.5
Packing pastries with tongs 5% 100%
Closing a toolbox lid 0% 93%
Unhanging a mug to the table 29% 75%
Untying ribbon bows 0% 70%
Hanging a cardigan 0% 62%
Sweeping trash into a dustpan 0% 52%
Folding pants 0% 40%
Packing a rice cooker 0% 40%

The model works with different robot bodies, including Apollo 2 and the Franka Duo two-arm platform with parallel grippers, DeepMind says. It also used its Gemini 4 Argon model as a simulation engine: from one real camera scene, Argon builds a digital twin, generates thousands of varied practice runs, and helps diagnose why tasks fail.

Safety claims and a new open benchmark

DeepMind says reasoning-time guidance raised safe task completion by 33% around obstacles and 43% around nearby people in its evaluations. It also released ASIMOV-Jailbreak on Hugging Face, an open benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes. On it, Gemini 4 Argon resisted jailbreaks 99.5% of the time versus 82% for the previous baseline, Gemini Robotics ER 2. Note that DeepMind attributes that score to Argon, not to Gemini Robotics 2.5 itself.

What the numbers don't show

  • No denominator for the 96%. Neither DeepMind nor The Decoder gives the number of trials, parts or shifts, or how a "success" was defined. DeepMind calls it a proof-of-concept pilot.
  • Self-comparisons only. Every gain is measured against DeepMind's own older models, not rival robot models from other labs.
  • Self-graded safety. The safety evaluations and the jailbreak benchmark were designed by DeepMind, The Decoder notes, so they aren't independent, though the benchmark being open means others can now test against it.
  • No speed versus humans. The "20% to 30% faster" figure compares against the old model, not against line workers or conventional automation.

Who can use it

Nobody outside Google's partners, for now. DeepMind says the model is rolling out to robotics partners, naming Apptronik, Mercedes-Benz and Agile. There's no public API, pricing or general-availability date.

What to watch: whether Mercedes moves the pilot into regular production, and whether outside labs publish ASIMOV-Jailbreak scores for their own robot models. Those would be the first numbers in this story that DeepMind didn't produce itself.

Sources

  1. Google DeepMind — Gemini Robotics 2.5: Accelerating physical intelligence deepmind.google
  2. The Decoder — Google DeepMind's Gemini Robotics 2.5 scores 96 percent sorting metal parts in a Mercedes-Benz assembly pilot the-decoder.com

#Google DeepMind#Gemini Robotics#Gemini Robotics 2.5#Apptronik#Mercedes-Benz#humanoid robots#robotics#AI safety