Google's 'physical AGI' can screw in lightbulbs — so why is safety the bigger story
google deepmind released gemini robotics 2, a multi-model system enabling humanoids to perform complex physical tasks like tidying shelves autonomously. this signals a major push towards 'physical agi' where ai controls hardware in unstructured environments, introducing new safety risks beyond digital errors.
google deepmind launched gemini robotics 2, an upgrade that combines a vlm for reasoning with two vla models to translate commands into physical movement for robots. the system successfully demonstrated autonomous capabilities, such as an apptronik apollo 2 robot tidying shelves, utilizing training data from teleoperation and simulations. this release underscores google's strategy to lead in 'physical agi' by building an ai os for hardware, differentiating its robotics focus from the text-based focus of competitors like anthropic and openai. however, giving ai models physical agency raises new safety concerns regarding unpredictable behavior, leading google to introduce the asimov-agentic benchmark to test for harmful outcomes in robotic control systems.