Humanoid robots are now working on live Mercedes-Benz production lines.

Google DeepMind has announced Gemini Robotics 2.5, which it calls its most capable vision-language-action (VLA) model to date. In a proof-of-concept pilot on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 humanoid picked and packed automotive parts — sorting individual metal brackets from bulk bins into sequencing trays in the exact order needed for each vehicle — with a 96% success rate, DeepMind reports.

The model was evaluated across more than 200 distinct tasks, from serving food with tongs to pipetting liquid. DeepMind reports success rates 30% to 100% higher than previous baselines depending on the task type, with tasks completed 20% to 30% faster.

Two models work together: Gemini Robotics ER (embodied reasoning) tracks the workflow, while Gemini Robotics 2.5 controls the hands and arms. Apollo 2 uses the five-finger SharpaWave hand, with 22 degrees of freedom.

Safety is a headline feature too. DeepMind reports safe task completion up 33% around obstacles and 43% around nearby people. It has also released ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts designed to test whether robotics models can be tricked into unsafe actions. Its Gemini 4 Argon model scored 99.5% jailbreak resistance, against 82% for the previous baseline.

This is a pilot, not a permanent deployment, and DeepMind has announced no public API, price or general availability. The model is rolling out to robotics partners including Agile, Apptronik, Boston Dynamics and Mercedes-Benz.