I am a PhD student in Intelligent Transportation at The Hong Kong University of Science and Technology (Guangzhou), where I work with Prof. Haoang Li in the IRPN Lab and am co-supervised by Prof. Ying-Cong Chen. My research focuses on foundation models for embodied intelligence, with current interests in Video-Action Models, World Models, and Video Generation.
Before starting my PhD, I received an M.Eng. in Mechanical Engineering and a second bachelor's degree in Computer Science and Technology from Xi'an Jiaotong University. I am interested in building models that can understand dynamics, predict future interactions, and support downstream reasoning and decision making for robotic systems.
arXiv Technical Report, 2026
An action-centered World Action Model that uses future visual dynamics during training while enabling efficient action-only inference for real-time robot control.
ECCV 2026
A shortcut video-action model that self-distills geometric and semantic foresight from diffusion-based generation to enable efficient and precise robot manipulation.
Preprint · under review at IEEE Transactions on Image Processing
A 4D scene generation framework from a single image and text prompt that combines scene extension and dynamics synthesis for contextually aligned and realistic physical interactions.
arXiv 2026
A post-training framework that bridges offline supervised fine-tuning and online reinforcement learning through on-policy distillation for vision-language-action models.
arXiv 2026
A dual-stream reasoning framework that couples visual and linguistic chain-of-thought with parallel inference for fine-grained robotic manipulation.
arXiv 2026
A large-scale robotic memory benchmark with 26 long-horizon tasks averaging over 1,000 steps, most of them memory-dependent, together with PrediMem, a dual-system VLA whose VLM planner manages a predictive memory bank.
ICRA 2026
A comprehensive benchmark and lightweight baseline that study how to make vision-language-action models more practical across diverse robot embodiments and real-world settings.
CVPR 2026
A scalable structure- and contact-aware representation for generating realistic hand-object interaction videos that generalize to open-world scenarios.
AAAI 2026 Outstanding Paper Award
A reconstructive vision-language-action model that improves robot perception by reconstructing task-relevant visual regions for downstream manipulation.
arXiv 2025
A visual chain-of-thought framework for motion reasoning in VLAs that predicts future dynamics before generating the final action.
Pacific Graphics 2024
A multimodal diffusion framework that uses gaze to improve stochastic human motion prediction.
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024
Photometric bundle adjustment with material and illumination awareness for challenging non-Lambertian scenes.
Reliability Engineering & System Safety, 2023
A graph-based anomaly detection framework designed for complex equipment under limited-data settings.
ISA Transactions, 2023
An unsupervised anomaly detection model that combines memory augmentation and multi-scale skip connections for rocket engine monitoring.
Journal of Manufacturing Systems, 2022
A graph-based framework that imputes missing sensor readings and performs anomaly detection on incomplete multi-sensor equipment data.
PhD in Intelligent Transportation, The Hong Kong University of Science and Technology (Guangzhou)
M.Eng. in Mechanical Engineering, Xi'an Jiaotong University
B.Eng. in Mechanical Engineering, Xi'an Jiaotong University
Second Bachelor's Degree in Computer Science and Technology, Xi'an Jiaotong University
Internship, Institute for Visualisation and Interactive Systems, University of Stuttgart
Remote Internship, Computer Vision Group, Technical University of Munich
National Scholarship, Ministry of Education, China
National First Prize, Chinese University Students Mechanical Innovation Competition
Recently, I have been enjoying rock climbing, especially bouldering at V2-V3 and routes around 5.10a-5.10b.