Google DeepMind just moved robots out of the lab and onto the factory floor. On Tuesday, November 10, the company introduced Gemini Robotics 2.5, its most capable vision-language-action model to date — and instead of a demo video, it brought receipts: Apptronik's Apollo 2 humanoid, running the model, sorted car parts on live Mercedes-Benz production lines with a 96% success rate.
What Gemini Robotics 2.5 Actually Is
A vision-language-action (VLA) model connects what a robot sees and what you tell it to the movements it makes. Gemini Robotics 2.5 is DeepMind's newest version of that idea, built to give robots broader skills and faster learning. It works across different robot bodies — from Apptronik's Apollo 2 humanoid, with its five-finger, 22-degree-of-freedom SharpaWave hand, to the Franka Duo platform with two-fingered grippers. DeepMind says Gemini 4 Argon, its frontier model, acted as a simulation engine during development, generating thousands of practice runs from real camera scenes to speed up training.
The Mercedes Test: 96% on a Live Line
The headline result came from a real factory, not a controlled demo. In a proof of concept on active Mercedes-Benz lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apollo 2 autonomously did automotive kitting — sorting individual parts from bulk bins into organized trays in the exact order each vehicle on the line needs. DeepMind reports a 96% success rate. Across more than 200 tasks tested, from serving food with tongs to tying a bow on a gift box and pipetting liquid, the company says Gemini Robotics 2.5 achieved 30% to 100% higher success rates than previous baselines while completing tasks 20% to 30% faster. All of those figures are DeepMind's own — no independent lab has verified them yet.
Safety Got a Serious Upgrade Too
As robots move into messy, everyday spaces, DeepMind says it improved physical safety steering: the model uses inference-time reasoning to steer around hazards in real time, boosting safe task completion around obstacles by 33% and around nearby people by 43%, per the company's evaluations. It also open-sourced ASIMOV-Jailbreak, a benchmark of hundreds of prompts designed to trick robots into unsafe actions. Gemini 4 Argon resisted 99.5% of those attempts, against 82% for the previous baseline, Gemini Robotics ER 2.
Pricing and Availability
Here's the catch: you can't use it yet. Gemini Robotics 2.5 is rolling out only to DeepMind's robotics partners — Apptronik, Agile Robots, Boston Dynamics and more than 100 trusted testers. There is no public API, no price, and no general-availability date. An early-access waitlist is open on DeepMind's model page.
The Verdict: The Lab Demo Era Is Ending
This is the second big robotics moment in a week —
NVIDIA's GR00T 2 just showed robots learning new tasks by watching humans — and it lands in the same week Anthropic dropped
Claude Fable 5.2 and Mythos 5.2, and as the
race to put an AI agent on every screen reaches the wrist. The pattern is clear: the race is no longer about who has the cleverest demo, but who can survive a real shift on a real factory floor. Gemini Robotics 2.5's 96% is a proof of concept, not a product, and the last 4% is where factory managers live. But a humanoid doing boring, physically taxing prep work while technicians focus on complex builds is exactly the pitch the industry has been waiting for. If DeepMind's numbers hold up outside its own tests, the factory floor just became AI's next workplace.
Comments
Post a Comment