Gemini Robotics 2: Embodied Reasoning and Whole-Body Intelligence for Robots

This Release Notes episode features a conversation with Google DeepMind team members at the Gemini Robotics lab about the launch of Gemini Robotics 2, a suite of models designed to bring embodied intelligence to robots. The discussion covers the new embodied reasoning model, whole-body intelligence, dexterity, multi-robot collaboration, the research arc leading to this release, technical challenges, deployment considerations, and model availability.

■ Gemini Robotics 2 capabilities
- Whole-body intelligence, dexterity, and multi-robot collaboration
- Embodied reasoning and action models for controlling various robot embodiments

■ Technical challenges and research arc
- Evolution from reinforcement learning and sim-to-real to vision-language-action models
- Data pyramid, physical interaction data, and dexterous manipulation as an unsolved problem

■ Availability and safety
- Embodied Reasoning 2 via AI Studio and Gemini Enterprise Agents Platform
- Action models through partners and on-device trusted testers
- Asimov Agentic safety benchmark for semantic physical understanding

Viewers interested in robotics, embodied AI, and foundation models will gain an overview of the current capabilities, open problems, and access paths for Gemini Robotics 2. The episode provides context for understanding the state of general-purpose robotics and the steps required for real-world deployment.