Qianjun Xia

Research

Physical intelligence: reading physics out of the world, and writing it back into machines.

My work sits between perception and physics. A video shows what a body did; a simulator needs to know what that body is — its stiffness, its mass, where its actuators sit, how its parts connect. I build models that close that gap, so that a machine can act on the physics rather than only observe it.

I am a Ph.D. student in the PhysAI Lab at UBC, supervised by Prof. Peter Yichen Chen.

Directions

World modeling from real interaction

Turning recordings of real robot-object interaction into simulatable digital twins, so that policies can be trained and evaluated against something that behaves like the world did. Vision-language agents recover geometry, object state and physical parameters without per-scene hand tuning — across rigid manipulation, deformable interaction and humanoid motion.

Soft-body parameter estimation

Deformable bodies have properties that vary across space, deform far past the linear regime, and hide their actuators where no camera can see them. I treat system identification as a translation problem: from a short grayscale video to a voxel-level parameter field a simulator can replay.

Modeling and control of slender structures

Elastic-rod models for magnetically actuated guidewires — bending, torsion, stretching and shear under a spatially varying field — and the control that turns those models into autonomous navigation.


Media and write-ups are on the projects page; papers are listed under publications.