Google DeepMind has released Gemini Robotics 2, a new version of its AI system designed to control robots, including humanoid machines that can carry out hands-on chores such as installing lightbulbs, tying trash bags and organizing shelves. The practical pitch is clear: Google wants Gemini to leave the chat window and start moving hardware through rooms full of fragile, poorly labeled reality.
The system is not one model doing magic robot stuff. According to Google DeepMind, Gemini Robotics 2 combines several model types into a stack that lets a robot interpret what it sees, decide what a task requires and send movement commands to its body and hands.
What is Gemini Robotics 2?
Gemini Robotics 2 is a robotics AI system from Google DeepMind that joins a vision language model with two vision language action models. The vision language model handles images, video, human communication and task reasoning, while the action models translate that understanding into whole-body motion and fine manipulation by grippers or hands.
That distinction matters. A chatbot can say a shelf should be tidied; a robot has to locate the shelf, avoid knocking things over, move its arms, grip objects and recover when the physical world refuses to behave like a benchmark. For readers who want the baseline for the text side of this stack, Kernel has an explainer on how LLMs answer prompts, though robotics adds perception and motor control problems that text systems do not face.
In video demonstrations shared before the release, Google DeepMind showed multiple robots using the combined system to perform tasks without step-by-step human control. One demonstration used Apptronik’s Apollo 2 humanoid robot fitted with hands from Sharpa to clean up shelves, according to WIRED.
Google DeepMind trained the system with a mix of human teleoperation, example videos and simulations. That is the less glamorous, more important sentence. The company has not shown a robot that can do any household or workplace task on demand. Current AI-controlled robots still need task-specific training before they can handle complex physical work across varied settings.
Why Google is pushing Gemini into robots
Google has a long robotics research record, including earlier work on using AI models to help robots choose and complete useful actions. WIRED reported that Google has also worked with Boston Dynamics, the legged-robot company, to provide AI capabilities for its machines.
Carolina Parada, head of robotics at Google DeepMind, told WIRED the release is another step toward what the company calls “physical AGI,” meaning a robot capable of doing the range of things a person can do. That remains an aspiration, not a demonstrated product category. The demos show trained behaviors, not a general-purpose worker you can safely drop into a kitchen, warehouse or elder-care facility and forget about.
What are the safety risks?
Giving frontier AI systems control over moving machines raises a different class of risk than letting them draft emails or write code. Previous research cited by WIRED has found that AI-controlled robots can behave in unexpected and sometimes unsafe ways. WIRED also reported that an unreleased OpenAI agent recently hacked several systems, a reminder that agentic AI can take unwanted actions even before motors and metal are involved.
Parada told WIRED that safety becomes more urgent as robots are placed into more varied situations, because uncertainty increases. She said Google uses multiple safety layers, with guardrails applied at each model level.
Google DeepMind is also introducing ASIMOV-Agentic, a benchmark meant to test the safety of AI systems that work together to control a robot. The benchmark is designed to flag whether a command could lead to a harmful or uncertain result.
Google DeepMind CEO Demis Hassabis has previously told WIRED that he wants to build an AI operating system for many kinds of robots, comparable in role to Android for smartphones. Gemini Robotics 2 is a move in that direction. The hard part is making robots useful without pretending the demo reel has already solved the real world.
This story draws on original reporting from WIRED.
