Skip to content
MaungaAI news, explained

Robotics

A robot running Google's AI prepares parts in Mercedes factories, still in testing

Google DeepMind has unveiled Gemini Robotics 2.5, an AI that makes robots more dexterous. A humanoid has tested it in Mercedes-Benz factories, but for now only partner companies are getting it.

2 min read2 sources, one officialTranslated with AI from the Spanish original

Share: WhatsApp · Email

AI-generated illustration for Maunga. Not a photograph.

The essentials

  • Google DeepMind has unveiled Gemini Robotics 2.5, the AI that controls its robots' movements, and is starting to offer it to partner companies (Google DeepMind Blog, Ars Technica).
  • Claimed by one partyAccording to Google DeepMind, an Apollo 2 humanoid prepared parts in Mercedes-Benz factories in Hungary and Germany with a 96% success rate (Google DeepMind Blog).
  • Claimed by one partyThe company says it gets things right between 30% and 100% more often than its earlier versions, depending on the task, and that it is faster (Google DeepMind Blog).
  • Google DeepMind is openly releasing ASIMOV-Jailbreak, a test to measure whether a robot can be tricked into doing something unsafe (Google DeepMind Blog, Ars Technica).

Why it matters

Robots tend to fail outside the lab, where nothing is in its proper place. If the figures hold up, this suggests humanoids are getting closer to doing repetitive jobs in real factories, such as preparing parts. But this was a pilot, the figures come from Google DeepMind, and no one outside the company has checked them yet.

The details

According to Google DeepMind, the system has two parts. One plans: it looks through the camera and decides on the steps. The other, Gemini Robotics 2.5, moves the arms and hands (Google DeepMind Blog, Ars Technica). The company tested it on more than 200 tasks, such as serving food with tongs, using a pipette, or tying a bow on a gift box (Google DeepMind Blog). Ars Technica highlights the bow, because ribbons and other objects that bend are among the hardest things for a robot to handle (Ars Technica).

In the factories, Apptronik's Apollo 2 robot picked metal brackets out of bulk bins. It then placed them on trays, in the order each car on the line needs them (Google DeepMind Blog). Google DeepMind says this version learns to find its way in real time to avoid obstacles and people. It is also openly releasing a test to check whether someone can trick the robot into doing something dangerous (Google DeepMind Blog).

What we don't know

  • Whether Mercedes-Benz will use these robots regularly or whether it was only a trial.
  • When the technology will reach companies beyond Google's partners, and at what price.
  • Whether anyone independent has checked the success rates the company reports.

Sources

This story is based on 2 sources. In the text, each fact names in brackets who reports it.

Was this story useful? Yes Not really

Spanish original

The thread: Google DeepMind · Gemini Robotics 2.5

  1. Signal passes an AI-run math exam, but only in its key parts
  2. A robot running Google's AI prepares parts in Mercedes factories, still in testing · you are here
  3. Any Google account will be able to make three AI videos a month for free
  4. Google keeps its most powerful AI for cybersecurity experts
  5. Google's virtual worlds now have sound and change on command, but almost no one can get in
  6. OpenAI and Anthropic ask for laws that would bind them; Google and Meta don't
  7. Google's top model is now available, but only if you pay