AI研讨会 | World Models for Spatial Intelligence: From Policy Learning to Data Generation

🎤 Speaker

Dr. JIANG Li

香港中文大学（深圳）数据科学学院助理教授，校长青年学者

📰 Title

World Models for Spatial Intelligence: From Policy Learning to Data Generation

⏰ Time

11:00 -11:50 am, Beijing Time

📅 Date

2 Sep. 2026, Wed.

📍 Venue

Lecture Hall B, GZ Campus

🤝 Online Zoom link

https://hkust-gz-edu-cn.zoom.us/j/97777467473?pwd=PxC8FbxXlMDg1MYTObQY2aBadPqqoa.1

Meeting ID: 977 7746 7473

Password: ait

World models learn how the world evolves, providing an essential foundation for intelligent agents to understand, predict, and interact with their environments. In the context of spatial intelligence, world models can serve both as world representation models that facilitate policy learning for autonomous driving and embodied intelligence, and as world generators that support scene reconstruction, training data generation, and closed-loop simulation. This talk will present some explorations along these two directions. First, from the perspective of policy learning, this talk will discuss how latent-space world modeling, foundation models for video generation, and future spatial prediction can enhance spatiotemporal understanding, dynamics forecasting, and decision-making in autonomous driving and robotic manipulation. Second, from the perspective of data and scene generation, this talk will introduce some recent research on long-horizon interactive video generation and spatiotemporal memory modeling. Through these studies, this talk will illustrate how world models can bridge policy learning and generation, and discuss their future potential for advancing spatial intelligence.

蒋理，香港中文大学（深圳）数据科学学院助理教授，校长青年学者，与深圳河套学院双聘。于2021年获得香港中文大学博士学位；其后在马克斯·普朗克研究所担任博士后研究员。其研究聚焦于计算机视觉与人工智能，具体方向包括三维感知与空间智能、世界模型、自动驾驶、具身智能、表征学习以及多模态学习等。其研究成果发表在CVPR、ICCV、NeurIPS、和TPAMI等顶级会议和期刊中，并多次获选口头报告与亮点论文，谷歌学术引用量超过18,000次，并担任顶会审稿人与领域主席。其关于自动驾驶运动预测的研究连续三次获得CVPR Waymo Open Dataset Motion Prediction Challenge全球冠军。多次入选由斯坦福大学与爱思唯尔联合发布的全球前2%顶尖科学家年度影响力榜单，并获得国家级青年人才项目资助。