Google DeepMind has released Gemini Robotics 2, a system that can help control several kinds of robots, including humanoids. WIRED reports that the release combines a vision-language model with vision-language-action models that handle body movement and hands or grippers.
In demonstrations, robots performed tasks such as tidying shelves, screwing in lightbulbs, and tying trash bags. The system was trained with human teleoperation, video examples, and simulations, which is an important limitation: it is not a general robot that can do any physical task without task-specific training.
The release shows Google’s bet that AI will move beyond chat and coding into physical environments. Robotics also makes safety questions sharper because mistakes can affect people and property, not just files or messages.
Google says it uses layered guardrails and is introducing ASIMOV-Agentic, a benchmark for evaluating whether commands given to AI-controlled robots could lead to harmful or uncertain outcomes. The practical test will be whether these systems remain reliable outside staged demos.