摘要
工业具身智能作为新一代人工智能的热点发展方向,可通过融合多模态大模型对世界的认知能力,驱动工业世界中的物理实体(如各类工业机器)实现“感知–决策–执行”的端到端闭环智能系统,从而为未来智能制造的实现提供重要支撑.然而,工业界各类复杂的应用场景具有操作对象多变性、操作任务多样性等特点,工业场景构建以及数据采集成本高昂,造成工业具身世界模型的训练场景与数据极度匮乏,成为该方向发展的瓶颈挑战.为此,本文提出了“数字族谱”的全新概念,并定义了基于数字族谱DG-DNA基因的生成机制.基于DG-DNA,能够生成从多样零件到多型产品的多代演化产品数字族谱,也能够生成各制造阶段中包含不同设备及环境配置的多样生产空间数字族谱.基于此可构建支撑工业具身智能世界模型训练的场景与数据,以驱动更具操作对象泛化性与空间适应性的工业具身智能的实现.首先,提出了数字族谱的定义与概念内涵;其次,提出了数字族谱系统体系架构;同时,提出了数字族谱的生成构建方法;然后,阐述了面向产品全生命周期的数字族谱典型应用场景,并给出了一个基于数字族谱DG-VAE生成方法的工业零件定制化创新设计应用案例;最后,探讨和展望了数字族谱未来的研究方向和在其他领域的应用延伸.本文为工业具身智能世界模型以及大规模高质量数据训练场的构建提供了一种新的思路与方法,为数字族谱这一全新概念的开创与发展,提供基础理论、关键技术和行业应用的全面指导.
Industrial embodied intelligence,as a cutting-edge development in artificial intelligence,can drive the realization of“perception-decision-execution”end-to-end intelligent systems for physical entities in the industrial world(such as various industrial machines)by integrating multimodal foundation models for world cognition.This provides crucial support for the future of smart manufacturing.However,complex industrial applications face challenges such as diverse operation objects,varied tasks,and high costs for scenario construction and data collection,leading to a shortage of training scenarios and data for industrial embodied world models.To address this,this paper introduces the novel concept of“digital genealogy”and defines the generation mechanism based on DG-DNA.DG-DNA enables the creation of multi-generation evolutionary product digital genealogies,from diverse parts to multi-type products,and the generation of production space digital genealogies that include various equipment and environment configurations at different manufacturing stages.This approach supports the training of industrial embodied intelligence world models,enhancing object generalization and spatial adaptability.The paper presents the definition and conceptual framework of digital genealogy,its generation methods,and typical applications throughout the product lifecycle with a case of industrial parts customized design via proposed DG-VAE in digital genealogy.It explores future research directions and potential applications in other domains.This work provides a novel idea and method to generate high-quality training scenes and data for industrial embodied world models.Also,comprehensive guidance is provided for the foundational theory,key technologies,and industry applications of digital genealogy.
作者
任磊
董家宝
曾宪超
王宇清
杨凌远
赖李媛君
张霖
李伯虎
Lei REN;Jiabao DONG;Xianchao ZENG;Yuqing WANG;Lingyuan YANG;Yuanjun LAILI;Lin ZHANG;Bohu LI(School of Automation Science and Electrical Engineering,Beihang University,Beijing 100191,China;Hangzhou International Innovation Institute,Beihang University,Hangzhou 311115,China;Shanghai Innovation Institute,Shanghai 200231,China)
出处
《中国科学:信息科学》
北大核心
2025年第7期1748-1765,共18页
Scientia Sinica(Informationis)
基金
国家杰出青年科学基金(批准号:62225302)
国家自然科学基金青年学生基础研究项目(批准号:624B2021)资助。
关键词
数字族谱
具身智能
数字族谱DNA
工业大模型
多模态大模型
世界模型
空间智能
生成式人工智能
工业智能体
智能制造
数字孪生
digital genealogy
embodied intelligence
digital genealogy DNA
industrial foundation model
multimodal foundation model
world model
spatial intelligence
generative artificial intelligence
industrial agent
smart manufacturing
digital twin
