The American company Apptronik’s humanoid ‘Apollo 2’, equipped with Gemini Robotics 2, ties a knot in a trash bag using its fingers. Google DeepMind YouTube capture
Google has introduced a next-generation physical artificial intelligence (AI) model. The company has presented a new “brain” that controls an entire robot body, quickly adapts to different robot forms with only several hours of data, and orchestrates collaboration among multiple robots.
On the 30th (local time), Google unveiled its physical AI model product line called ‘Gemini Robotics 2’.
Gemini Robotics 2 is a vision-language-action (VLA) model that converts video and language commands into motor control signals for robots. It can be applied to full-body movements of humanoids, including robot arms and grippers. Previous Gemini Robotics models were mostly limited to controlling the upper body and arms in fixed locations.
The American company Apptronik’s humanoid ‘Apollo 2’, equipped with Gemini Robotics 2, closes a zipper bag. Google DeepMind YouTube capture
In a demonstration video, Apptronik’s humanoid ‘Apollo 2’ equipped with Gemini Robotics 2 picked up a watering can, walked to a shelf and placed it in a designated space, replaced a stand light bulb using its five fingers, tied a knot, and closed a zipper bag. Evaluations so far suggest it is still difficult to compare its task execution speed or accuracy directly with that of humans.
Gemini Robotics ER 2 acts as the brain that, after understanding a command, breaks a task down into multiple steps and executes appropriate movements. Because ER 2 continuously observes camera footage and assesses task progress, it does not restart the entire task from the beginning every time it fails.
Based on the physical AI Gemini Robotics 2, Apptronik’s humanoid Apollo 2 (right) and a robot arm (left) collaborate on a task. Provided by Google
The key achievement is the stepwise integration of a general-purpose AI, which handles long-term planning and situational judgment, with a behavior model that rapidly controls actuators. The higher-level AI demonstrates a unified robot operating system structure that can simultaneously control robots from different manufacturers.
Safety, however, has not been fully verified. According to Google’s explanation, it explicitly states that Gemini Robotics must not be used in safety-critical domains such as medicine or transportation, where malfunctions could lead to loss of life or property damage.
The American company Apptronik’s humanoid ‘Apollo 2’, equipped with Gemini Robotics 2, bends its waist to pick up a watering can from the floor. Provided by Google
Copyright ⓒ DongA Science. All rights reserved.