Frontier of Artificial Network
A Series of Invited Talks @ FAN Group, CityU

Biography: Zeyu Zhang is a Senior Researcher at the Beijing Institute for General Artificial Intelligence (BIGAI) and serves as a doctoral co-advisor in the Tong Program in collaboration with Peking University and Huazhong University of Science and Technology. He received his Ph.D. in Computer Science from the University of California, Los Angeles (UCLA), under the supervision of Professor Song-Chun Zhu. He is a recipient of the Beijing Overseas Young Talent Award. He posits that human intelligence is grounded in fundamental representations and a unified cognitive architecture. His research seeks to uncover and computationally realize this architecture to enable higher levels of autonomy and intelligence in robotics by bridging how robots perceive the world, reason over abstract structures, and act in complex environments. His work has been published in leading journals and conferences in the field, including IJCV, TPAMI, Engineering, T-RO, R-AL, ICRA, IROS, and ICCV, with over 20 publications to date. He has long served as a reviewer for top-tier international conferences and journals.

Abstract: In this talk, I will present a series of our works on robot manipulation in 3D scenes that aim to discover a unified representation bridging perception and action. I will begin with how we model the embodiment of the robot and the manipulated object through a unified kinematic representation using a Virtual Kinematic Chain. I will then introduce our function-oriented scene reconstruction that enables robots to perceive environments as actionable structures for manipulation. Next, I will show how we jointly infer perception, task plans, and motion generation in an object cutting example by combining neural signals with symbolic recursive reasoning. Finally, I will present our recent work on mobile manipulation that generalizes these ideas to coordinated whole-body behaviors across diverse 3D scenes.

Biography: Guibo Luo is an Assistant Professor and Ph.D. Supervisor at the School of Electronic and Computer Engineering, Peking University. He received his Ph.D. degree from Peking University, and subsequently conducted postdoctoral research at Harvard Medical School and Massachusetts General Hospital. His research focuses on privacy-preserving computation and foundation model training, with an emphasis on discovering and quantifying scientific patterns from heterogeneous real-world data while ensuring privacy and security in collaborative settings. Recently, he has been investigating efficient multi-institutional collaborative intelligence without exposing private data. To this end, he has developed a systematic research framework that connects (i) the creation of real-world multi-center datasets, distribution-shift characterization, and benchmark evaluations, (ii) communication-efficient privacy-preserving learning and secure computation paradigms, and (iii) privacy-preserving collaboration between foundation models and lightweight edge models under stringent communication and compute constraints. His work further emphasizes reliability and accountability in real deployments, and has led to practical deployments in healthcare, embedded systems, and embodied intelligence. He has published more than 80 papers in leading journals and conferences, including IEEE TPAMI, IEEE JBHI, Radiology: AI, IEEE TCSVT, IEEE TCDS, Journal of Digital Imaging, CVPR, ECCV, AAAI, KDD, ICSE, IJCAI, and MICCAI. He also serves as a reviewer for journals and conferences such as IEEE TIP, IEEE TNNLS, Knowledge-Based Systems, CVPR, ICCV, ICLR, NeurIPS, KDD, and MICCAI.

Abstract: Real-world clinical settings, medical imaging data is distributed across institutions and inherently non-IID, governed by strict privacy and regulatory constraints. This makes centralized training difficult to deploy, while conventional multi-round federated learning (FL) introduces substantial communication cost and a broader privacy attack surface. We provide an end-to-end roadmap from evaluation → training → deployment: (1) we establish reproducible benchmarks with multi-center, multi-modal data foundations and quantitative heterogeneity metrics; (2) we propose low-interaction / one-shot generative collaboration, reframing aggregation from “sharing gradients/parameters” to “sharing generative, controllable knowledge,” thereby reducing communication and exposure while approaching centralized performance; and (3) we develop bi-directional foundation–edge collaboration to transfer capability to lightweight models and improve foundation-model adaptation to edge distributions, enabling long-term, compliant, and deployable multi-party training in practice.