Google has unveiled Gemini Robotics 2.0, a suite of AI models intended to make robots more capable of handling complex, changing real-world tasks. The company says the system improves dexterity, whole-body control, real-time reasoning and cooperation between robots, advancing its longer-term aim of creating general-purpose “physical AGI”.

The central model, Gemini Robotics ER 2, interprets instructions and live video from a robot’s cameras, enabling it to monitor progress and correct individual failed steps without restarting a task. Google says it identifies key moments in video with almost 90% accuracy and assesses task completion frames with nearly 60% accuracy, while the action-generating Gemini Robotics 2 and offline On-Device 2 versions remain available only to a limited set of testers.