• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (17): 189-211.doi: 10.3901/JME.260754

• 特邀专栏:人形机器人与具身智能 • 上一篇    

扫码分享

面向任务可迁移的具身智能系统架构:一种基于功能抽象的框架实例

解杨敏1,2, 杨巧妮3, 姜逸川3, 师文帅3, 张浩4, 许毅伟3, 刘锴3, 张敬寒3, 施宇杰3, 杨雨生1,2   

  1. 1. 上海市智能制造及机器人重点实验室 上海 200444;
    2. 上海大学特种人形机器人研究院 上海 200444;
    3. 上海大学机电工程与自动化学院 上海 200444;
    4. 上海大学未来技术学院 上海 200444
  • 收稿日期:2025-12-30 修回日期:2026-04-27 发布日期:2026-11-20
  • 作者简介:解杨敏,女,1984年出生,教授、博士生导师。主要研究方向为自主导航、环境重构、操控决策、集群智能、具身智能。E-mail:xieym@shu.edu.cn;杨雨生(通信作者),男,1992年出生,博士,讲师。主要研究方向为操控决策、路径规划、自主导航。E-mail:yysshu@shu.edu.cn
  • 基金资助:
    国家自然科学基金资助项目(62173220,62303294)。

A Task-transferable Architecture for Embodied Intelligence Systems: A Function-abstraction-based Framework Instance

XIE Yangmin1,2, YANG Qiaoni3, JIANG Yichuan3, SHI Wenshuai3, ZHANG Hao4, XU Yiwei3, LIU Kai3, ZHANG Jinghan3, SHI Yujie3, YANG Yusheng1,2   

  1. 1. Shanghai Key Laboratory of Intelligent Manufacturing and Robotics, Shanghai 200444;
    2. Bionic Intelligent Robot Dreamworks for Specialpurpose at Shanghai University, Shanghai 200444;
    3. School of Mechatronic Engineering and Automation, Shanghai University, Shanghai 200444;
    4. School of Future Technology, Shanghai University, Shanghai 200444
  • Received:2025-12-30 Revised:2026-04-27 Published:2026-11-20

摘要: 针对具身智能系统在多任务、多平台应用中普遍存在的软件复用性不足与任务迁移成本高的问题,本文提出一种基于功能抽象的任务可迁移具身智能系统架构。该架构在任务需求与具体算法实现之间引入稳定功能层,将任务变化转化为功能组合与配置问题,实现任务逻辑、系统功能与算法实现的解耦。围绕该思想,构建了一套包含边缘感知、大脑系统与小脑系统的具身智能架构实例,并定义了统一的数据接口与功能-算法映射机制,以支撑算法替换与系统演进。通过在移动机械臂自主抓取与人形机器人跨楼层巡逻两类复杂任务上的实验验证,展示了该架构在短流程高耦合任务与长时序多阶段任务中的适用性。同时,提出任务迁移指数(Task transfer index, TTI),量化分析不同任务间的系统迁移新增工作量。实验结果表明,相较于传统基于算法模块的组合方式,所提出的功能抽象架构能够显著降低任务迁移的新增工作量,为真实场景中具身智能系统的快速构建与复用提供了一种可行的软件架构路径。

关键词: 具身智能, 功能抽象, 系统架构设计, 任务迁移性

Abstract: This study presents a task-transferable system architecture based on functional abstraction to address the low software reusability and high task migration costs in embodied intelligence systems. By introducing a functional layer between task requirements and algorithm implementations, task variations are reduced to problems of functional composition and configuration, decoupling task logic, system functionality, and algorithms. Based on this principle, we construct an embodied intelligence system with edge perception, cognitive, and motor control systems, defining unified data interfaces and a function-algorithm mapping mechanism for algorithm interchangeability and system evolution. The architecture is validated through experiments on two complex tasks: autonomous grasping with a mobile manipulator and cross-floor patrol by a humanoid robot. A Task Transfer Index (TTI) is introduced to quantify the additional effort required during task migration. Experimental results show that the proposed architecture significantly reduces development costs compared to traditional module-based designs, offering a practical solution for rapid system construction and reuse.

Key words: embodied intelligence, functional abstraction, system architecture design, task transferability

中图分类号: