Date, time, and room will be added once confirmed.
The talk presents a world model that integrates generation, navigation, and control, addressing limitations in previous models through innovations in representation and physical perception. A framework combining flow matching and DPO reinforcement learning enhances continuous generation and physical understanding, achieving state-of-the-art results in open-source video generation. The researchers also developed a comprehensive benchmark system to evaluate physical perception across multiple domains, and demonstrated the integration of a world model with robotic systems for real-world applications.