身智能的学习数据走进日常生活 上海交通大学卢策吾团队及穹彻智能研发团队 与语言模型不同，具身智能的学习严重依赖真实世界中的交互数据。语言模型可以利用互联网上天然存 在的海量文本进行训练，这些数据无需额外成本即可持续增长；但具身智能面对的是高维、动态且充满 不确定性的物理环境——从二维平面扩展到包含关节、力觉、接触等在内的八十多维状态空间。这意味 着，它需要大量与真实物体互动的经验才能学会拧瓶盖、叠毛巾或刮胡子这类看似简单却极其复杂的动 作。 机器期待来自全社会的数据 数据就像是具身智能的"粮食"，但这份粮食在过去是稀缺且昂贵的。比如实验室里最常见的做法是搭 建"数据牧场"：圈出上千平方米场地，请来专职操作员，让机器人在固定布景里反复抓取、推拉、旋 拧，四周布满动作捕捉相机和六维力台，一条一分钟的数据成本高达数美元。牧场模式养出的数据干 净、标注精细，却天然带着"天花板"——场地面积有限、人力工时有限、物体品类有限，更关键的是它 无法复制真实世界：塑料袋的静电吸附、药盒上的易撕口、果蔬表面看不见的弧度，这些细枝末节都成 了拦路虎。 团队估算过，想让通用机器人达到人类水准，至少需要几十亿小时的操作片段，相当于把全 ...

Investment Rating - The report does not explicitly provide an investment rating for the industry or specific companies discussed Core Insights - The 2025 World Robot Conference (WRC) highlighted advancements in bionic design, agricultural automation, embodied intelligence AI capabilities, operational cognition optimization, and industrialization pathways [1][14] - Companies like FESTO, Tevel, iFLYTEK, Qiongtor Intelligence, and Auda Voice Medical showcased innovative solutions and technologies in robotics [1][14] Summary by Relevant Sections Bionic Design - FESTO has developed bionic flying robots, including the "BionicBee," which weighs a few grams, has a wingspan of 12 cm, and achieves a wingbeat frequency of 15-20 times per second, showcasing the integration of natural design with engineering [2][15] Agricultural Automation - Tevel addresses the low automation in fruit picking despite a 200% increase in global fruit output over the past 30 years, utilizing flexible gripping and AI decision-making to create a scalable robotic harvesting system [2][16] AI Capability Upgrades - iFLYTEK emphasizes the need for AI to possess capabilities such as autonomous execution, improved human-machine interaction, and global deployment support, integrating VR/AR with large models to enhance task execution [2][17] Operational Cognition - Qiongtor Intelligence introduced the "digital gene" concept to improve operational understanding in robotics, advocating for programmatic language over natural language to enhance generalization and robustness in task execution [3][18] Future Education - The report discusses the transformation of education through AI Agents, emphasizing the need for interdisciplinary skills and adaptability to work alongside AI systems, which is crucial for meeting the demands of the embodied intelligence era [5][21]