🤖Google DeepMind Gemini Robotics 2.5 Hits 96% Success in Mercedes Factories

Your warehouse robots just got a lot less clumsy

Google DeepMind just pushed Gemini Robotics 2.5, and the robots are finally doing real work. Gemini-powered Apollo 2 robots kitted automotive parts in several European factories, including Mercedes-Benz plants, hitting a 96% success rate. If you're building anything with physical AI, the architecture here is the story. DeepMind splits the stack into Gemini Embodied Reasoning, a vision-language model that reads live video and plans the task, and a vision-language-action model that turns that plan into actuator moves. The catch? The 2.5 release still runs on cloud servers, so latency is a real constraint for anything that needs fast reflexes. The numbers: 2.5 posts a 30 to 100 percent higher success rate than its predecessor across more than 200 activities, and it runs 20 to 30 percent faster. Training leaned on digital twins built with Argon, which generated thousands of practice variations from real camera data. Safety got attention too. When obstacles appeared, the robot stopped or waited, completing tasks 33 percent more often, and success climbed 43 percent when people were in the way. DeepMind also released ASIMOV-Jailbreak, an open source database of 1,500 prompts meant to trick a robot into unsafe actions. With Argon in the loop, the system hits 98.8% jailbreak safety, and Gemini Robotics ER 2 alone scores 73.5% on Google's safety benchmark.