Google Deep Mind Unveils New Robot AI Models to Tackle Dexterity Challenges
Google DeepMind has unveiled a new artificial intelligence model for robots, saying it can help humanoid robots coordinate movements across their entire bodies. The release is the latest effort to extend Google's Gemini AI technology into robotics. The company aims to give robots reasoning and multi-step planning capabilities while helping them adapt to human environments. The new system, Gemini Robotics 2, enables robots to walk, squat and manipulate objects while using reasoning capabilities to complete tasks autonomously. Google said that, unlike earlier models focused mainly on controlling a robot's upper body, Gemini Robotics 2 can control the full body of a humanoid robot. In a prerecorded demonstration, DeepMind's model controlled Apptronik's humanoid robot Apollo as it moved through a room, picked up a watering can and placed it on a shelf while avoiding obstacles along the way. DeepMind also released two other robot AI models that can work together or operate independently. Gemini Robotics 2 converts camera images and natural-language instructions into commands for a robot's motors. Gemini Robotics ER 2 serves as the robot's reasoning system, planning multi-step tasks and coordinating multiple robots working toward the same goal. Google also demonstrated improvements in robot dexterity. Researchers said the system achieved a 92% success rate in tests of a task involving unscrewing a lightbulb. Its success rate remained relatively low on more complex operations, such as tying up a trash bag and sealing a Ziplock bag. Google introduced a new robotics safety-evaluation benchmark to test whether robots can recognize uncertain situations and refuse instructions that pose safety risks. Google said Gemini Robotics ER 2 is its safest robot model to date, particularly in following instructions and avoiding people. Gemini Robotics ER 2 will be available through Google's AI Studio developer platform and offered in private preview on its enterprise AI platform, Google said. The more specialized Gemini Robotics 2 and On-Device 2 models will initially be made available to early partners and more than 100 trusted testing organizations. The company also announced a waitlist for robot developers and said it is working with key partners including Apptronik, Agile Robots SE and Boston Dynamics. Carolina Parada, vice president of DeepMind's robotics business, said: “Our goal is to bring AI into the physical world and build a layer of intelligence that can be used by all robots.” Kanishka Rao, director of DeepMind's robotics business, acknowledged that achieving humanlike dexterity remains a long way off. Robots still move slowly and cautiously because they must stop during execution to reason through decisions that humans make intuitively, Rao said. Rao added that robots still learn far less efficiently than humans. People can often adjust their behavior after only one or two mistakes, while robots remain far from that level. The release builds on Google's launch of Gemini Robotics in 2025, a robotics version of its flagship Gemini AI model that can convert language and visual information into physical actions by robots. The product also marks a renewed push in Google's robotics strategy, which has continued for more than a decade. Alphabet acquired several robotics startups but later scaled back the business and shut down its Everyday Robots division in 2023. Google's rivals OpenAI and Nvidia are also expanding in robotics AI. OpenAI has been exploring a general-purpose robotics foundation model that combines vision, language and action, while Nvidia provides software platforms to help developers train AI robots.
The release is the latest effort to extend Google's Gemini AI technology into robotics. The company aims to give robots reasoning and multi-step planning capabilities while helping them adapt to human environments.
The new system, Gemini Robotics 2, enables robots to walk, squat and manipulate objects while using reasoning capabilities to complete tasks autonomously. Google said that, unlike earlier models focused mainly on controlling a robot's upper body, Gemini Robotics 2 can control the full body of a humanoid robot.
In a prerecorded demonstration, DeepMind's model controlled Apptronik's humanoid robot Apollo as it moved through a room, picked up a watering can and placed it on a shelf while avoiding obstacles along the way.
DeepMind also released two other robot AI models that can work together or operate independently. Gemini Robotics 2 converts camera images and natural-language instructions into commands for a robot's motors. Gemini Robotics ER 2 serves as the robot's reasoning system, planning multi-step tasks and coordinating multiple robots working toward the same goal.
Google also demonstrated improvements in robot dexterity. Researchers said the system achieved a 92% success rate in tests of a task involving unscrewing a lightbulb. Its success rate remained relatively low on more complex operations, such as tying up a trash bag and sealing a Ziplock bag.
Google introduced a new robotics safety-evaluation benchmark to test whether robots can recognize uncertain situations and refuse instructions that pose safety risks. Google said Gemini Robotics ER 2 is its safest robot model to date, particularly in following instructions and avoiding people.
Gemini Robotics ER 2 will be available through Google's AI Studio developer platform and offered in private preview on its enterprise AI platform, Google said. The more specialized Gemini Robotics 2 and On-Device 2 models will initially be made available to early partners and more than 100 trusted testing organizations.
The company also announced a waitlist for robot developers and said it is working with key partners including Apptronik, Agile Robots SE and Boston Dynamics.
Carolina Parada, vice president of DeepMind's robotics business, said: “Our goal is to bring AI into the physical world and build a layer of intelligence that can be used by all robots.”
Kanishka Rao, director of DeepMind's robotics business, acknowledged that achieving humanlike dexterity remains a long way off. Robots still move slowly and cautiously because they must stop during execution to reason through decisions that humans make intuitively, Rao said.
Rao added that robots still learn far less efficiently than humans. People can often adjust their behavior after only one or two mistakes, while robots remain far from that level.
The release builds on Google's launch of Gemini Robotics in 2025, a robotics version of its flagship Gemini AI model that can convert language and visual information into physical actions by robots.
The product also marks a renewed push in Google's robotics strategy, which has continued for more than a decade. Alphabet acquired several robotics startups but later scaled back the business and shut down its Everyday Robots division in 2023.
Google's rivals OpenAI and Nvidia are also expanding in robotics AI. OpenAI has been exploring a general-purpose robotics foundation model that combines vision, language and action, while Nvidia provides software platforms to help developers train AI robots.