Gemini Robotics 2.5 accelerates real-world general dexterity

Google DeepMind has introduced Gemini Robotics 2.5, which it calls its most capable vision-language-action model so far, and it's rolling out to robotics partners. The headline claims: 30% to 100% higher success rates than prior baselines depending on the task, 20% to 30% faster completion, and evaluation across 200 distinct tasks from serving food with tongs to tying a bow on a gift box to pipetting liquid.

The number that will travel is the 96% success rate in a proof-of-concept with Apptronik's Apollo 2 humanoid doing automotive kitting on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. That's a factory, not a demo table, and that's where robotics claims finally get tested.

We're genuinely encouraged by this one. Kitting, sorting parts from bulk bins into trays in the exact order each vehicle needs, is repetitive, physically taxing work. Handing it to a robot so technicians can focus on complex build tasks is the kind of automation we'd argue for: dull, hard on bodies, and well suited to a machine that doesn't get tired.

But let's be precise about what 96% means. It's a pilot, and on a production line the remaining 4% isn't a rounding error. It's stoppages, human correction and cost. The real question for Mercedes-Benz isn't whether the robot can do the task; it's whether the failure modes are cheap enough to live with at scale. DeepMind calls this promising progress, which is the right register.

The part I'm most focused on is Gemini 4 Argon. DeepMind says the frontier model helped accelerate this work by acting as a simulation engine, generating thousands of varied practice runs from a real camera scene, building digital twins, and helping researchers hypothesize why failures happen from video. That's AI speeding up AI research, now spilling into the physical world. It's a big deal for the pace of robotics, and it compounds: better models produce better training data for robots, which produce more real-world data.

Safety gets real attention here, and it should. DeepMind reports inference-time guidance that improved safe task completion around obstacles by 33% and around nearby people by 43%. It's releasing ASIMOV-Jailbreak, an open benchmark of disguised unsafe requests, and says Argon reached 99.5% jailbreak resistance versus 82% for its earlier baseline. An open benchmark is exactly the kind of thing that lets outsiders check the work.

The downside if this scales is the obvious one. Robots that can be talked into things are a different category of risk than chatbots that can be talked into things, and a 0.5% gap matters when the system has arms. And the labor shift won't stop at kitting. We'd like to see the partners say as much about how displaced work gets redeployed as they do about success rates.

Gemini Robotics models work across different robot bodies, including Apptronik's Apollo 2 with a five-finger, 22 degree-of-freedom hand and the Franka Duo platform with parallel grippers. In the factory pilot, the Gemini Robotics ER embodied reasoning model tracked the workflow while the 2.5 actions model controlled the hands and arms. Apptronik CEO Jeff Cardenas said the models improved dramatically within weeks of post-training and carried over to production with little drop-off.