Google DeepMind Unveils Gemini Robotics 2: Controlling Humanoid Robots from Head to Toe
Google DeepMind has released Gemini Robotics 2, a groundbreaking family of AI models capable of controlling humanoid robots from their feet to fingertips. These models can coordinate multiple robots simultaneously and adapt to new robot bodies in just a few hours, marking a significant step towards "physical AI."
A Leap Towards Physical AI
Gemini Robotics 2 represents a clear indication of the industry’s push towards physical AI. Wired describes it as a major milestone in achieving Artificial General Intelligence (AGI). The concept is simple: an AI brain that can control not just a robot’s upper body, but its entire frame, enabling tasks like walking across rooms and tidying up.
Controlling Every Part of the Humanoid
Unlike DeepMind’s earlier models that primarily moved an upper body to perform tabletop tasks, Gemini Robotics 2 drives the entire humanoid robot. It allows the robot to walk, crouch, stretch, and balance while manipulating objects in human-sized spaces.
In a demonstration, Apptronik’s Apollo 2 robot received the instruction to place a watering can in a green bin on a shelf. The robot successfully navigated to the table, picked up the can, moved to the shelves, bent down, and executed the task. This feat showcases the complexity of coordinating legs, torso, arms, and hands with a single prompt.
Improved Dexterity and Precision
The AI model has also enhanced dexterity, enabling the robot to perform tasks like tying knots and sealing ziplock bags using its five-fingered, 22-joint hand. Additionally, it can operate simpler two-fingered grippers on other platforms.
Three Models in One System
The release comprises three distinct models:
- Gemini Robotics 2: This model translates visual and auditory inputs into motor commands, essentially the robot’s physical doer.
- Gemini Robotics ER 2 (Reasoning Layer): Acting as a high-level brain, this layer plans multi-step jobs, tracks its progress in real-time, and can even steer other robots like Boston Dynamics’ Spot to fetch items. It also facilitates collaboration among different robots.
- On-Device 2: Capable of running locally without an internet connection, this model can be installed on a new robot body with minimal examples (fewer than 200) and training time.
Acknowledging Limitations
DeepMind was transparent about the current limitations of their system. While impressive, the AI’s success rate in tasks like unscrewing a light bulb is only around 92%, highlighting areas for improvement.