Hang Dai 戴航

I am an undergraduate student in Information and Computing Science at Peking University (graduating June 2027). I want robots to learn from 4D worlds that are physically correct, not just visually realistic, so I treat geometry, articulation, and dynamics as first-class targets when reconstructing or generating the world.

From the real world, I reconstruct simulation-ready articulated objects from a free-moving monocular RGB-D video (FreeArtGS, CVPR 2026) and build physics-aware digital twins for zero-shot real-to-sim-to-real policy transfer (TwinAligner). From imagination, I build agentic systems that generate minute-long, physically plausible 4D scenes from text (DyCraft). My goal is to make such worlds a scalable source of training data, simulation, and evaluation for robot learning. Building DyCraft also showed me that current multimodal LLMs are still poor at understanding 4D information from video, a gap I am interested in studying.