Google DeepMind, a division of Alphabet, has launched a cutting-edge artificial intelligence model for robots that allows humanoid machines to synchronize movements throughout their entire structure. This breakthrough signifies a major advancement in the expansion of Gemini AI technology, equipping robots with the ability to reason, plan complex tasks, and adjust to human environments.
The recently introduced system, Gemini Robotics 2, enables robots to walk, crouch, and interact with objects while reasoning through multiple tasks. Unlike its predecessor that primarily managed the robot’s upper body, Gemini Robotics 2 can control the entire humanoid form, from head to toes.
“Our aim is to incorporate AI into the physical world and create an intelligence layer that is relevant to every robot,” stated Carolina Parada, vice president of robotics at Google DeepMind, during a press conference.
At the same time, Kanishka Rao, director of robotics at Google DeepMind, admitted that achieving genuine dexterity remains a distant objective; robot movements are often slow and deliberate because these machines need to pause for decision-making that humans do instinctively.
This announcement is a continuation of Google’s 2025 launch of Gemini Robotics, a variant of its flagship Gemini AI model that was able to convert language and visual input into robotic actions.
ADVERTISEMENT
CONTINUE READING BELOW
The previous release rejuvenated the company’s long-standing robotics ambitions. Following the acquisition of several robotics startups in the early 2010s, Alphabet later curtailed much of that work, including closing its Everyday Robots unit in 2023.
Rivals like OpenAI and Nvidia are also engaged in developing AI models and software for robotic purposes. OpenAI is investigating general-purpose robot foundation models that merge vision, language, and action, while Nvidia offers software solutions for training AI-driven robots.
In a recorded demonstration for reporters, DeepMind’s new AI model successfully controlled Apptronik Inc.’s Apollo humanoid robot as it navigated a room, picked up a watering can, and placed it on a lower shelf, all while avoiding obstacles.
DeepMind also presented two additional robotics AI models on Thursday, capable of collaboration or functioning independently. Gemini Robotics 2 converts camera input and natural-language commands into motor directives, while Gemini Robotics ER 2 acts as the robot’s reasoning system, adept at planning multi-step tasks and coordinating multiple robots to reach a unified goal.
ADVERTISEMENT:
CONTINUE READING BELOW
Despite these advancements, Rao noted that robots continue to learn significantly less efficiently than humans, who can adapt their behavior after just one or two mistakes.
Google also highlighted progress in robotic dexterity, claiming that its system can successfully unscrew a light bulb 92% of the time during tests. However, other complicated tasks, such as tying a trash bag and sealing a Ziplock bag, showed comparatively lower success rates.
The company introduced new benchmarks designed to evaluate whether robots can recognize uncertainty and refuse unsafe requests. Google claims that Gemini Robotics ER 2 is its safest robotics model thus far, particularly in terms of adherence to instructions and its ability to navigate around people.
Gemini Robotics ER 2 will be available through Google’s developer platform, AI Studio, and will be in private preview on Google’s enterprise AI platform. The more specialized Gemini Robotics 2 and On-Device 2 models are being provided to early-access partners and over 100 trusted testers. Additionally, Google announced a waitlist for its models intended for robotics developers and is collaborating with key partners, including Apptronik, Agile Robots SE, and Boston Dynamics.
© 2026 Bloomberg
