Google unveils Gemini Robotics 2.0, promising better dexterity and safety

Google DeepMind has released Gemini Robotics 2.0, a suite of AI models designed to enhance the capabilities of physical robots. The new system aims to create generalist robots that can perform a wide range of tasks, moving beyond the narrow, pre-programmed actions like running or backflipping. The 2.0 release can control entire humanoid robots, including complex hands, with greater dexterity. A key upgrade is Gemini Robotics ER 2, an embodied reasoning model that DeepMind says is a significant advance over the previous 1.6 version. It is integrated with the Gemini Live API and is available to developers starting today. ER 2 is a vision language model that can process live video from the robot's cameras, allowing it to track progress during tasks. Google reports that ER 2 can classify video frame completeness with almost 60 percent accuracy, a substantial improvement over both the 1.6 release and competing models. Videos of robots performing various actions have been common for years, but these were typically narrowly programmed; the goal of Gemini Robotics is to achieve a form of 'physical AGI' where a robot can understand and execute human-like instructions. The Gemini 2.0 suite includes three new sub-models.