Google Says Its New Robotics Model Is Ready to Move Beyond the Lab

Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action model designed to help robots learn physical tasks faster, operate more reliably, and handle messier real-world environments. The announcement signals that the AI race is expanding beyond chat interfaces and code generation into robots that can work with their hands.

The model was evaluated across more than 200 tasks, including handling delicate objects, using tools, serving food with tongs, tying bows, pipetting liquid, and manipulating flexible materials. According to Google, Gemini Robotics 2.5 improved success rates by 30% to 100% versus previous baselines while completing tasks 20% to 30% faster.

The more compelling proof point comes from a Mercedes-Benz pilot using Apptronik’s Apollo 2 humanoid robot. On active automotive production lines in Hungary and Germany, the robot handled “kitting”: sorting metal parts from bulk bins into sequenced trays for vehicle assembly. Google reports that the system reached a 96% success rate.

The goal is not to replace every worker with a humanoid robot overnight. The use case is much narrower and more practical: offload repetitive, physically taxing preparation tasks so technicians can focus on more complex assembly and problem-solving work.

Safety is also a major part of the release. Google says the model uses real-time reasoning guidance to steer around physical obstacles and nearby people. It reported a 33% improvement in safe task completion around obstacles and a 43% improvement around humans. DeepMind also introduced an open-source benchmark called ASIMOV-Jailbreak to test whether robotics systems can resist unsafe or disguised instructions.

The real challenge for physical AI is reliability. A chatbot can give a bad answer and be corrected. A robot handling factory equipment, hot materials, or expensive components needs to make far fewer mistakes. That is why the Mercedes-Benz pilot matters more than a flashy demo.