Keynote Speakers
Xin Jin
Keynote Title
Build and Leverage World Model in the Latent Representation Space
Abstract
World model is becoming a core technological paradigm connecting the digital and physical worlds. Their value lies in endowing intelligent agents with the ability to understand, reason, and interact in space, enabling machines to truly "enter" the physical world. This report will systematically outline the complete technological chain of spatial intelligence from 3D perception and world model construction to embodied task execution, and highlight the team's representative work in this field: the UniScene unified generation framework for autonomous driving and robotic scenarios; the DreamVLA model, which introduces world models into the VLA architecture to enhance long-range reasoning and action prediction capabilities through "world representation embedding"; benchmark datasets such as InterVLA & SceneScribe-1M, filling the gaps in embodied interaction data and spatial geometric video scene data from the first-person perspective; the DeFI framework, which solves the forward-backward dynamics model learning process; and VLA-JEPA, which endows VLA models with efficient reasoning capabilities from the latent space level, etc.
Biography
Xin Jin received his PhD degree from the University of Science and Technology of China (USTC). He is currently an Assistant Professor at the Eastern Institute of Technology, Ningbo, China. He was a visiting scholar at the Learning and Vision Lab of the National University of Singapore (NUS) from 2021 to 2022.
He has received several academic awards, including 2024 VSPC Rising Star, 2026 MSA TC Best Paper Award, the ACM SIGAI China Doctoral Dissertation Award, the Chinese Academy of Sciences President's Scholarship—Special Award, and the “Rising Star of Microsoft Research Asia” title. He has published over 40 papers in top-tier conferences and journals, including CVPR, ICCV, ECCV, NeurIPS, ACM MM, ICIP, ICME, ISCAS, VCIP, TIP, TMM, TCSVT, and Pattern Recognition, with nearly 9,000 Google Scholar citations. He was recognized as one of the top 2% of scientists worldwide by Stanford University in 2024&2025.
He has actively contributed to international and domestic standardization efforts in image/video compression, such as MPEG VCM and China DCM (Data Coding for Machine). He has organized multiple tutorials and special sessions at major conferences, including CVPR 2024, ECCV 2024, VCIP 2024, and ICME 2025, focusing on generative AI, diffusion models, and visual disentanglement. He also served as a session chair at ICIP 2024. In CVPR 2024, ECCV 2024, ICCV 2025, he organized tutorial sessions and workshop related to “Visual Representation Learning of Disentanglement and Compositionality”. In VCIP 2024, he also built up a special session about “Generative AI for Image/Video Coding”.
Patrick Le Callet
Keynote Title
Should We Cross This Frontier? Embodied AI and Human Sustainability
Abstract
The world is aging. As demographic shifts reshape our societies, maintaining health, autonomy, and quality of life has become one of the defining challenges of the coming decades. Artificial intelligence is increasingly presented as a technological response, with embodied AI expected to assist, collaborate with, and augment human capabilities in the physical world.
Recent advances are transforming embodied AI. Beyond classical imitation learning, new paradigms allow robots to learn through interaction and to autonomously acquire increasingly sophisticated behaviors. These developments promise more adaptive and resilient machines, capable of operating in complex, real-world environments. But they also invite a more fundamental question: what kind of embodied intelligence are we actually building, and for what purpose?
True embodiment cannot be reduced to placing a foundation model inside a robot. Following Gibson’s ecological theory of perception, intelligence emerges from the continuous coupling between an organism and its environment through action. Similarly, the Free Energy Principle describes cognition as an active process of self-maintenance, shaped by the perception–action loop and by the finitude of living systems, their limited energy, vulnerability, and need to survive. These principles suggest that intelligence is not simply computed but enacted through the dynamic relationship between body, environment, and self.
Yet today’s embodied AI remains, by design, an imitation of these processes. The more convincing this imitation becomes, the greater the temptation to attribute to machines properties such as agency, understanding, consciousness, or a genuine self. This may be the real frontier of embodied AI. The challenge is not only to develop increasingly capable systems, but also to preserve the conceptual distinction between simulating the dynamics of life and reproducing life itself. If embodied AI is to support human sustainability, we must ask not only what we can build, but also which frontier we should choose not to cross.
Biography
Prof Patrick Le Callet (Fellow, IEEE) received the M.Sc. and Ph.D. degrees in image processing from the Ecole Polytechnique de l’Université de Nantes. He was an Assistant Professor and a full-time Lecturer with the Department of Electrical Engineering, Technical Institute of the University of Nantes, from 1997 to 1999 and from 1999 to 2003, respectively. He led the Image and Video Communication Laboratory, CNRS IRCCyN, from 2006 to 2016. Since 2015, he has been the Scientific Director of the cluster Ouest Industries Creatives, a five-year program gathering over ten institutions (including three universities). He was one of the five members of the Steering Board of CNRS from 2013 to 2016. Since 2017, he has been one of the seven members of the Steering Board of the CNRS LS2N Laboratory (450 researchers) as a Representative of the Polytech Nantes. He is mostly involved in research dealing with the application of human vision modeling in image and video processing.
