Yiwei Liu  |  刘益伟

I am an independent scholar, working in Tencent Hunyuan as a VLM Evaluation Engineer. Before this, I received my B.E. degree from College of Software Engineering, Sichuan University in 2025, supervised by Prof. Lin Shao (邵林) at National University of Singapore.

Email  /  GitHub  /  WeChat

profile photo

Research

My research interests lie in 🎞️ Computer Vision, 🤖 Embodied AI, and 🏦 Fintech.
Lately, I've been immersed in the engineering implementation of agent benchmarks (e.g., GUI agents), with a particular focus on how to improve the evaluation automation efficiency.
I'm open to collaborations on agent-related projects and welcome any online coffee chat!

Publications (* denotes equal contribution)

Bi-Adapt Bi-Adapt: Few-shot Bimanual Adaptation for Novel Categories of 3D Objects via Semantic Correspondence
Jinxian Zhou*, Ruihai Wu*, Yiwei Liu, Yiwen Hou, Xunzhe Zhou, Checheng Yu, Licheng Zhong, Lin Shao
arXiv  /  Website  /  Bibtex
ICRA 2026  International Conference on Robotics and Automation
ICRA 2026 Best Paper Finalist on Robot Manipulation and Locomotion
Oral Presentation, ICRA 2026
TL;DR: Bi-Adapt integrates category-specific object features, task-level manipulation goals, and a few-shot adaptation strategy based on contact point selection to enable bi-manual manipulation.
Uncertainty-based Ensemble Learning in CMR Semantic Segmentation Uncertainty-based Ensemble Learning in CMR Semantic Segmentation
Yiwei Liu, Liang Zhong, Lingyi Wen, Yuankai Wu
arXiv  /  Code  /  Bibtex
ICASSP 2026  International Conference on Acoustics, Speech, and Signal Processing
Poster Presentation, ICASSP 2026
TL;DR: We propose an ensemble learning paradigm, Streaming, to enhance end-slice accuracy while maintaining near-SOTA DSC performance on cardiac CMR datasets.
Manual2Skill Manual2Skill: Learning to Read Manuals and Acquire Robotic Skills for Furniture Assembly Using Vision-Language Models
Chenrui Tie*, Shengxiang Sun*, Jinxuan Zhu, Yiwei Liu, Jingxiang Guo, Yue Hu, Haonan Chen, Junting Chen, Ruihai Wu, Lin Shao
arXiv  /  Website  /  Bibtex  /  Media (机器之心)
RSS 2025  Robotics: Science and Systems
Oral Presentation, CVPR 2025 @ 3D Vision Language Models for Robotic Manipulation
TL;DR: Manual2Skill enables robots to autonomously perform complex assembly tasks in real-life guided by high-level manual instructions.
MetaFold MetaFold: Language-Guided Multi-Category Garment Folding Framework via Trajectory Generation and Foundation Model
Haonan Chen*, Junxiao Li*, Ruihai Wu, Yiwei Liu, Yiwen Hou, Zhixuan Xu, Jingxiang Guo, Chongkai Gao, Zhenyu Wei, Shensi Xu, Jiaqi Huang, Lin Shao
arXiv  /  Website  /  Bibtex  /  Media (机器之心)
IROS 2025  Intelligent Robots and Systems
Oral Presentation, IROS 2025
TL;DR: MetaFold employs language-guided point cloud trajectories for task planning and a foundation model for action prediction, enabling better generalization across garments and instructions.

Education

National University of Singapore 2024.08 - 2025.05

NGNE Program Exchange Student, School of Computing.
Summer Workshop First Prize
An Interesting Project: Motion Sensing FPS Game (CS3247 Game Development)

Sichuan University 2021.09 - 2025.06

B.E. in Software Engineering
Mentor:Yuankai Wu (伍元凯), Computer Science
Xuebin Chen (陈学彬), Economics

Employment

Tencent, Shenzhen, China 2026.09 - Present

VLM Evaluation Engineer, Hunyuan

ByteDance, Beijing, China 2025.07 - 2026.08

VLM Evaluation Engineer, AI Data & Safety

Siemens Industrial Automation Products (SEWC), Chengdu, China 2024.01 - 2024.06

Embodied AI Intern, IdeaWorks
An Interesting Demo (early attempt on language-guided pick&place task)

Miscellanea
Awards
  • ICRA 2026 Best Paper Finalist on Robot Manipulation and Locomotion 2026
  • Meritorious Winner in Mathematical Contest in Modeling (MCM) by COMAP 2024
  • National Second Prize of "Citi Cup" Financial Innovation and Application Competition by CSTC 2024
Stories

Thanks for your visiting😊! Feel free to contact me if you have any problems.
This website is designed based on Jon Barron and Jingxiang Guo. Last Update: