Google DeepMind has introduced Gemini Robotics 2.5, its most capable vision-language-action (VLA) model for robots, and proven it outside the lab: Apptronik Apollo 2 humanoids running the model picked and packed automotive parts on active Mercedes-Benz production lines with a 96% success rate. Why it matters: this is the clearest sign yet that general-purpose robot models are moving from demos to real factory work.
November 10, 2026 — DeepMind is rolling Gemini Robotics 2.5 out to its robotics partners. No public API or price has been announced, so for now it reaches the world through partner robots rather than a developer console.
What Gemini Robotics 2.5 does better
A VLA model turns what a robot sees and what it is told into movements of its arms and hands. DeepMind says version 2.5 widens both the range of skills a robot can learn and the speed at which it picks them up. The lab evaluated it across more than 200 distinct tasks — serving food with tongs, tying a bow on a gift box, pipetting a liquid — and reports success rates 30% to 100% higher than its previous baselines depending on the task type, with tasks completed 20% to 30% faster. Those are DeepMind’s own results; there is no independent evaluation yet.
The model is not tied to one robot body. DeepMind names Apptronik’s Apollo 2 humanoid, fitted with the five-finger SharpaWave hand (22 degrees of freedom), and the Franka Duo, a two-armed platform with simple two-fingered grippers. It follows Gemini Robotics 2, which brought whole-body control to humanoids in July.
The Mercedes-Benz factory test
The headline proof is a proof-of-concept on live production lines at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 performed automotive kitting: pulling metal brackets from bulk bins and laying them into sequencing containers in the exact order each car on the line needs. DeepMind’s embodied-reasoning model, Gemini Robotics ER, tracked the workflow while the Gemini Robotics 2.5 actions model controlled the robot’s hands and arms, correcting for gaps as it worked. The pilot reached a 96% success rate. DeepMind calls it a proof of concept, and neither partner has said whether humanoids will stay on the lines permanently.
Safety: a benchmark for tricking robots
DeepMind also shipped safety work alongside the model. Inference-time reasoning guidance steers the robot around physical hazards in real time, which the lab says boosted safe task completion by 33% around obstacles and 43% around nearby people. And it released ASIMOV-Jailbreak, an open-source benchmark on Hugging Face whose prompts disguise unsafe requests across dozens of scenes. Gemini 4 Argon — the frontier model DeepMind credits with accelerating 2.5’s development by generating simulated practice runs — resisted 99.5% of the jailbreak attempts, against 82% for the earlier Gemini Robotics ER 2 baseline.
Why it matters
Robotics is where the frontier-model race meets physical proof. A language model can bluff its way through a benchmark; a humanoid either picks the bracket or it does not. DeepMind’s 96% figure is a company-reported proof of concept, not a production contract — but doing it on active Mercedes-Benz lines, in two countries, is a stronger claim than another lab video. With NVIDIA’s GR00T 2 also courting robot builders this month, the fight to be the default “brain” inside other companies’ robots is now a real market, and Google just put a factory receipt on the table.
Frequently asked questions
What is Gemini Robotics 2.5?
It is Google DeepMind’s newest vision-language-action model — the software that turns camera input and instructions into robot arm and hand movements. DeepMind calls it its most capable VLA model to date.
What did the model achieve at Mercedes-Benz?
In a proof-of-concept, Apptronik Apollo 2 humanoids running Gemini Robotics 2.5 sorted and packed automotive parts on active production lines in Hungary and Germany with a 96% success rate, as reported by DeepMind.
Can developers use Gemini Robotics 2.5 today?
Not directly. DeepMind is rolling it out to robotics partners and has not announced a public API or pricing.
Sources: Google DeepMind (announcement, November 10, 2026), Unite.AI, Ars Technica via coverage of the launch.