Google DeepMind unveils Gemini Robotics 2 for whole-body robot control

Google DeepMind announced Gemini Robotics 2, an AI model that expands control of humanoid robots from the upper body to whole-body motions, including walking, crouching, and stretching. The update follows the previous model’s focus on upper-body control and is designed to enable a wider range of physical actions.

The new model also improves dexterity by supporting five-fingered hands, allowing robots to perform tasks such as sealing a Ziploc bag, tying a trash bag, or unscrewing a lightbulb. Videos shared by Google demonstrate these capabilities using Apptronik’s Apollo 2 robot, which can bend over to pick up objects and retrieve specific items from a shelf.

Google DeepMind simultaneously updated Gemini Robotics ER (embodied reasoning), a vision-language model that helps robots analyze surroundings, process instructions, and execute multi-step tasks. The updated ER 2 model enhances long-horizon task completion, recognizes task boundaries, and supports coordination among multiple robots of different types.

In one demonstration, Apollo 2 instructs Google’s dual-arm robot to place tools in a bin during a garage cleaning task. Google describes ER 2 as its "safest robotics model to date," citing improved detection of nearby humans, safety tool calls, and the ability to halt operations if a person approaches too closely.

The company also improved its Gemini Robotics On-Device Model, which can run locally on a robot without an internet connection. This version adapts more quickly to new robot embodiments, including those with significantly different shapes, sensors, and degrees of freedom.