ZurichRobotics #3

Michael Neunert (Google Deepmind): Building AI agents for the physical world The realization of truly general-purpose robots relies on three fundamental capabilities: a deep understanding of the physical world, advanced reasoning, and highly dexterous control. We will explore the development and impact of two groundbreaking models: Gemini Robotics 1.5, a multi-embodiment Vision-Language-Action (VLA) model, and Gemini Robotics-ER 1.5, our state-of-the-art Embodied Reasoning (ER) model. We will examine three major innovations driving this new generation of physical agents: - Multi-Embodiment Learning via Motion Transfer (MT): How a novel architecture allows the VLA model to learn seamlessly from heterogeneous data across different types of robotic hardware - "Thinking Before Acting": How interleaving actions with a multi-level internal reasoning process in natural language greatly improves a robot's ability to decompose complex tasks while remaining interpretable to human users. - State-of-the-Art Embodied Reasoning: How Gemini Robotics-ER 1.5 establishes new benchmarks in critical areas…