Zhewei Gui

I am a robotics researcher and engineer working on dexterous manipulation, reinforcement learning, and Sim-to-Real system. I received my M.Sc. in Robotics from the National University of Singapore and my B.Eng. from Tongji University.

My research focuses on low-level dexterous foundation controllers. I have conducted academic research on robot learning for dexterous manipulation at the Center on Frontiers of Computing Studies (CFCS) in the School of Computer Science at Peking University, advised by Prof. Hao Dong, and at the LinS Lab at the National University of Singapore, advised by Prof. Lin Shao.

In industry, my research focused on reinforcement learning and Sim-to-Real system at JD, AgiBot, and Fourier Intelligence.

I am currently seeking full-time industry roles and Ph.D. opportunities in robot learning, dexterous manipulation, and embodied AI.

/ZheweiGui@outlook.com

Research & Publications

Dexterous manipulation

DexHOT configuration transfer and long-horizon tasks

DexHOT: Learning Hand–Object Configuration Transfer for Dexterous Manipulation

Zhewei Gui*, Anthony Bolton*, Jiyao Zhang‡, Zeyuan Chen, Qiyang Yan, Tianyu Wang, Hao Dong†

arXiv preprint · 2026 First author

[paper][website]

DexHOT learns a single relation-conditioned policy for dexterous hand–object configuration transfer from synthetic feasible configuration pairs, without human data during policy training. At deployment, configuration subgoals enable Any2Any transfers and human demonstration-guided long-horizon tasks. On DexHOTBench with 120 objects, it achieves 80.5% / 69.4% transfer success on seen / unseen objects in simulation, 61.5% over 200 real-robot transfers, and 90% task success across six long-horizon tasks.

DexSinGrasp singulation and grasping results in dense, random, and irregular clutter

DexSinGrasp: Learning a Unified Policy for Dexterous Object Singulation and Grasping in Densely Cluttered Environments

Lixin Xu*, Zixuan Liu*, Zhewei Gui, Jingxiang Guo, Zeyu Jiang, Tongzhou Zhang, Zhixuan Xu, Chongkai Gao, Lin Shao

IEEE Robotics and Automation Letters (RA-L) · 2025

Contribution: Participated in the RL setting design and completed the sim-to-real system.

[paper][website][code]

DexSinGrasp learns a unified policy for dexterous object singulation and grasping in clutter. Curriculum learning improves generalization across diverse clutter conditions, while policy distillation produces a deployable vision-based strategy. Experiments show higher efficiency and grasping success, particularly in dense clutter.