• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (17): 81-101.doi: 10.3901/JME.260801

• 特邀专栏:人形机器人与具身智能 • 上一篇    

扫码分享

数据与模型驱动的具身感知技术:进展、挑战与展望

韦锦宇1,2, 闫虹坤1,2, 李栋1,2, 华天宇1,2, 常皓楠1,2, 李峰1,2, 王振华1,2, 孙立宁1,2, 张驰3,4, 陈国栋1,2   

  1. 1. 苏州大学机电工程学院 苏州 215021;
    2. 江苏省具身智能机器人技术重点实验室 苏州 215021;
    3. 杭州海康机器人股份有限公司 杭州 310052;
    4. 哈尔滨工程大学智能科学与工程学院 哈尔滨 150001
  • 收稿日期:2025-12-26 修回日期:2026-04-23 发布日期:2026-11-20
  • 作者简介:韦锦宇,男,2001年出生。主要研究方向为具身智能,多模态感知。E-mail:20245229010@stu.suda.edu.cn;闫虹坤,男,2001年出生,硕士研究生。主要研究方向为具身智能,多模态感知。E-mail:20245229028@stu.suda.edu.cn;李栋,男,1997年出生,博士研究生。主要研究方向为智能机器人技术,多模态感知。E-mail:20254029004@stu.suda.edu.cn;华天宇,男,2002年出生,硕士研究生。主要研究方向为移动机器人,多模态融合定位。E-mail:20245229008@stu.suda.edu.cn;常皓楠,男,2002年出生,硕士研究生。主要研究方向为机械臂路径规划。E-mail:20245229017@stu.suda.edu.cn;李峰,男,2001年出生,硕士研究生。主要研究方向为机器人自动控制,多传感器融合定位。E-mail:20245229036@stu.suda.edu.cn;王振华,男,1974年出生,博士,教授,博士生研究导师。主要研究方向为机器视觉,多模态感知,机器人智能化技术、智能制造技术等。E-mail:Wangzhenhua@suda.edu.cn;孙立宁,男,1964年出生,博士,教授,博士生研究导师。主要研究方向为工业机器人及机电一体化装备、微纳米操作机器人与装备、医疗与特种机器人等。E-mail:zfsh751228@163.com;张驰,女,1987年出生,博士研究生。主要研究方向为机器人技术、人工智能技术等。E-mail:zhangchi3@hikrobotics.com;陈国栋(通信作者),男,1983年出生,博士,教授,博士研究生导师,主要研究方向为机器视觉,多模态感知,机器人智能化技术,具身智 能等。E-mail:chenguodong@suda.edu.cn
  • 基金资助:
    国家重点研发计划(2022YFB4702202)、江苏省重点研发计划(BE2021009-02)、泰安市“双十工程”科技创新(2024JSGG05)、国家自然科学基金重点(U25A20318)和长三角科技创新共同体联合攻关计划(2025CSJGG01200)资助项目。

Data and Model Driven Embodied Perception Technology: Progress, Challenges, and Prospects

WEI Jinyu1,2, YAN Hongkun1,2, LI Dong1,2, HUA Tianyu1,2, CHANG Haonan1,2, LI Feng1,2, WANG Zhenhua1,2, SUN Lining1,2, ZHANG Chi3,4, CHEN Guodong1,2   

  1. 1. School of Mechanical and Electrical Engineering, Soochow University, Suzhou 215021;
    2. Jiangsu Key Laboratory of Embodied Intelligent Robot Technology, Suzhou 215021;
    3. Hangzhou Hikvision Robotics Co., Ltd., Hangzhou 310052;
    4. College of Intelligent System Science and Engineering, Harbin Engineering University, Harbin 150001
  • Received:2025-12-26 Revised:2026-04-23 Published:2026-11-20

摘要: 具身感知是连接智能体行为决策与复杂物理环境交互的关键纽带。近年来,具身感知技术不断融合机器人感知、语义理解与认知推理等领域成果,逐渐形成具有明确研究目标和方法体系的研究方向。然而,现有研究多聚焦于特定任务场景与单点技术突破,尚缺乏对具身感知技术体系的系统梳理。为此,围绕具身感知技术的发展历程,从三个维度系统总结其最新研究进展:在数据源获取层面,梳理了遥操作采集、仿真合成与互联网跨域迁移三类主流方法;在关键技术突破层面,剖析了多模态时空同步、跨模态语义对齐、主动感知与世界模型构建等核心机制;在评估基准建设层面,分类总结了跨平台通用、精细操作及仿真基准等典型数据集。最后,明确指出高维语义空间与低维高频物理交互空间的维度错位是制约技术工程化落地的核心瓶颈,并从虚实融合数据范式、交互式世界模型推演及感知决策端侧协同等维度探讨了未来演进路径,相关探讨可为具身感知的理论研究与工程应用提供体系化参考。

关键词: 具身智能, 具身感知技术, 数据获取方法, 具身数据集

Abstract: Embodied perception is the crucial link connecting an agent's behavioral decision-making with its interactions in complex physical environments. In recent years, embodied perception technology has continuously integrated achievements from fields such as robot perception, semantic understanding, and cognitive reasoning, gradually forming a distinct research direction with clear objectives and methodological frameworks. However, existing research predominantly focuses on specific task scenarios and isolated technological breakthroughs, lacking a systematic review of the embodied perception technological framework. To this end, based on the development of embodied perception technology, this study systematically summarizes the latest research progress from three dimensions. In terms of data acquisition, it reviews three mainstream methods: teleoperation collection, simulation synthesis, and Internet-based cross-domain transfer. Regarding key technological breakthroughs, it analyzes core mechanisms such as multi-modal spatio-temporal synchronization, cross-modal semantic alignment, active perception, and world model construction. Concerning evaluation benchmarks, it categorizes typical datasets including cross-platform universal benchmarks, fine manipulation benchmarks, and simulation benchmarks. Finally, it explicitly identifies the dimensional misalignment between the high-dimensional semantic space and the low-dimensional, high-frequency physical interaction space as the core bottleneck restricting the technology's engineering application. Furthermore, it explores future development paths from the perspectives of virtual-real integrated data paradigms, interactive world model deduction, and the edge-side coordination of perception and decision-making. This discussion aims to provide a systematic reference for both the theoretical research and engineering applications of embodied perception.

Key words: embodied intelligence, embodied perception technology, data acquisition methods, embodied datasets

中图分类号: