Tying a bow sounds easy. You probably learned it as a kid. For a robot, though, a floppy ribbon is a nightmare: it bends, twists and never sits the same way twice.
So when Google DeepMind showed a humanoid robot tying a bow this week, people paid attention. The robot runs on Gemini Robotics 2.5, a new AI model that DeepMind announced on November 10. Yet the bow isn’t the most interesting part. The real story comes from a factory floor at Mercedes-Benz.
So what is Gemini Robotics 2.5?
Gemini Robotics 2.5 is a vision-language-action model, or VLA. In plain words, it looks at the world through cameras, understands an instruction, and then turns that into actual arm and hand movements. DeepMind calls it its most capable VLA model to date.
It doesn’t work alone, though. A sister model, Gemini Robotics ER, handles the planning side, which DeepMind calls embodied reasoning. Think of ER as the foreman who tracks the job, while 2.5 is the pair of hands doing it. According to Dataconomy, Google says the system is “approaching human-level performance on some real-world robotic tasks.”
DeepMind tested it on two kinds of robots. One is Apptronik’s Apollo 2 humanoid, fitted with a five-finger hand that has 22 degrees of freedom. Degrees of freedom are the separate ways a joint can move, so more means finer control. Then there’s the Franka Duo, a two-armed setup with simple two-finger grippers.
This builds on Gemini Robotics 2, which brought whole-body control to robots back in July. Right now, though, Gemini Robotics 2.5 is only rolling out to DeepMind’s robotics partners. There’s no public API and no price yet.
The Mercedes test is the real headline
Lab demos are nice, but factories are where robots earn their keep. So DeepMind, Apptronik and Mercedes-Benz ran a pilot at two Mercedes plants: Kecskemét in Hungary and Sindelfingen in Germany.
The job was automotive kitting. That means picking parts out of big bulk bins and sorting them into trays, in the exact order each car on the line needs them. It’s repetitive and fiddly, and it has to be right every time.
The result, according to DeepMind’s announcement, was a 96% success rate. ER tracked the workflow, while Gemini Robotics 2.5 controlled the hands and arms.
Apptronik’s co-founder and CEO, Jeff Cardenas, was upbeat. “Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training,” Cardenas said. Post-training means extra tuning for a specific robot and task, on top of the base model. Cardenas also called it “the foundation for bringing humanoid robots to industrial and commercial environments at scale.”
Honestly, 96% is impressive, but it’s not the finish line. It still means some picks fail, and somebody has to catch them. We saw a similar milestone when Xiaomi’s humanoid worked a full factory shift, and the pattern’s getting clear. Robots are moving from demo videos to real work.
Why bows and cardigans are a big deal
Here’s the thing about robots. Rigid objects like boxes and bolts keep their shape, so they’re fairly predictable. Soft, floppy things like ribbon or cloth don’t, and Dataconomy notes they’ve been a long-running headache for robots.
That’s why DeepMind’s before-and-after numbers jump off the page. Compared with Gemini Robotics 2, here’s how the new model did on a few tasks:
- Packing pastries with tongs: 5% to 100%
- Closing a toolbox lid: 0% to 93%
- Untying ribbon bows: 0% to 70%
- Hanging a cardigan: 0% to 62%
- Folding pants: 0% to 40%
Across more than 200 distinct tasks, DeepMind says success rates rose by 30% to 100% over previous baselines, depending on the task. The robots also finished jobs 20% to 30% faster. For example, tasks ranged from serving food with tongs to using a pipette, the little dropper tool you see in labs.
Part of the trick is practice. DeepMind used Gemini 4 Argon as a simulation engine, generating thousands of varied practice runs from a single real camera scene. That’s a different route from Korea’s humanoid data factory, where people teach robots by remote control.
A quick reality check, though. These are DeepMind’s own numbers, not independent tests. Still, going from zero to 62% on hanging a cardigan makes home robots feel a little less like science fiction.
Can you talk a robot into something dangerous?
Chatbots get “jailbroken” a lot. That’s when someone tricks an AI into ignoring its safety rules with a cleverly worded request. Now picture the same trick on a robot holding a kitchen knife. Not fun.
So DeepMind released ASIMOV-Jailbreak, a new open-source benchmark for exactly this problem. It packs hundreds of prompts that disguise unsafe requests across dozens of scenes. On jailbreak resistance, Gemini 4 Argon scored 99.5%, up from 82% for Gemini Robotics ER 2.
The robots also got better at staying safe in messy spaces. DeepMind says safe task completion improved by 33% around obstacles and by 43% when people were nearby. That second number matters most, because a factory robot that works next to humans has to be safe around them first.
We also like that DeepMind made the benchmark open-source. Other robot makers can now test their own systems against the same tricky prompts. Shared tests are a good first step toward shared safety standards.
So should you care about Gemini Robotics 2.5?
If you’re hoping to buy a bow-tying robot, not yet. Gemini Robotics 2.5 is partners-only for now, and the most impressive figures come from DeepMind itself. We’d love to see an independent team run the same tests.
But the direction is hard to ignore. The robot “brain” is turning into software that many different robot bodies can share, a bit like an operating system. That’s the same bet investors made when they valued Physical Intelligence at $20 billion. If that approach wins, the software layer becomes the prize.
For a country like Armenia, that’s actually good news. You don’t need to build humanoids from scratch to take part. Software, simulation and robot training are areas where smaller tech ecosystems can compete, and that’s where this race seems to be heating up.
So here’s the question we’ll leave you with. When a robot can hit 96% on a real car line, how long before that last 4% stops being a reason to wait?