This week: Google DeepMind released Gemini Robotics 2.5, and a humanoid robot running it spent time on real Mercedes-Benz production lines sorting car parts into trays.

Elsewhere: What the factory test did and didn’t show, a new way to check whether robots can be talked into bad ideas, and three fresh AI drops worth a skim.

The New Prime Signal

Most robot videos are filmed in a lab, with good lighting, a tidy table, and a researcher standing just out of frame. This week's is different. On Tuesday, Google DeepMind introduced Gemini Robotics 2.5, its newest vision-language-action model—the kind of AI that turns what a robot sees and what it's told into arm and hand movements. To show it off, DeepMind teamed up with Apptronik and Mercedes-Benz, putting Apptronik's Apollo 2 humanoid robots on active car production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The job wasn't glamorous—it was sorting metal brackets. That's sort of the point.

What shipped: DeepMind says Gemini Robotics 2.5 is its most capable robot model to date, and it tested the model on more than 200 distinct tasks, from serving food with tongs and pipetting a liquid to tying a bow on a gift box. Compared with previous baselines, DeepMind reports a 30% to 100% higher success rate depending on the type of task, while finishing tasks 20% to 30% faster. The model works with different robot bodies, including Apollo 2 with its five-finger SharpaWave hand (22 degrees of freedom) and the Franka Duo platform with simple two-finger grippers. In the Mercedes-Benz proof-of-concept, Apollo 2 did automotive kitting: pulling parts from bulk bins and placing them into trays in the exact order each vehicle on the line needs. DeepMind reports a 96% success rate. The model is rolling out to DeepMind's robotics partners, not to the public.

Why it matters: Kitting is repetitive, physically tiring prep work that keeps an assembly line fed, and it's exactly the kind of job companies would love to hand off so technicians can focus on harder build tasks. Factories have used robots for decades, but those are usually programmed for one narrow job. Gemini Robotics is trying to be a general-purpose brain that can learn many jobs, and a live production line is a far tougher test than a lab bench. Under the hood, two models split the work: Gemini Robotics ER keeps track of the overall workflow, while Gemini Robotics 2.5 controls the hands and arms, reaching into bins and correcting for gaps as it goes. DeepMind also says its Gemini 4 Argon model sped up development by generating thousands of simulated practice runs from a single real camera scene.

The fine print: 96% sounds great until you do the math. It means roughly one miss in every 25 tries, and as Ars Technica points out, a person doing the same repetitive task gets it right virtually every time. DeepMind director Kanishka Rao told Ars that closing the last 4% will be a "steep road." TechTimes notes that DeepMind didn't say how many attempts the 96% covers, how it defines a success, or how often a person stepped in, and Ars reports the factory video is sped up 2x. Every number here comes from DeepMind, the lab results are DeepMind's own tests, and this is a proof-of-concept, not humanoids rolling out across Mercedes-Benz plants.

A general-purpose robot doing real, boring work on a real line is genuine progress, and the honest "steep road" quote is more reassuring than any polished demo. But factories run on 99.9%, not 96%. For now, think of Apollo as a promising new hire on probation: useful, improving fast, and still needing someone to check the trays.

Picking parts is one challenge. Making sure a robot near people stays safe, even when someone tries to trick it, is another.

How it works: Gemini Robotics 2.5 uses what DeepMind calls inference-time reasoning guidance, meaning the robot thinks about physical hazards while it's acting and steers around them. Ars Technica describes it as the robot stopping or waiting until it's safe to continue. In DeepMind's evaluations, this raised safe task completion by 33% around obstacles and by 43% around nearby people. DeepMind is also releasing ASIMOV-Jailbreak, an open-source benchmark with hundreds of prompts that disguise unsafe requests across dozens of scenes—the robot version of someone sweet-talking a chatbot into misbehaving. DeepMind says Gemini 4 Argon, its reasoning model that leads its agentic robotics safety evaluations, resisted 99.5% of those attempts, up from 82% for its previous baseline, Gemini Robotics ER 2. That score belongs to Argon, not to the new robot-control model.

Try it: You can't download Gemini Robotics 2.5 yet, but you can watch it work. DeepMind's announcement has clips of the tong, pipette, and bow-tying tasks plus the Apollo 2 factory footage, along with details on ASIMOV-Jailbreak for anyone building or studying robots. For a skeptical read on the numbers, see TechTimes' breakdown of what the 96% leaves out, and for a quick summary from Europe, heise online covers the pilot. A good question to ask of any robot demo, this one included: out of how many tries?