具身智能的研究与应用  

Research and application of embodied intelligence

作　　者：张伟男[1] 刘挺[1] ZHANG Weinan;LIU Ting(Faculty of Computing,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]哈尔滨工业大学计算学部,黑龙江哈尔滨150001

出　　处：《智能系统学报》2025年第1期255-262,共8页CAAI Transactions on Intelligent Systems

基　　金：国家重点研发计划项目(2022YFF0902100);国家自然科学基金项目(92470205);黑龙江省自然科学基金项目(YQ2021F006).

摘　　要：随着深度学习和大模型技术的不断增强,人工智能技术从研究简单、封闭的虚拟场景,发展到研究更为复杂、开放的现实场景。研究焦点也从早期的小规模语料库和网络文本数据集处理,发展到多模态一体化的处理架构和研究范式。与此同时,以OpenAI Sora为代表的物理世界近似和仿真模型的出现,标志着人工智能再次向通用人工智能迈进了一步。然而,若要让人工智能真正达到通用人工智能的标准,成为类人的智能,需要当今的人工智能体具备与物理世界交互学习的能力,即具身智能。因此,本文主要关注具身智能的研究内容和进展,具体包括具身感知、具身认知和具身行为优化3个方面。同时结合近期人形机器人的发展,概述具身智能技术在人形机器人等载体上的应用,并对未来的研究及应用进行展望。With the continuous enhancement of deep learning and large model technologies,artificial intelligence has evolved from studying simple,closed virtual environments to exploring more complex,open real-world scenarios.The research focus has shifted from early small-scale corpus and web text data processing to multimodal integrated processing architectures and research paradigms.Meanwhile,the emergence of physical world approximation and simulation models,represented by OpenAI Sora,marks another step forward in the progress towards general artificial intelligence(AGI).However,for AI to truly reach the standard of AGI and become human-like intelligence,it must possess the ability to interact and learn from the physical world—this is known as embodied intelligence.Therefore,this paper primarily focuses on the research content and progress of embodied intelligence,specifically including embodied perception,embodied cognition,and embodied action optimization.Additionally,considering recent developments in humanoid robots,this paper outlines the application of embodied intelligence technologies in carriers such as humanoid robots and provides an outlook on research and applications in this field.

关 键 词：具身智能 具身感知 具身认知 具身行为优化 深度学习 人工智能 仿真环境 人形机器人 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP242.6[自动化与计算机技术—计算机科学与技术]