AI Podcast › 卡罗琳娜·帕拉达 › 本期从强化学习到 Gemini Robotics：具身智能的进化之路 From Reinforcement Learning to Gemini Robotics: The Evolution of Embodied AI
卡罗琳娜·帕拉达 Carolina Parada · Google DeepMind · 2025-05-22 · 约 46 分钟 · 原视频 ↗
打开互动全文版（中英对照 + 朗读 + 问答）→
本期速览 · Overview Google DeepMind 机器人研究负责人 Karolina Parada 探讨机器人技术的戏剧性演变，从教机器人叠积木到最新的 Gemini Robotics 模型，将多模态理解带入物理世界。
Karolina Parada discusses the dramatic evolution of robotics at Google DeepMind, from teaching robots to stack blocks to the recent Gemini Robotics model that brings multimodal understanding to the physical world.
要点 · TL;DR AI 让通用机器人能够推理并适应各种任务。 AI enables general-purpose robots that reason and adapt across tasks. Gemini Robotics 将多模态理解融入物理动作。 Gemini Robotics integrates multimodal understanding into physical actions. 机器人技术的进步从几十年缩短到 5-10 年。 Robotics progress has accelerated from decades to 5-10 years. 核心观点 · Key points AI 是构建能推理和适应的通用机器人的关键。 AI is key to building general-purpose robots that reason and adapt. Gemini Robotics 将多模态理解带入物理动作。 Gemini Robotics brings multimodal understanding to physical actions. 机器人现在结合了慢速推理和快速反应系统。 Robots now combine slow reasoning with fast reactive systems. 通过遥操作和扩散模型提高了灵巧性。 Dexterity improved via teleoperation and diffusion models. 安全需要分层方法，包括语义物理安全。 Safety requires layered approaches including semantic physical safety. 机器人技术的进展已从几十年缩短到 5-10 年。 Robotics progress has shifted from decades to 5-10 years. 反共识 · Contrarian takes 机器人无需显式训练即可泛化到未见任务。 Robots can generalize to unseen tasks without explicit training. 简单的机器人硬件通过 AI 可实现复杂操作。 Simple robot hardware can achieve complex manipulation via AI. 多摄像头视图无需显式深度感知即可理解深度。 Multiple camera views enable depth understanding without explicit depth sensing. 强化学习对全身控制仍然至关重要。 Reinforcement learning remains crucial for whole-body control. 模拟到现实的差距仍然存在，但许多任务已缩小。 Sim-to-real gap persists but is reduced for many tasks. 机器人可以通过最少的人类示范在工作中学习。 Robots can learn on the job with minimal human demonstrations. 本期章节 · Chapters（共 19） 引言与机器人进化 Introduction and Robotics Evolution 当前能力与未来目标 Current Capabilities and Future Goals 用玩具展示概念理解 Robot's conceptual understanding demonstrated with toys 交互场景与模型适应性 Interactive scenarios and model adaptability 机器人多模态理解 Multimodal Understanding in Robotics 概念理解的重要性 Why Conceptual Understanding Matters 指向与边界框 Pointing and Bounding Boxes 具身推理与标准推理 Embodied Reasoning vs Standard Reasoning 从 2D 到 3D 理解 From 2D to 3D Understanding 增强物理与运动理解 Enhancing physical reasoning and motion understanding 灵巧性与遥操作 Dexterity and Teleoperation 技能的意外涌现 Surprising Emergence of Skills 莫拉维克悖论与学习速度 Moravec's Paradox and Learning Speed 泛化与专业化 Generalization vs Specialization 强化学习的作用 Role of Reinforcement Learning 全身控制的强化学习 Reinforcement Learning for Whole-Body Control 实际部署与安全 Real-World Deployment and Safety 物理安全数据集 ASIMOV ASIMOV Dataset for Physical Safety 结论与行动号召 Conclusion and Call to Action 阅读全文双语转录 →