Wendi Chen | 陈文迪

I'm a Computer Science Ph.D. student at Shanghai Jiao Tong University (SJTU) and Shanghai Innovation Institute (SII). I'm a member of Machine Vision and Intelligence Group (MVIG) supervised by Prof. Cewu Lu (卢策吾). I also work closely with Prof. Chuan Wen on my research. Before that, I received my B.E. degree of Computer Science at SJTU.

My research interests focus on Embodied AI and Robotics. Specifically, I am interested in how to extend robotic manipulation to a wider range of tasks, such as those involving deformable objects, contact-rich interactions, and dexterous tasks, by improving hardware, data, and algorithms.

I'm always open to meeting people from diverse backgrounds. If you're interested in my research or just want to chat, free feel to reach out to me via Email or WeChat.

Email  /  WeChat (微信)  /  Twitter (X)  /  Google Scholar  /  Github

profile photo

News

Publications

Representative papers are highlighted.

GIF of ImplicitRDP Demos
click to view the proposed framework
ImplicitRDP Framework
ImplicitRDP: An End-to-End Visual-Force Diffusion Policy with Structural Slow-Fast Learning
Wendi Chen, Han Xue, Yi Wang, Fangyuan Zhou, Jun Lv, Yang Jin, Shirun Tang, Chuan Wen\( \dagger \), Cewu Lu\( \dagger \) (\( \dagger \)equal advising)
project page

ImplicitRDP is a unified end-to-end visual-force diffusion policy that integrates visual planning and reactive force control. By leveraging Structural Slow-Fast Learning, it performs closed-loop adjustments at high frequency while maintaining temporal coherence. Additionally, Virtual-target-based Representation Regularization prevents modality collapse, enabling adaptive attention to visual and force modalities.

GIF of SOE Demos
click to view the proposed framework
SOE Framework
SOE: Sample-Efficient Robot Policy Self-Improvement via On-Manifold Exploration
Yang Jin, Jun Lv, Han Xue, Wendi Chen, Chuan Wen\( \dagger \), Cewu Lu\( \dagger \) (\( \dagger \)equal advising)
arXiv preprint, 2025
project page / paper / arXiv / bibtex

We propose a plug-and-play module that enables constrained exploration on the manifold of valid actions for robotic policies. In this way, our model can generate diverse and consistent actions, supporting sample-efficient policy self-improvement.

GIF of Right-Side-Out Demos
click to view the proposed framework
Right-Side-Out Framework
Right-Side-Out: Learning Zero-Shot Sim-to-Real Garment Reversal
Chang Yu*, Siyu Ma*, Wenxin Du, Zeshun Zong, Han Xue, Wendi Chen, Cewu Lu, Yin Yang, Xuchen Han, Joseph Masterjohn, Alejandro Castro, Chenfanfu Jiang, (*equal contributions)
arXiv preprint, 2025
project page / paper / arXiv / bibtex

Right-Side-Out is a zero-shot sim-to-real framework that turns garments right-side out by decomposing the task into keypoint-parameterized primitives and scaling training via high-fidelity GPU-parallel MPM simulation.

GIF of RDP Demos
click to view the proposed framework
TactAR and RDP Framework
Reactive Diffusion Policy: Slow-Fast Visual-Tactile Policy Learning for Contact-Rich Manipulation
Han Xue*, Jieji Ren*, Wendi Chen*, Gu Zhang\( \dagger \), Yuan Fang\( \dagger \), Guoying Gu, Huazhe Xu\( \ddagger \), Cewu Lu\( \ddagger \) (*equal contributions, \( \dagger \)equal contributions, \( \ddagger \)equal advising)
Robotics: Science and Systems (RSS), 2025
🔥Best Student Paper Finalist
🔥Best Paper @ Beyond P&P Workshop in ICRA 2025
project page / paper / arXiv / tweet / code / bibtex

We propose TactAR and Reactive Diffusion Policy (RDP). TactAR is a teleopration system that uses AR to provide tactile / force feedback. RDP is a slow-fast policy learning method which enables closed-loop tactile / force control via fast policy while maintaining the capability of modeling complex action distributions via slow policy.

GIF of DeformPAM Demos
click to view the proposed framework
DeformPAM Framework
DeformPAM: Data-Efficient Learning for Long-horizon Deformable Object Manipulation via Preference-based Action Alignment
Wendi Chen*, Han Xue*, Fangyuan Zhou, Yuan Fang, Cewu Lu (*equal contributions)
International Conference on Robotics and Automation (ICRA), 2025
🔥Best Paper Finalist @ RMDO Workshop in ICRA 2025
project page / paper / arXiv / tweet / code / video / bibtex

Inspired by RLHF, DeformPAM enhances learning efficiency and mitigates distribution shift in deformable object manipulation by selecting action through a preference-based implicit reward model.

Selected Awards and Honors

The website template is from Jon Barron.