Gemini Robotics 2 Brings Google's AI Into the Physical World

Google DeepMind released a new AI model called Gemini Robotics 2 that can control different types of robots, including humanoids capable of performing complex physical tasks like screwing in lightbulbs and tying trash bags. The system combines multiple AI models, including vision language models that understand images and video, and vision language action models that control a robot's movements and hand operations. Google has trained the model using human teleoperation, video examples, and simulations. However, giving frontier AI models access to robots introduces safety risks, as previous research has shown such systems can produce unexpected and sometimes dangerous behavior. Google is addressing these concerns with a multi-layered safety approach and introducing a new benchmark called ASIMOV-Agentic to measure the safety of AI systems controlling robots.