Google DeepMind announced Gemini Robotics 2 on Thursday, a new artificial intelligence model that provides complete whole-body control for humanoid robots. Previously, DeepMind's robotics AI focused on upper-body movements for tabletop tasks. The Gemini Robotics 2 model now extends control from a robot's feet to its fingertips, allowing for a wider range of physical actions.
The AI model enables robots to reason through their movements, facilitating tasks like walking, crouching, stretching, and manipulating objects. In demonstrations, the Apptronik Apollo 2 robot, powered by Gemini Robotics 2, executed instructions such as retrieving a watering can and placing it into a specific shelf. This represents a significant step toward more adaptable and generally capable robots.
Gemini Robotics 2 is comprised of three distinct models. The primary 'Robotics 2' model converts vision and language input into motor control for action. A reasoning and planning model, 'Robotics ER 2', analyzes surroundings and sequences tasks, even recovering from errors and collaborating with other robots. The third model, 'Robotics On-Device 2', allows the AI to operate locally without an internet connection, adapting to new robotic bodies with minimal training data.
This new AI can also coordinate multiple robots to work together on complex workflows. For example, different robot types, such as a wheeled robot and a humanoid, can divide tasks through shared understanding. The system also incorporates safety features, including halting a robot if a person approaches and resuming only when the area is clear.
Beyond whole-body control, Gemini Robotics 2 enhances dexterity in robotic hands and grippers. It can operate a five-fingered hand to perform delicate actions like tying knots or sealing bags, and also manage standard grippers for complex tasks.
Apptronik's Apollo 2 robot serves as a platform for testing and developing this AI. Apollo 2 is a modular humanoid robot designed for real-world operations, featuring advanced actuation and dexterous object handling. Through partnerships like the one with Google DeepMind, data collected from Apollo 2 robots in various environments helps train and refine the Gemini Robotics models.
Google states that Gemini Robotics 2 moves toward "physical AI," aiming for robots that can intelligently interact with the world and perform tasks previously requiring human intervention. This development is seen as a stride towards robots that can operate more autonomously and adaptably in human environments.