Yongsheng Ou, Professor
Dalian University of Technology, China
Biography: Ou Yongsheng, professor in Dalian University of Technology. In 2011, he received the support of the 100 Talents Program of the Chinese Academy of Sciences, and was selected into the National "Ten Thousand Talents Program" in 2020. Experts of the 13th Five-Year Key Intelligent Robot Group of the Ministry of Science and Technology, key project evaluation expert of Natural Science Fund Committee, and evaluation expert of national Science and Technology Award plan. He has undertaken many national projects, such as key projects of the National Natural Science Foundation of China, national major research plan projects, and the 863 project of the Ministry of Science and Technology. With the interdisciplinary background of robotics and intelligent control, in the past five years, he has published more than 180 EI / SCI papers in the field of artificial intelligence and robotics. He has cited 5047 times and published more than 30 articles in JCR 1, Google H-index 55. He led the team to develop a number of core key technologies and realize the transformation. The low-cost navigation technology developed by him is used for the new product Xinluo sweeping robot developed by Shenzhen Silver Star Intelligent Company, and the humanoid service robot of Shenzhen UBTECH Technology Co., Ltd. His research won the 2016 Wu Wenjun Artificial Intelligence Technology Progress Award (ranked first), the 2018 Shenzhen Municipal Science and Technology Progress Award (ranked first), and the 2020 Guangdong Provincial Technology Invention Award (ranked first).
Speech Title: Research on the Complex Skill Learning Method of Embodied Intelligence Robots
Abstract: At present, there is still a big gap between robots and humans in terms of mechanism, perception and control, and it is difficult for robots to adapt to changing scenes or tasks. In contrast, humans are better at handling missing, contradictory, and vague information. Therefore, integrating with human beings and the environment is an important means to solve the problem of insufficient robot intelligence. For complex scenarios, in-depth exploration of robot theory and practical problems under the situation of "human-in-loop", and robot multi-mode perception, knowledge reasoning, accurate and compliant control and other methods are studied. The main research contents include: (1) human operation intention understanding and behavior prediction in dynamic unstructured environment;(2) efficient migration and generalization technology of robot collaboration skills driven by knowledge reasoning; and (3) Human-machine collaborative compliance force control based on multi-mode information fusion.
Ning Wang, Professor
Dalian Maritime University, China
Biography: Ning Wang received his B. Eng. degree in Marine Engineering and the Ph.D. degree in control theory and engineering from the Dalian Maritime University, Dalian, China in 2004 and 2009, respectively. From September 2008 to September 2009, he was financially supported by China Scholarship Council to work as a joint-training PhD student at the Nanyang Technological University (NTU), Singapore. In view of his significant research at NTU, he received the Excellent Government-funded Scholars and Students Award in 2009. From August 2014 to August 2015, he worked as a Visiting Scholar at the University of Texas at San Antonio. His research interests include self-learning autonomy, unmanned (marine) vehicles, and autonomous systems.
Dr. Wang was honored with the National High-level Talents Program for Young Top-notch Talents. He has published more than 200 SCI-indexed journal papers. He is a Fellow of IET, IMarEST, IIAV, and ISCM. He currently serves as Associate Editors of IEEE Transactions on Intelligent Transportation Systems, IEEE Signal Processing Letters, IEEE Systems Journal, Journal of Field Robotics, Ocean Engineering, International Journal of Robust and Nonlinear Control, ISA Transactions, Transactions of the Institute of Measurement and Control, and IET Electronics Letters. He was recognized as the Highly Cited Researcher by Clarivate Analytics, and the Highly Cited Chinese Researcher in Naval Architecture and Ocean Engineering by Elsevier.
Hejun Wu, Professor
Sun Yat-sen University, China
Biography: Hejun Wu, Professor, PhD Supervisor, School of Computer Science and Engineering, Sun Yat-sen University. Deputy Director, Guangdong Key Laboratory of Big Data Secretary-General, Working Committee, New Engineering Alliance of the Ministry of Education Senior Member, China Computer Federation (CCF) Chairman, Emergency Medicine Branch, Guangdong Medical Association. Professor Hejun Wu received his PhD degree from the Hong Kong University of Science and Technology in 2008. His current research interests include embodied AI, and cross-disciplinary applications of artificial intelligence in healthcare and transportation. He has undertaken four national-level projects and numerous provincial and municipal science and technology programs. He has published more than 60 academic papers in international journals and conferences, three of which have won Best Paper Awards. He also holds more than 20 authorized invention patents. In 2023, Professor Wu was honored as an "Outstanding Teacher in Computer Science among National Higher Education Institutions."
Speech Title: Generative Policy Optimization for Embodied Intelligence
Abstract: Generative policies enable robots to learn complex behaviors from demonstrations, but their effectiveness depends on representation quality, the ability to improve through interaction, and the computational cost of generating actions. This talk presents our recent findings on these challenges, focusing on diffusion policies and one-step generative policies for robot control.
We first investigate representation collapse in diffusion policies, where similar observations produce indistinguishable representations, obscuring subtle differences that are critical for precise manipulation. We show how dispersive regularization during pretraining improves representation learning and how subsequent online fine-tuning with proximal policy optimization improves task performance. We then examine the challenge of reducing action generation costs while maintaining policy quality. By combining native one-step generation, representation spreading, and online reinforcement learning, we develop policies that support fast inference and further improvement through environment feedback. Experiments on simulation benchmarks and a physical robot demonstrate progress in manipulation performance and control efficiency.
The talk connects these findings through a common perspective: pretraining should establish useful representations, policy optimization should improve behavior through interaction, and action generation should meet the demands of real-time execution. We conclude by discussing how these components work together to support the practical deployment of generative policies.