Gemini Robotics 2: A Revolution in Robot Intelligence

A Major Advancement in Robotic Intelligence

For decades, the idea of robots capable of seamlessly integrating into our daily lives has fueled the collective imagination. Today, this vision is becoming a reality thanks to a significant advancement in the field of artificial intelligence applied to robotics. Most current robots are limited by rigid programming, often designed to perform repetitive and specific tasks. Their ability to learn autonomously or adapt to changing environments has remained limited. Furthermore, transferring skills from one robot to another has proven complex. To address current challenges, it is crucial to equip robots with AI models capable of thinking and acting intelligently while safely interacting with their environment.

Gemini Robotics 2: A New Era for Robotics

Gemini Robotics 2 represents a major advancement in adaptable robotics. This intelligence system allows robots to control their entire bodies, enhancing their dexterity and enabling effective collaboration among them. This innovation paves the way for a new generation of robots capable of adapting to various tasks, such as walking, squatting, or manipulating objects to clean a room. Thanks to this advanced intelligence, robots can now operate autonomously on the device, quickly adapting to new robotic bodies in just a few hours. This flexibility is made possible by three high-performance models:

• Gemini Robotics 2: This vision-language-action (VLA) model converts visual and linguistic inputs into movements, allowing precise motor control of humanoid and bi-articulated robots. It also offers skilled manipulation with hands and grippers, thereby increasing the robots' ability to interact with their environment in a more sophisticated manner.
• Gemini Robotics ER 2: This embodied reasoning (ER) model enables robots to understand and plan complex tasks while facilitating communication with humans. It also introduces the capability for robots to work as a team, which is essential for accomplishing tasks that require coordination and collaboration among multiple units.
• Gemini Robotics On-Device 2: Optimized for local operation, this model quickly adapts to new robotic bodies, even with radically different configurations. It is designed to operate efficiently without requiring a network connection, which is crucial for applications where latency and connectivity are major concerns.

Humanoids in Motion: Comprehensive Task Management

The world we live in is designed for human movements, requiring capabilities for reaching, bending, and balancing. Previous models focused on controlling the upper body of humanoid robots. Now, with Gemini Robotics 2, AI extends to the entire body, allowing for complete coordination. Take the example of the Apollo 2 humanoid robot from Apptronik. Thanks to Gemini Robotics 2, it can receive complex instructions such as "place the watering can in the green box on the bottom shelf." Apollo is capable of processing this instruction, moving to the table, grasping the watering can, and placing it precisely in the designated spot. Although the speed of movement can still be improved, this capability marks a significant advancement toward accomplishing complex tasks that require full-body coordination.

Towards Advanced Dexterity

To be truly useful in domestic and professional environments, robots must possess great finesse in their movements. Gemini Robotics 2 achieves a new level of physical dexterity, whether with hands or grippers. The model can control the SharpaWave five-finger hand, equipped with 22 degrees of freedom, to perform delicate actions such as tying knots or sealing a ziplock bag. It can also manipulate standard two-finger parallel grippers on a Franka Duo platform to carry out complex tasks. The goal is to achieve precision and speed comparable to that of humans.

Agentic Reasoning and Robot Collaboration

Most real-world tasks require multiple steps and planning over an extended period. The embodied reasoning model, Gemini Robotics ER 2, acts as the robot's brain, processing instructions, observing the environment, and coordinating necessary actions. This update enables robots to manage longer task sequences, involving hundreds of decisions. Gemini Robotics ER 2 now understands when tasks start and end and can identify key events, thereby improving the understanding of progress. Additionally, collaboration among robots is introduced, allowing different types of robots to work together to solve complex workflows that a single robot could not accomplish alone.

Rapid Adaptation and Local Operation

In many applications, robots must operate without relying on a network connection. Gemini Robotics On-Device 2 is designed to meet these constraints, enabling efficient local operation. This model can quickly adapt to new robotic bodies, even with different shapes and sensors, in just a few hours. This is made possible through advanced motion transfer techniques inherited from Gemini Robotics 1.5. This ability to adapt rapidly is essential for environments where flexibility and quick adaptation are crucial.

Commitment to Safe and Responsible Robotics

Safety is a fundamental aspect of robotics research. As robots gain capabilities, it is essential to ensure their safety and alignment with safety standards. Gemini Robotics 2 integrates ASIMOV-Agentic, a new benchmark for agentic safety. It assesses the robots' ability to refuse dangerous actions and request human intervention in cases of uncertainty. With enhanced embodied reasoning, Gemini Robotics ER 2 is the safest model to date, capable of detecting human proximity and triggering appropriate safety measures. This multi-layered approach to safety is crucial to ensure that robots can interact safely with humans and their environment.

Towards General-Purpose Physical Artificial Intelligence

Gemini Robotics 2 represents a key step toward achieving general artificial intelligence (AGI) in the physical world. By going beyond the automation of single tasks, this advancement aims to develop general-purpose intelligence capable of working alongside humans to tackle complex challenges. The goal is to create a foundational intelligence that will allow robots to adapt to a variety of situations and tasks, making their integration into our daily lives smoother and more efficient.