Abstract: Understanding, navigating, and interact with the 3D physical real world has long been a central challenge in the development of artificial intelligence. This talk presents a fundamental framework for generating explorable 3D worlds that enable AI systems to reason more deeply and act more intelligently. By simulating outcomes and refining internal beliefs, these generative worlds foster informed embodiment, where agents learn to plan, interact, and adapt within the physical world. I further illustrate how this paradigm applies to real-world domains such as robotics and medical treatment planning.
Bio: Jieneng Chen is a Ph.D. candidate in Computer Science at Johns Hopkins University, advised by Alan Yuille and working closely with Rama Chellappa. He is interested in building scalable multimodal models to understand and interact with the physical world. His first-authored neural architecture TransUNet has been cited more than 8000 times, with huge impacts on image understanding and diffusion. His paper was ranked among the top 15 most cited AI papers of 2021 and the top 3 most cited papers from ECCV over the past five years. He has received honors including the Siebel Scholar Award, the DAAD AInet Fellowship, a nomination for the Schmidt Science Fellowship, the MICCAI Doctoral Thesis Runner-up Award, the MICCAI Best Paper Runner-up Award, CVPR Highlights, the KDD 2025 CCC Best Paper Award, and the 2025 NVIDIA Grant Award. His publications have garnered over 17,000 citations.