Google DeepMind’s push to move AI from screens into factories has reached a new milestone: Gemini Robotics 2.5, its latest vision-language-action model, has been tested on Mercedes-Benz production lines in Hungary and Germany.

DeepMind’s case is straightforward. The company says the system broadens the tasks robots can learn while improving speed, dexterity and safety. Across more than 200 evaluations—including using tongs, pipetting liquid and tying bows—it reported success rates 30% to 100% higher than earlier baselines, with tasks completed 20% to 30% faster.¹ The model is being rolled out to robotics partners rather than made generally available.

The factory trial is the headline result. Apptronik’s Apollo 2 humanoid sorted automotive parts from bulk bins into sequenced trays, a repetitive process known as kitting, and reached a 96% success rate. Apptronik CEO Jeff Cardenas called it “useful work alongside people in real industrial settings,” arguing the results provide a foundation for deploying humanoids at industrial and commercial scale.¹ DeepMind also promoted the launch as enabling robots to learn “hundreds of everyday skills,” from folding laundry to sorting car parts.Demis Hassabis (demishassabis)²

But the distinction between a strong demonstration and factory-ready dependability is where the debate sharpens. Independent coverage notes that human workers performing such repetitive tasks get them right virtually every time; DeepMind director Kanishka Rao described closing the remaining four-point gap as a “steep road.”³

DeepMind’s robotics chief Carolina Parada frames that tension as a trade-off. A robot can be optimized for one application, she said, or developers can pursue broad generalization and “maybe not get to 99.9 in every application.”³ That matters because the company’s ambition is not another tightly programmed factory machine, but a robot able to adapt to uncertain environments.

Safety is part of the same unresolved equation. DeepMind says its reasoning guidance raised safe completion by 33% around obstacles and 43% around people, while its new ASIMOV-Jailbreak benchmark probes whether robots refuse unsafe instructions.¹ Yet the current model remains cloud-based, meaning latency still separates a polished pilot from the seamless, near-instant response a busy factory floor requires.³