Google DeepMind Unveils Gemini Robotics 2 for Humanoids
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Gemini Robotics 2 combines a vision language model (VLM) with two vision language action (VLA) models into a single system that lets robots perceive surroundings, reason about tasks, and control both full-body movement and gripper/hand actions.
- Apptronik's Apollo 2 humanoid, fitted with hands from a company called Sharpa, autonomously tidied shelves in a Google DeepMind demo; the model was trained using human teleoperation, video examples, and simulations.
- Google DeepMind is launching ASIMOV-Agentic, a new benchmark that detects whether commands given to AI-driven robots could produce harmful or uncertain outcomes.
- Carolina Parada, head of robotics at Google DeepMind, called the release a milestone toward "physical AGI" and told WIRED the safety question is "even more pressing" when AI models act physically, citing prior research showing frontier-AI robot control can produce unexpected and dangerous behavior.
- CEO Demis Hassabis has previously said he wants to build an Android-like operating system for many different robot types, and Google already partnered with Boston Dynamics to supply AI brains for its legged robots.
- While Anthropic and OpenAI lead in chatbots and AI coding tools, WIRED notes Google has a stronger robotics research track record, positioning Gemini Robotics 2 as its bid to push AI out of the digital realm.
Why it matters: Google is staking out a physical-AI lead where Anthropic and OpenAI have weaker footholds, and the ASIMOV-Agentic benchmark shows the company is taking the unique risks of physical AI seriously. For partners like Apptronik, Boston Dynamics, and Sharpa, Hassabis has explicitly framed the model as the basis for an Android-style robot operating system spanning many hardware makers.
