Our research focuses on human-centric physical AI, aiming to develop intelligent systems that understand humans, learn from human experiences, and safely interact with humans in the physical world. We advance fundamental methods for modeling, understanding, and generating human behaviors and their interactions with the environment. Our long-term vision is to enable intelligent agents to acquire generalizable knowledge from human observations, reason about the physical world, and translate this knowledge into reliable and adaptive behavior in real-world environments. We pursue this vision through three complementary research directions:
- (Representation) Human Motion and Interaction Representation Learning: How can machines build comprehensive and physically grounded representations of humans and their interactions?
- (Understanding) Video Understanding and Generation: How can machines understand the semantics and physics of videos and generate videos that reflect this understanding?
- (Interaction) Human-centric Embodied AI: How can AI systems leverage human knowledge to interact with physical world intelligently and safely?
-
Human Motion and Interaction Representation Learning
Learn a precise, human-centric motion representation from noisy videos utilizing self-supervised learning and biomechanics modeling.
-
Video Understanding and Generation
How can machines understand the semantics and physics of videos and generate videos that reflect this understanding?
-
Human-centric Embodied Intelligence
Develop embodied AI systems that leverages human knowledge to interact with the physical world intelligently and safely.
