Google DeepMind is taking AI beyond screens into the real world with Gemini Robotics, a new model built on Gemini 2.0 that allows robots to see, understand, and interact with their surroundings. This update introduces two AI-driven models, Gemini Robotics and Gemini Robotics-ER which are designed to improve robots' ability to perceive, interact with, and respond intelligently to their surroundings.

For any robots to be truly useful, they need more than just computation; they must be able to see, understand, and act in dynamic environments. Gemini Robotics is built on a vision-language-action (VLA) framework, enabling it to interpret the world, respond to commands, and take physical action.

Key Features:

1. Generalization: Gemini Robotics now has the ability to adapts to new objects, instructions, and environments including the ones it hasn't seen before, making it highly flexible.
2. Interactivity: It understands natural language commands, adjusts to real-time changes, and responds conversationally in multiple languages.
3. Dexterity: It can manipulate delicate objects and perform fine motor tasks, making it valuable for both household and industrial applications.
4. Multiple embodiments: Gemini Robotics is designed to seamlessly integrate to different robot of different sizes and shapes

Gemini Robotics-ER: Enhancing Spatial Intelligence

Alongside Gemini Robotics, DeepMind is introducing Gemini Robotics-ER, a model designed to improve spatial understanding and reasoning in robotics. This extension allows robots to:

1. Interpret 3D environments with enhanced perception.
2. Plan and execute precise movements, such as grasping or manipulating objects.
3. Autonomously complete complex tasks, leveraging real-world data to refine decision-making.

By bridging high-level AI reasoning with low-level robot control systems, Gemini Robotics-ER enables roboticists to integrate its capabilities into existing robotic frameworks more efficiently.

Advancing AI Safety in Robotics

As AI-driven robots become more capable, safety remains a priority. DeepMind is implementing multiple layers of protection:

1. Physical Safety: Ensuring robots avoid collisions, maintain stability, and limit contact forces.
2. Semantic Safety: Developing a data-driven constitution to guide robot behavior and prevent unsafe actions.
3. Responsible Development: Collaborating with trusted testers including Boston Dynamics, Agile Robots, Agility Robotics, and Enchanted Tools to rigorously evaluate Gemini Robotics-ER’s real-world impact.

DeepMind is also introducing ASIMOV, a dataset designed to evaluate and improve AI safety in robotic applications, reinforcing a structured approach to responsible AI development.

The future of robotics isn’t just about automation—it’s about AI that perceives, learns, and collaborates with us in everyday life.