Google DeepMind has unveiled Gemini Robotics 2, an advanced AI platform enabling "intelligent whole-body control" for robots. A demonstration video showcased robots performing tasks like cleaning trash, picking up watering cans, inserting tapes into boomboxes, screwing in lightbulbs, and tying garbage bags.

This new iteration builds upon the original Gemini Robotics system, which focused on robotic arms and hands for tasks like folding origami. Gemini Robotics 2 integrates multiple AI models, including a vision language model and two vision language action models, to achieve a multimodal understanding of their environment and control full-body and hand movements. Google emphasizes that the video features "real-time footage" of "fully autonomous" robots, contrasting with demonstrations of other systems that have been shown to use teleoperation.

While these robots are not yet general-purpose, the model was trained for the specific video tasks using a combination of human teleoperation, video examples, and simulations. Google DeepMind views this as a step towards "physical AGI." Safety is a major consideration, with a multi-layered approach including guardrails and a new benchmark, ASIMOV-Agentic, to detect potentially harmful commands. Despite these advancements, Google is not planning immediate consumer releases, unlike rivals like Elon Musk, who has ambitious timelines for Tesla's Optimus robots.