My research operates at the intersection of egocentric vision and embodied AI. I aim to incorporate human common sense and behavioral patterns into machine learning to build embodied agents capable of learning structured models from human-world interactions and generalizing to diverse real-world settings. I tackle these challenges through three core lenses:

• Act via Prediction: Determining the optimal abstractions for representing past experiences from partial human activity data, modeling underlying task structures, and leveraging multimodal integration to boost downstream performance.
• Interact via 3D Reconstruction: Capturing the nuances of dynamic human interactions from an egocentric viewpoint to reconstruct detailed 3D physical models for robotic manipulation.
• Collaborate via Understanding: Creating intelligent robots that seamlessly interface with humans by decoding underlying intent and estimating human goals from multimodal observational cues.

I support slow science.

---

---

---

Publication

---

---

Personal

I was born and grew up in a small city (with only a bit more than 1 million people) in China. Five years of my life was spent in Shanghai for study and I had a wonderful time living in Berlin while pursuing my PhD. I enjoy travelling and taking pictures on the way. I also enjoy playing guitar and have been in a rock band during my studies in Shanghai. Since 2014 I started playing jazz, and discovered the powerful secret of playing jazz -- no matter how it sounds, you can always entitle it as improvisation. My favorite jazz guitarists are Pat Metheny and Bill Frisell (their music can easily give me goosebumps). The rest of time is spent cooking, learning to play drum-sets and tap dancing, a perfect combination of jazz and rhythm ;).