Google launches Gemini Robotics 2, enabling robots to tie trash bags and change light bulbs

Google has unveiled a next-generation AI model for humanoid robots capable of controlling the entire body, signaling a new phase in the race to build capable robotic AI. The model enables precise hand movements, autonomous navigation and multi-robot collaboration.

Google DeepMind on Thursday introduced Gemini Robotics 2, a next-generation AI model featuring full-body control and precision manipulation.

Unlike its predecessor, Gemini Robotics 1.5, which focused on upper-body control, the new model integrates control across the robot's entire body — from head to toe.

For example, when given the instruction "put the object in the green box on the lower shelf," the robot can walk to the location on its own, pick up the item, bend down and place it precisely in the designated box.

Hand dexterity has also improved significantly. The model moves each of the five fingers independently to tie knots, seal zip-lock bags and replace light bulbs — delicate tasks that were previously difficult for robots to perform.

Google also unveiled two companion models alongside Gemini Robotics 2: Gemini Robotics-ER2, a reasoning model that serves as the robot's "brain," and Gemini Robotics On-Device 2, a lightweight model that runs directly on the robot without an internet connection.

ER2 can independently plan and execute complex, multi-step tasks lasting several minutes, recognize errors mid-task and correct them, and enable multiple robots to share information and collaborate on intricate workflows.

On-Device 2 is a lightweight model designed to handle new robot types within hours using fewer than 200 training data samples. It runs AI directly on the robot without a cloud connection.

Carolina Parada, vice president of robotics at DeepMind, called Gemini Robotics 2 "a major milestone in the journey to bring artificial general intelligence to the physical world," adding that "for decades, we have dreamed of robots that naturally come into our lives and help us."

The models are expected to compete against Nvidia's robot AI model Isaac GR00T and Tesla's Optimus.

Google will initially offer the models through a private preview on its enterprise agent platform, with the on-device model to be rolled out sequentially to key robot manufacturing partners.

Google also introduced a new safety benchmark for robotic AI models called Asimov-Agentic.

The benchmark measures whether a robot can refuse unsafe tool calls, predict whether a task is feasible before attempting it, and proactively request human intervention in uncertain situations.