Google has unveiled an artificial intelligence (AI) model to power next-generation humanoid robots.
Google said on the 31st it unveiled an AI model, "Gemini Robotics 2," that strengthens full-body control and precision manipulation for robots.
Unlike its predecessor, "Gemini Robotics 1.5," which could control only the upper body, Gemini Robotics 2 can integrate control of the entire body, including fingers and toes. It also performs delicate tasks such as tying knots or sealing zip bags by moving all five fingers individually.
Google also unveiled the reasoning model "Gemini Robotics-ER2," which serves as the robot's brain, and "Gemini Robotics On-Device 2," which runs directly on robot hardware without network consolidation.
ER2 can plan and execute multi-step tasks that continue for several minutes. It also has a function to identify the cause of failure and autonomously correct its actions if problems arise during a task. In addition, it enables collaboration by exchanging information among multiple robots and jointly performing complex processes.
On-Device 2 is a lightweight model that can operate new types of robot hardware within hours with fewer than 200 training data samples.
Karolina Parda, vice president (VP) overseeing robots at DeepMind, said, "For decades, we have dreamed of robots that naturally enter our lives and help," and introduced Gemini Robotics 2 as "a key milestone on the journey to realizing artificial general intelligence (AGI) in the physical world."
Google will first offer these models as a private preview on its agent platform for corporations. The on-device model will be provided sequentially to major robot manufacturing partners.
They are expected to compete in the market with Nvidia's robot-driving model "Isaac GROOT" and Tesla's "Optimus."
Google also unveiled a new performance metric (benchmark) to measure the safety of robot AI models, "Asimov-Agentic."
Google said the metric can measure whether the model can refuse unsafe tool calls, predict whether a task is feasible before attempting an action, and request human intervention in advance in uncertain situations.