I am a Ph.D. candidate in the Department of Automation at the University of Science and Technology of China, expected to graduate in June 2027. I am advised by Associate Professor Yang Cao (曹洋) and Professor Zheng-Jun Zha (查正军). My research experience includes embodied intelligence, multimodal learning, and out-of-distribution detection.
I am currently an Applied Research Intern at Robbyant（蚂蚁灵波科技）, supervised by Kecheng Zheng (郑可成), where I contribute to the research and development of vision-language-action (VLA) foundation models for embodied intelligence. My current work involves large-scale VLA models for real-world robotic manipulation, including model architecture construction, training infrastructure, systematic ablations, and large-scale pretraining.
Deeply involved in the research and development of LingBot-VLA and LingBot-VLA 2.0, large-scale vision-language-action foundation models for real-world robotic manipulation. My contributions include model architecture construction, codebase infrastructure design, large-scale pretraining, and the implementation and optimization of an MoE-based action expert architecture for LingBot-VLA 2.0.
Conducted research on multimodal learning, including CLIP pre-training with long captions, VLM training with generative visual representations, and the development of a VLM benchmark for evaluating the quality of detailed image captions.
Started my master's studies in the Department of Automation in September 2022 and transitioned to the Ph.D. stage in September 2024 through an integrated M.S.–Ph.D. program. My early research focused on out-of-distribution detection.
My broader research interests include embodied intelligence, multimodal learning, and large-scale pretraining.
Student first author.
LingBot-VLA 2.0 is a practical vision-language-action foundation model designed to move from large-scale pretraining toward reliable real-world robot applications. As a student first author, I focused on model architecture construction, codebase infrastructure design, and the implementation and optimization of an MoE-based action expert architecture.
Co-first author; leading student contributor.
LingBot-VLA is a large-scale vision-language-action foundation model for real-world robotic manipulation. I contributed as a co-first author and leading student contributor, focusing on model architecture construction, systematic ablations, codebase infrastructure design, and large-scale pretraining.
A practical VLA foundation model for real-world robot applications, with expanded cross-embodiment capabilities and an MoE-based action expert.
A large-scale vision-language-action foundation model for real-world robotic manipulation.
Email: lufan@mail.ustc.edu.cn / lufanzwj@gmail.com WeChat: luDZ310711