Jing-Wen (Joycelyn) Chen

I am an incoming PhD student in the Department of Electrical Engineering at National Taiwan University, where I will join the Robot Learning Lab, advised by Shao-Hua Sun. My research focuses on building general-purpose robotic systems that can understand the physical world and perform complex, long-horizon tasks in everyday environments — particularly leveraging large-scale foundation models to advance robot perception, reasoning, and decision-making toward capable, adaptable embodied AI.

I completed my MSc in Computer Science at the University of British Columbia (Okanagan), where I was advised by Alex S. Hill and Mohamed S. Shehata and worked on computer vision for scientific discovery — most recently 3D transformer segmentation of supernova-driven superbubbles in the interstellar medium, which grew into a first-author paper in The Astrophysical Journal. Before that, I earned my BSc in Computer Science at National Chi Nan University, Taiwan, where I worked on multi-object detection and tracking and facial recognition systems.

Email  /  Google Scholar  /  GitHub  /  LinkedIn  /  CV

Jing-Wen (Joycelyn) Chen
National Taiwan University

NTU

University of British Columbia

UBC

National Yang Ming Chiao Tung University

NYCU

National Chi Nan University

NCNU

News

Mar 2026 Our superbubble segmentation paper is published in The Astrophysical Journal 🎉
Nov 2025 Released AstroUNETR, the open-source codebase behind the paper.
2025 Graduated with an MSc in Computer Science from UBC Okanagan 🎓
2025 Astro-VOS presented at the International Telecommunications Conference (ITC-Egypt 2025).
Jan 2023 Our multi-vehicle tracking paper appeared at IEEE ICCE 2023.

Research

I am interested in developing physically intelligent robot systems, with a focus on robot vision, perception, manipulation, and robot learning. My work emphasizes interactive perception and utilizing VLMs to enable general-purpose robots to perform long-horizon tasks.

Superbubble segmentation
Segmenting Superbubbles in a Simulated Multiphase Interstellar Medium Using Computer Vision
Jing-Wen Chen, Alex S. Hill, Anna Ordog, Rebecca A. Booth, Mohamed S. Shehata
The Astrophysical Journal, 2026
journal / arXiv / code

We use 3D transformer models to segment and track superbubbles in magnetohydrodynamic simulations of the supernova-driven interstellar medium, replacing hand-tuned rule-based detection. The resulting masks let us follow a single bubble's growth, energy retention, and interaction with the surrounding medium across its full evolution.

Video object segmentation for supernova tracking
Astro-VOS: Tracking Supernova Evolution Using Video Object Segmentation
Jing-Wen Chen, Mohamed S. Shehata, Alex S. Hill
International Telecommunications Conference (ITC-Egypt), 2025
code

Treats a time series of simulation slices as video and adapts video object segmentation to propagate a single annotated frame across the whole sequence, making supernova remnant tracking possible with minimal manual labelling.

3D volumetric segmentation
Automating Superbubble 3D Segmentation in a Multiphase Interstellar Medium Using Computer Vision
Jing-Wen Chen
MSc Thesis, University of British Columbia, 2025
UBC Library / pdf

My master's thesis: an end-to-end pipeline that takes raw MHD simulation volumes to 3D superbubble masks and time-resolved measurements, including the data preparation, model design, and evaluation protocol the later publications build on.

Multi-vehicle detection and tracking
Disentanglement-Based Multi-Vehicle Detection and Tracking for Gate-Free Parking Lot Management
Chia-Hao Cheng, Jing-Wen Chen, Wei-Hao Su, Chung-Chi Huang
IEEE International Conference on Consumer Electronics (ICCE), 2023

A disentanglement-based approach to detecting and tracking multiple vehicles in parking lot surveillance video, aimed at gate-free entry and exit management.

Experience

Team Lead — Active Exploration of Hidden Physical Properties in Unstructured Environments via Code Policies
Robot Learning Lab, National Taiwan University
Oct 2025 – Present

Developing a curiosity-driven robot exploration framework to efficiently infer hidden physical properties in unseen, unstructured environments for robust task execution.

Team Lead — Supernovae 3D Segmentation and Tracking Using Computer Vision
Department of Computer Science, Mathematics, Physics and Statistics, University of British Columbia, Okanagan
Sep 2022 – May 2025

Engineered a three-fold data pipeline to iteratively construct a superbubble dataset, enabling precise segmentation and tracking of superbubble morphology and evolution via video object segmentation (Astro-VOS) and 3D instance segmentation (Astro-UNETR). Applied these methods to study superbubble evolution in 3D magnetohydrodynamic simulations, refining our understanding of the solar neighborhood's history; presented findings at CASCA 2023 (Penticton), ECCV 2024 (Milan), and CVPR 2025 (Nashville). Published the Astro-UNETR data loader and training scripts as an open-source PyPI package, enabling third-party use in other astro-CV projects.

Disentanglement-Based Multi-Vehicle Tracking for Gate-Free Parking Lot Management
National Yang Ming Chiao Tung University, Taiwan
Feb 2022 – Jul 2022

Enhanced multi-object tracking for parking-lot applications by coupling the YOLOv5 detection model with the DeepSORT tracking algorithm, and implemented GMM-based background subtraction to reduce ID switching during occlusions, improving tracking accuracy by 10%. Trained, tested, and deployed the system on real-world parking-lot video, contributing to a paper published at IEEE ICCE 2023.

Projects

Selected code from my research and side work. More on GitHub.

Beyond Research

Mindfulness

A steady meditation practice — how I stay grounded and clear-headed through a fast-moving field.

Reading

Cognitive science, and the philosophical questions that come with building intelligent systems.

Outdoors

Hiking and photography — the Okanagan is a good place for both.

Making things

Chess, cooking my way through other people's cuisines, and the occasional hackathon.