Research Scientist NVIDIA Research
I am a Research Scientist at NVIDIA Research, developing generalizable and agentic systems for robot learning and control. My work connects world–action modeling, embodied reasoning, and action to help robots understand physical dynamics, anticipate the consequences of their decisions, and adapt their behavior across tasks and environments.
I received my Ph.D. from National Taiwan University in July 2023, advised by Prof. Yu-Chiang Frank Wang. Previously, I was a research intern at NVIDIA Research (Feb.–Aug. 2023) and a Ph.D. program researcher at ASUS AICS (Sep. 2020–Oct. 2022), working on model personalization, vision-language models, and visual transfer learning.
I received my bachelor’s degree in Electrical Engineering from National Taiwan University in 2018.
My research focuses on generalizable and agentic robot intelligence. I study world modeling and world–action modeling to understand physical dynamics, predict action outcomes, and reason through imagined futures. By connecting these predictive models with agentic robotics, I aim to build robots that plan over long horizons, learn from interaction, and adapt to unfamiliar tasks and environments.
A motion-centric video encoder that separates appearance from temporal dynamics using compact motion tokens. Temporal Transfer layers and Diff Compression improve motion-sensitive recognition with about half the encoder FLOPs of VideoMAE, V-JEPA 2, and DisMo.
Scalable online 3D reconstruction from long videos using multi-reference relative pose queries and online pose-graph optimization, with only about 1% additional parameters on a frozen backbone.
A slow–fast framework for real-time streaming video anomaly understanding. Lightweight detection preserves anomalous evidence in persistent memory and triggers heavyweight reasoning only for suspicious events.
Compressing textual chain-of-thought into compact, verbalizable latents for efficient vision-language-action reasoning, long-horizon planning, and real-time robotic control.
An open vision-language-action foundation model for generalist humanoid robots, connecting language and visual observations to manipulation skills across embodiments.
A reasoning VLA framework that thinks before acting. Reinforced visual latent planning enables long-horizon manipulation, few-shot adaptation, and self-correction.
Reconstructing 3D scenes from casual, long, unposed videos and rendering novel views without camera calibration.
Customizing multiple subjects and their motion in text-to-video generation through disentangled appearance–motion learning and spatial-temporal composition.
Reliable concept erasing for text-to-image diffusion models using lightweight erasers.
Selective knowledge transfer from two teachers for continual learning with vision-language models.
Reinforced rationale prompting for natural-language explanations in visual question answering.
Language-guided transformer learning for multi-label image classification in a federated setting.
Client-specific prompt generation for efficient model personalization in federated learning.
Adversarial teacher–student representation learning for generalization to unseen visual domains.
ICLR 2027, ICRA 2027, CoRL 2026, NeurIPS 2026, ECCV 2026, CVPR 2026, ICLR 2026, AAAI 2026, WACV 2026, BMVC 2026, NeurIPS 2025, ICCV 2025, ICML 2025, CVPR 2025, ICLR 2025, ICLR 2025 WS SCOPE, AAAI 2025, ACM MM 2025, NeurIPS 2024, ECCV 2024, ICML 2024, CVPR 2024, AAAI 2024, ACCV 2024, ICIP 2024, NeurIPS 2023, ICCV 2023, CVPR 2023, AAAI 2023, WACV 2023, ICIP 2023, ACCV 2022, CVPR 2022, AAAI 2022, WACV 2022, AAAI 2021, ICIP 2020, AAAI 2020
Visitor statistics & locations ↗