The latest version of the Gemini Robotics artificial intelligence model, developed by Google DeepMind, is now capable of completely controlling the bodies of humanoid robots. While the previous edition focused only on the upper part, the new Gemini Robotics 2 now allows integral movements, from feet to hands, as recently announced by the company. This evolution represents a significant advance, bringing robots closer to human action and allowing a greater range of interactions with the environment.
This improvement enables humanoid robots to perform a wide range of movements, such as walking, crouching, stretching and manipulating objects with precision. In videos released by Google, it is possible to observe Apptronik’s Apollo 2 robot bending down to pick up a watering can and selecting and removing specific items from a shelf, demonstrating the versatility of the new system.
Although Google DeepMind recognizes that the movement speed of its robots still needs to be improved, the update is considered an essential step towards developing skills needed in real-world tasks that require complex whole-body coordination.
Learn more: Google DeepMind innovation: new AI model controls entire robot body
Additionally, Gemini Robotics 2 has been enhanced to provide superior dexterity, now controlling more sophisticated hands equipped with five fingers. This capability allows robots to perform delicate activities such as zipping plastic bags, tying garbage bags or unscrewing a light bulb with greater precision.
Google DeepMind also updated Gemini Robotics ER (embodied reasoning), a vision and language system that helps robots interpret the environment, process instructions, and complete multi-step tasks. The Gemini Robotics ER 2 version demonstrates greater efficiency in carrying out prolonged activities, being able to clearly identify the beginning and end of each action.
On the same topic: Artificial intelligence drives new era of autonomous robots in factories and homes, but challenges persist
The company highlights that this update also allows multiple robots of different configurations to collaborate in carrying out tasks. A video illustrates Apollo 2 directing Google’s two-armed robot to store tools in a bin while organizing a garage.
The Gemini Robotics ER 2 is named by Google DeepMind as the company’s “safest robot model ever developed”, given its improved ability to identify the presence of people nearby, activate safety devices and safely stop the robot’s operation if someone gets too close.
At the same time, Google DeepMind improved the Gemini Robotics On-Device model, which now operates directly on the robot’s hardware, without the need for an internet connection. This feature optimizes adaptation to new configurations, including those with “significantly different shapes, sensors and degrees of freedom.”