Alphabet Inc’s Google DeepMind unveiled a new artificial intelligence (AI) model for robots that allows humanoid machines to coordinate movements across their entire bodies, marking its latest effort to extend its Gemini AI technology to allow robots to reason, plan multi-step tasks and adapt to human environments.
The new system, Gemini Robotics 2, enables robots to walk, crouch and manipulate objects while reasoning through tasks. Unlike the previous model, which primarily controlled a robot’s upper body, the company said Gemini Robotics 2 can direct an entire humanoid from top to bottom.
"Our goal is to bring AI into the physical world and then build the intelligence layer that can be used by every robot,” said Carolina Parada, vice president of robotics at Google DeepMind, in a press briefing with reporters.
At the same time, Kanishka Rao, director of robotics at Google DeepMind, said that true dexterity remains a distant goal; robot movements remain slow and deliberate because the machines must pause to think through decisions that humans make intuitively.
The announcement builds on Google’s 2025 debut of Gemini Robotics, a version of its flagship Gemini AI model that was capable of translating language and visual information into robotic actions.
That release revived robotics ambitions that have stretched back more than a decade for the company. After acquiring a string of robotics startups in the early 2010s, Alphabet later wound down much of that work, including by shuttering its Everyday Robots unit in 2023.
Google rivals OpenAI and Nvidia are also developing AI models and software for robots. OpenAI has also explored general-purpose robot foundation models that combine vision, language and action, while Nvidia provides software that helps developers train AI-powered robots.
In a pre-taped demonstration for reporters on July 28, DeepMind’s new AI model controlled Apptronik Inc’s Apollo humanoid robot as it walked across a room, picked up a watering can and placed it on a lower shelf, all while navigating around obstacles.
DeepMind also released two other robotics AI models on July 30, which can work together or independently. Gemini Robotics 2 converts camera input and natural-language instructions into motor commands. Gemini Robotics ER 2, the robot’s reasoning system, plans multi-step tasks and can coordinate multiple robots working toward the same goal.
Even so, Rao, the DeepMind director, said robots still learn far less efficiently than humans, who can adjust their behaviour after just one or two mistakes.
Google also highlighted improvements in robotic dexterity, with its researchers saying the system can successfully unscrew a light bulb 92% of the time in tests. Other complex tasks, including tying a trash bag and sealing a Ziplock bag, had relatively low success rates.
The company also introduced new benchmarks to evaluate whether robots can recognise uncertainty and refuse unsafe requests. Google said Gemini Robotics ER 2 is its safest robotics model yet, compared to its prior models, particularly in its ability to follow instructions and navigate around humans.
Gemini Robotics ER 2 will be available through Google’s developer platform, AI Studio, and in private preview on Google’s enterprise AI platform, while the more specialised Gemini Robotics 2 and On-Device 2 models are being offered to early-access partners and more than 100 trusted testers, Google said. The company added that it is opening up a waitlist for its models to robotics developers, and is working with a range of core partners including Apptronik, Agile Robots SE and Boston Dynamics. – Bloomberg
