Google DeepMind has released Gemini Robotics 2, an AI system designed to control different robots, including humanoids performing dextrous physical tasks. The development matters because it moves Google’s Gemini models beyond chat and software into real-world environments, advancing its ambition for broadly capable “physical AGI”.

The system combines a vision-language model for understanding scenes and instructions with two models that translate this understanding into full-body and hand movements. Demonstrations included an Apptronik Apollo 2 robot tidying shelves, but the robots still require task-specific training using teleoperation, videos and simulations. Google says it applies safety guardrails at multiple layers and has introduced the ASIMOV-Agentic benchmark to identify harmful or uncertain robot actions.