Google DeepMind Unveils AI Model Capable of Full Robot Body Control

Google DeepMind has unveiled its latest iteration of the Gemini Robotics AI model, which now boasts the ability to “control entire humanoid robots.” Unlike its predecessor, which was limited to managing the upper body, Gemini Robotics 2 enables full-body movements, allowing robots to engage in actions from walking to manipulating objects with precision.

This upgrade significantly expands the capabilities of humanoid robots, enabling them to walk, crouch, stretch, and interact with their environment. Demonstrations featuring Apptronik’s Apollo 2 robot showcase its ability to bend down and pick up a watering can, as well as retrieve specific items from a shelf.

While Google DeepMind acknowledges that there’s still room for improvement in movement speed, they emphasize that this update marks a crucial advancement toward equipping robots with the skills necessary for more complex, real-world tasks that require coordinated whole-body movements.

Gemini Robotics 2 also enhances dexterity, allowing for the control of intricate five-fingered hands. This means robots can now tackle tasks such as sealing a Ziploc bag, tying up a trash bag, or unscrewing a lightbulb.

In addition, Google DeepMind is enhancing its Gemini Robotics ER (embodied reasoning) model, which aids robots in analyzing their surroundings, processing instructions, and executing multi-step tasks. The new Gemini Robotics ER 2 is designed to perform tasks over longer durations and is now capable of recognizing when tasks start and finish.

This update also facilitates collaboration among different types of robots, as demonstrated in a video where Apollo 2 directs Google’s dual-arm robot to organize tools in a bin while cleaning a garage.

Furthermore, the company claims that Gemini Robotics ER 2 is its “safest robotics model to date,” with improved capabilities for detecting nearby humans, triggering safety protocols, and halting operations if someone approaches too closely.

Lastly, Google DeepMind has made strides with its Gemini Robotics On-Device Model, which can operate locally on a robot without needing an internet connection. This model is now quicker to adapt to new embodiments, even those with significantly different shapes, sensors, and degrees of freedom.