**Introducing Gemini Robotics: Advanced Models for Robotics** At Google DeepMind, we have made significant strides in our Gemini models’ ability to tackle complex problems through multimodal reasoning, encompassing text, images, audio, and video. However, to be truly beneficial in the physical world, AI must exhibit “embodied” reasoning, which allows it to perceive and interact with its environment effectively. Today, we are unveiling two innovative AI models based on Gemini 2. 0, which will pave the way for a new era of helpful robotic applications. 1. **Gemini Robotics**: This is a state-of-the-art vision-language-action (VLA) model that integrates physical actions, enabling direct control of robots. 2. **Gemini Robotics-ER**: This model enhances spatial understanding and allows roboticists to implement their programs utilizing Gemini’s embodied reasoning capabilities. Together, these models enable a broader range of robots to undertake various real-world tasks. We are collaborating with Apptronik to develop the next generation of humanoid robots powered by Gemini 2. 0 and are also working with selected trusted testers to refine Gemini Robotics-ER. **Key Features of Gemini Robotics** - **Generality**: Gemini Robotics utilizes Gemini's world knowledge to adapt and handle unforeseen tasks and environments efficiently.

It performs more than twice as effectively on generalization benchmarks compared to other leading models. - **Interactivity**: The model can interact seamlessly in dynamic settings, understanding natural language commands in various languages, and promptly adapting its actions based on real-time environmental changes. - **Dexterity**: Unlike previous models, Gemini Robotics handles intricate tasks requiring fine motor skills—such as origami or packing snacks—showing significant advancements in physical manipulation. - **Multiple Embodiments**: Designed for versatility, Gemini Robotics trained primarily on the ALOHA 2 bi-arm platform but is capable of controlling other platforms, including those used in academic labs and humanoid robots like the Apollo from Apptronik. **Enhancements with Gemini Robotics-ER** Alongside Gemini Robotics, we introduce Gemini Robotics-ER, focusing on advanced spatial reasoning. This model refines Gemini’s abilities such as 3D detection and grasping, enabling it to carry out tasks like safely lifting a coffee mug by utilizing an appropriate grip. Gemini Robotics-ER can accomplish critical functions like perception and planning with a 2x-3x success rate over Gemini 2. 0. It utilizes context-based learning from human demonstrations to enhance its problem-solving capabilities. **Safety in AI and Robotics** We take a comprehensive approach to ensure safety and address concerns in robotics, from motor control to understanding complex actions. Our robotic systems incorporate traditional safety measures alongside advanced understanding to assess the safety of potential actions. To promote research in robotics safety, we are releasing a new dataset aimed at evaluating semantic safety in embodied AI. Building on concepts such as a Robot Constitution inspired by Asimov's laws, we have developed a framework for creating data-driven rules to guide robot behavior. Furthermore, we engage with our Responsible Development and Innovation team and the Responsibility and Safety Council to ensure ethical and safe AI applications. We collaborate with industry leaders such as Agile Robots, Boston Dynamics, and Enchanted Tools as we refine our technology. As we continue to develop these next-generation robots, we are eager to uncover and enhance the capabilities of our models in creating impactful robotic solutions.