• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (17): 40-64.doi: 10.3901/JME.260636

• 特邀专栏:人形机器人与具身智能 • 上一篇    

扫码分享

面向复杂操作的人形灵巧手:结构、感知与控制关键技术

唐志强1, 杨骑郡1, 王兴松1, 王柏村2, 訾斌3, 王乾乾1   

  1. 1. 东南大学机械工程学院 南京 211189;
    2. 浙江大学机械工程学院 杭州 310058;
    3. 西安电子科技大学机电工程学院 西安 710071
  • 收稿日期:2025-11-26 修回日期:2026-05-14 发布日期:2026-11-20
  • 作者简介:唐志强,男,1994年出生,博士,副教授。主要研究方向为软体机器人智能控制。E-mail:zhiqiang.tang@seu.edu.cn;王乾乾(通信作者),男,1992年出生,博士,教授,博士研究生导师。主要研究方向为外场驱控微机器人、医工交叉机器人系统。E-mail:qqwang@seu.edu.cn
  • 基金资助:
    国家自然科学基金资助项目(52205590,52575652)。

Anthropomorphic Dexterous Hands for Complex Manipulation: A Review of Key Technologies in Structure, Sensing, and Control

TANG Zhiqiang1, YANG Qijun1, WANG Xingsong1, WANG Baicun2, ZI Bin3, WANG Qianqian1   

  1. 1. School of Mechanical Engineering, Southeast University, Nanjing 211189;
    2. School of Mechanical Engineering, Zhejiang University, Hangzhou 310058;
    3. School of Mechatronic Engineering, Xidian University, Xi'an 710071
  • Received:2025-11-26 Revised:2026-05-14 Published:2026-11-20

摘要: 在智能制造、医疗康复与人机协作等需求推动下,具有人手形态与精细操作能力的人形灵巧手成为关键发展方向。以“五指仿生、面向复杂操作”为主线,回顾由结构仿生迈向智能操控的演进路径,强调结构复杂度、任务泛化与感知控制耦合之间的协同优化。首先,文章从设计理念走向物理实现,系统阐明如何将结构参数、欠驱动和柔顺设计等原则,与功能材料和增材制造相结合,形成高性能灵巧手结构落地方案。其次,在感知层面归纳角度与应变、触觉与压力等多模态方案,强调紧凑化视觉触觉一体化模组与信息融合对控制策略学习的促进作用。再次,在控制层面对比基于解析模型的经典控制方法、数据驱动的学习控制方法、混合模型与数据的控制方法三类路径,突出残差强化学习、物理可微仿真与大规模抓取数据在样本效率、鲁棒性与可迁移性方面的协同价值。最后,面向工程化部署提出模块化与标准接口以及软硬件协同封装与统一总线等集成范式,以缓解接口异构与布线协议冲突带来的落地障碍,支持跨平台复用与快速迭代。综上,形成以结构参数设计、紧凑多模态感知、融合式控制与模块化集成为核心的技术框架,为灵巧手在复杂操作与人机共融场景中的规模化应用提供方法参考与实践路径。

关键词: 人形灵巧手, 结构设计, 增材制造, 多模态感知, 融合控制

Abstract: Anthropomorphic dexterous hands are a central focus for fine manipulation. Demand arises from intelligent manufacturing, medical rehabilitation, and human–robot collaboration. This work follows the theme of five-finger bionics for complex manipulation and traces the shift from structural bionics to functional intelligence. It emphasizes coordinated trade-offs among structural complexity, task generalization, and sensing–control coupling. Building on this foundation, the path from design concepts to physical realization is clarified. Structural parameter design, underactuation, and compliance are linked with functional material selection and multi-material additive manufacturing. These links enable high-performance implementations with reliable grasping and stable manipulation. On the perception side, joint angle and strain sensing, tactile and pressure sensing, and vision are organized into a compact multimodal architecture. Vision-tactile integrated modules and early or late fusion strategies enhance high-dimensional representation and support policy learning. On the control side, analytical modeling, data-driven learning, and hybrid modeling-learning are compared within a unified view. Residual reinforcement learning, physics-differentiable simulation, and large-scale grasping datasets jointly improve sample efficiency, robustness, and transferability. For engineering deployment, modular architectures and standardized interfaces are advocated. Co-packaged hardware and software with unified buses reduce interface heterogeneity and wiring or protocol conflicts, which supports cross-platform reuse and rapid iteration. These elements establish a coherent framework for the scalable deployment of dexterous hands in complex manipulation and human-robot collaboration.

Key words: anthropomorphic dexterous hands, structural design, multi-material manufacturing, multimodal perception, integrated control

中图分类号: