FreshSynapse

AI trends & daily updates, in plain English

AI Research

Gemini Robotics 2.5 Puts Humanoids on Mercedes-Benz Lines — 96% Success, With the Fine Print Attached

By FreshSynapse Editorial · · 3 min read

Two bright yellow industrial robotic arms on metal worktables in a bright factory laboratory, with control panels, glass partitions and a purple banner behind them

Photo: Hippopx (public domain). Illustrative image.

What happened. Google DeepMind introduced Gemini Robotics 2.5 on Tuesday 10 November 2026, its most capable vision-language-action (VLA) model to date — a model that takes camera images and a plain-language instruction and outputs robot movements. The headline test was not in a lab: DeepMind put Apptronik’s Apollo 2 humanoid robots on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, where they picked metal automotive brackets from bulk bins and placed them into sequencing containers in the order each vehicle needs. DeepMind reports a 96% success rate.

What the model does. DeepMind tested Gemini Robotics 2.5 across more than 200 distinct tasks — serving food with tongs, tying a bow on a gift box, pipetting liquid — and says it succeeds 30% to 100% more often than its previous models, depending on the task type, and finishes tasks 20% to 30% faster. The gains are DeepMind’s own results, measured against baselines its announcement does not name, and there is no independent evaluation yet. The model runs on different robot bodies without a full rebuild: DeepMind names Apollo 2, fitted with the five-finger SharpaWave hand, and the Franka Duo, a two-armed platform with simple two-fingered grippers. In the Mercedes pilot, a second model — Gemini Robotics ER, an embodied reasoning model — tracked the workflow while Gemini Robotics 2.5 controlled the hands and arms. DeepMind credits much of the progress to Gemini 4 Argon, a frontier Gemini model whose robotics research abilities sped up development.

The 96% needs its fine print. A proof of concept is not a deployment, and the number as published cannot be judged precisely. DeepMind does not say how many trials the 96% covers, how it defines a success, or how often a human stepped in. Twenty-four successes in 25 attempts and 960 in 1,000 both round to 96%, but they mean very different things. No Mercedes-side statement confirms the figure, and neither company has said whether the humanoids will stay on the lines or what a failure looks like on a production line. The model still runs on cloud servers rather than on the robot itself, so some delay remains between seeing and acting. And the model is rolling out to DeepMind’s robotics partners — Agile Robots, Apptronik and Boston Dynamics — not to the public: no API, no price, no general availability date.

The safety side. DeepMind also released ASIMOV-Jailbreak, an open-source benchmark for testing whether robots can be talked into unsafe actions. On it, DeepMind reports that Gemini 4 Argon resists 99.5% of jailbreak attempts, against 82% for Gemini Robotics ER 2 — its own figures, but the benchmark itself is now public for others to run.

Why it matters. The robotics race has a new shape. Nvidia’s GR00T 2, unveiled the same week, teaches robots by watching human video; Gemini Robotics 2.5 is the rival bet, built on a frontier language model and tested on a real car line. A humanoid doing kitting — repetitive, tiring work — at 96% in a pilot is genuinely notable, because it is the first time the claim comes from an active production line rather than a demo floor. But the missing denominator is the story: until DeepMind publishes trial counts, intervention rates and sustained-run data, 96% is a promising pilot number, not a business case. The next proof worth watching is not another video — it is a second factory, a longer run, and numbers someone else has checked.

Sources

  • Google DeepMind
  • Gemini Robotics
  • robots
  • humanoid robots
  • Mercedes-Benz

Related articles