Building Interactable 3D Spaces from Imperfect ObservationsBuilding Interactable 3D Spaces from Imperfect ObservationsBuilding Interactable 3D Spaces from Imperfect ObservationsBuilding Interactable 3D Spaces from Imperfect ObservationsBuilding Inter

Release Time：2026-07-26Number of visits：10

Speaker:   Angela Dai, Associate Professor of Technical University of Munich

Time:       10:20, July. 29th

Location: 1A200, SIST

Host:        Prof. Lan Xu

Abstract:

Our observations of the world are always incomplete. Images and range sensors capture only fragments of real environments, leaving significant portions of scenes unobserved. Yet many applications, ranging from digital content creation to embodied AI systems that must reason and act in the physical world, require complete, structured, and interactable models of 3D environments. This talk explores approaches for learning generative 3D world models from imperfect observations, capturing not only what a scene looks like, but also how it can be interacted with.  We examine how self-supervised formulations for scene completion can leverage visibility reasoning together with modern generative techniques such as flow matching to infer plausible geometry in unobserved regions of real scans. Ultimately, usable 3D worlds must support interaction. To this end, we leverage image generative model priors to synthesize plausible articulations and human interactions within 3D scenes. Together, these ideas move toward generative 3D world models that can be complete, realistic, and interacted with.

Bio:

Angela Dai is an Associate Professor at the Technical University of Munich where she leads the 3D AI Lab. Angela's research aims to enable machines to understand, model, and generate real-world 3D environments. She focuses on enabling the creation of rich, semantically grounded, and interactable 3D worlds that allow machines not only to perceive physical spaces, but to reason about them and act within them. She received her PhD in computer science from Stanford in 2018, advised by Pat Hanrahan, and her BSE in computer science from Princeton in 2013. Her research has been recognized through an ECVA Young Researcher Award, ERC Starting Grant, Eurographics Young Researcher Award, German Pattern Recognition Award, Google Research Scholar Award, and an ACM SIGGRAPH Outstanding Doctoral Dissertation Honorable Mention. She has also served as Program Chair for Eurographics 2025 and CVPR 2026.