Google reveals Gemini Robotics 2.0, promising improved dexterity and safety (arstechnica.com)

Google has announced the launch of Gemini Robotics 2.0, enhancing its AI-powered robots to perform more complex tasks while adapting to dynamic environments. This advancement, centered around a trio of new sub-models—including the public release of the Gemini Robotics ER 2—marks a significant step toward achieving "physical AGI," where robots can execute a broad range of human-like activities. The improved capabilities allow for increased collaboration between robots and enable them to control humanoid forms with greater dexterity, including those equipped with intricate hands. A key technical feature of Gemini Robotics ER 2 is its integration with the Gemini Live API, allowing developers to leverage its advancements in embodied reasoning. The model acts as a vision language model (VLM), capable of processing live video feeds to track tasks in real-time. While it currently achieves around 60 percent accuracy in classifying video frame completeness, this is a notable improvement over previous iterations and positions Google's robots ahead of competitors in visual understanding. These developments signify a pivotal moment for AI in robotics, enhancing the potential for versatile and intelligent machines in various applications.