基于世界模型的通用机器人策略学习-清华大学ISRLab在读博士生、星动纪元实习研究员郭彦江.pdfVIP

基于世界模型的通用机器人策略学习

YanjiangGuo

ISRLab,IIIS,TsinghuaUniversity

ResearchIntern,RobotEra

2025-4-7ResearchSummary1

About

Info:YanjiangGuo(郭彦江）Page:https://robert-gyj.github.io/

Education

•2018-2022:bachelor’sdegreefromDept.EE,TsinghuaUniversity

•2022-present:PhDstudentatIIIS,TsinghuaUniversity

ResearchTarget:Buildingfoundationagentthatcansolvesvarioustasks.

•Vision-Language-ActionModels

•VideoGenerationModels(WorldModels)forgeneralistpolicylearning

•ReinforcementLearning

Advisor:Prof.JianyuChen,alsothefounderofstartupRobotEra(星动纪元)

2025-4-7ResearchSummary2

Outline

•通用机器人最大的难点：通用操作模型

•数据稀缺

•如何利用互联网数据，仿真数据

•基于视觉语言模型微调的VLA模型

•+基于语言数据集预训练，擅长语义理解，抽象符号的推理

•-不擅长表达物理世界的动态变化

•基于视频生成基础模型的通用机器人策略

•+基于视频数据集预训练，更擅长表示物体的发展变化（dynamics）

•-不擅长复杂推理

2025-4-7ISRLab,InstituteforInterdisciplinaryInformationSciences,TsinghuaUniversity3

OnbuildingFoundationAgent:Fastandslowthinking

•Thinking,Fastandslow

•System1forfastactions/actuations.

•System2forslow,logicalplanning

2025-4-7ISRLab,InstituteforInterdisciplinaryInformationSciences,TsinghuaUniversity4

OnbuildingFoundationAgent:Fastandslowthinking

•System2ismuchmorepowerfulthanSystem1atthismoment.

•VLM/LLMagentforsystem2.

•MyresearchesmainlyfocusonSystem1learning.

•Navigation,locomotion,manipulation…

2025-4-7ISRLab,InstituteforInterdisciplinaryInformationSciences,TsinghuaUniversity5

OnbuildingFoundationAgent:Fastandslowthinking

•“Maybeweunderscorehowmuchintellig