Yipeng Chen 陈祎芃

I’m a master’s student at Nanyang Technological University (opens in a new tab). Before that, I studied Physics at UCL (opens in a new tab). I’m currently working with Professor Na Zhao (opens in a new tab) and Dr. Xu (opens in a new tab), who are both excellent researchers at the Singapore University of Technology and Design (SUTD) (opens in a new tab).

I’ve worked on continual learning for radar perception, radar–camera fusion, and AI agents. Next, I want to work on world models: learning how the physical world changes and using that to plan actions.

Email / GitHub (opens in a new tab) / LinkedIn (opens in a new tab)

Research

I’m interested in continual learning, embodied intelligence, and world models.

Street scene from the RadarScenes dataset with cars, a van and pedestrians highlighted in colour.

Continual Learning for Automotive Radar Semantic Segmentation
Yipeng Chen, Ziwei Zhang, Jun Liu
IEEE Sensors Journal, 2026
paper

Teaching a radar segmentation model new object classes without forgetting the old ones. The previous model is kept frozen and guides the new one through knowledge distillation. Tested on RadarScenes (six classes, three incremental tasks) with four different network architectures.

Camera image with 3-D bounding boxes drawn around detected cars.

RC-HyFusion: Radar–Camera Hybrid Fusion for 3-D Object Detection in Autonomous Driving
Liguo Chen*, Yipeng Chen*, Hongsi Liu, Jie Zhang, Xiang Zhang, Chang Chen, Jun Liu, Weidong Chen
IEEE Transactions on Aerospace and Electronic Systems, vol. 62, pp. 15895–15909, 2026
paper

Combines camera images and radar for 3-D object detection. Image and bird’s-eye-view features are fused at several stages, with radar guiding where to sample. Evaluated on nuScenes. *Equal contribution.

Rendered view of a potted fern plant reconstructed from three photos.

Depth-Guided Sparse-View Gaussian Reconstruction
Yipeng Chen
Project, 2026
code

Reconstructing a 3D scene from only three photos. I added local depth-ranking supervision to D²GS, using relative depth from a frozen Depth Anything V2 model. On two LLFF scenes, averaged over three seeds, LPIPS dropped by 3.71% on trex and 6.15% on fern compared with my D²GS baseline (lower is better). Some views still show visible ghosting.

Experience

ByteDance, AI Agent Intern, July 2026 – present
Making long AI agent conversations faster and more reliable: context compression, caching token counts, and preparing conversation summaries in the background. Also checking how well automated scoring of presentations matches human ratings.