I study how to reconstruct humans and animals in 4D and understand how they move and interact—with each other and their environments. My research connects body assets, motion, and data to visual reconstruction and controllable generation, working toward interaction-aware models of humans and animals.
Toward human–animal interaction understanding.Jointly reason about bodies, motion, contact, and the scene—not isolated subjects.
FIG. 01 A research program, not a completed pipeline. Existing reconstruction work anchors the map; shared data, feed-forward models, generation, and human–animal interaction are directions to develop. Select a direction to explore.
Let each result become a foundation.
02 / The research program
Not a collection of unrelated papers. A proposed research loop in which assets and recovered motion support data, data supports reconstruction, and learned structure supports generation.
01 / Build
Assets
Reusable, animatable animal bodies.
02 / Recover
Motion
Articulated movement from visual observations.
03 / Combine
Data
Pair bodies and motion with rendered observations.
04 / Learn
Reconstruction
Feed-forward recovery of unseen bodies and motion.
05 / Create
Generation
Controllable, language-aligned motion and scenes.
The scientific bridge
Humans and animals share questions about articulation, motion, and contact, but not identical anatomy. The long-term goal is to reconstruct both and understand their relationships within a shared scene.
The evidence gate
Recovered motion is an estimate, not real-world ground truth. Validate asset–motion compatibility first, then test whether the resulting data improves generalization on independent real observations.
The research agenda
03 / Research directions
Each direction starts from a concrete limitation and asks what a better representation would let us do.
I / AnimalAgentPLANNED CVPR SUBMISSION
Frontier AI agents for animatable animal assets.
AnimalAgent investigates how frontier AI agents can be used to create animatable animal assets. This is an ongoing research project with a planned submission to CVPR. Within the broader program, the resulting assets could be paired with recovered motion to support data creation, reconstruction, and generation.
Questions to investigate
What makes an animal asset useful beyond its original reconstruction?
Can recovered motion transfer without breaking articulation, proportions, or contact?
Which asset variations provide useful supervision for unseen animals?
What would count as progress?
One asset and one recovered motion that combine reliably enough to generate useful training examples. Scaling the asset library comes after validating that connection.
Not just where an animal is. How it moves through the world.
WildAni4D connects monocular video to temporally coherent animal meshes, articulated motion, and world-space trajectories. The broader question is how such observations can support reusable motion representations and controllable generation.
Questions to investigate
How do we separate body motion from camera motion and ambiguous observations?
Which motion structure transfers across animals with different proportions?
How can language and visual priors guide motion without losing anatomy or temporal consistency?
What would count as progress?
Motion that stays consistent over time and supports a new camera, body, or condition. Reconstruction fidelity, physical plausibility, and downstream usefulness should be evaluated separately.
III / Human Motion UnderstandingRELATED WORK + ONGOING
Understand the movement. Preserve the interaction.
Human motion is more than a sequence of poses. I connect controllable avatars, pose analysis, and ongoing multi-dancer generation research with a broader question: how can we recover and understand human movement in relation to other people, animals, and the scene?
Questions to investigate
How can body identity and motion remain consistent under occlusion?
What changes when an action is defined by a human–human or human–animal interaction?
Which human motion priors transfer to animals—and which assumptions break?
What would count as progress?
Temporally stable bodies and coherent interactions under difficult observations. Related avatar and try-on publications are foundations, not claims of a completed motion-understanding system.
Multi-Dancer GenerationIn preparation · Sony × UT Austin collaboration
Questions worth staying with
05 / Across the three directions
When does explicit 3D / 4D earn its complexity?
When a body can be reused for a new viewpoint, motion, edit, or interaction. A better rendered frame alone is not the only test.
What should transfer between human and animal motion?
Shared ideas about articulation, temporal structure, and contact are useful. Morphology, anatomy, and behavior make blind transfer unreliable.
Can generative priors fill in evidence without inventing the body?
Generation can propose missing views and motion. The research challenge is to preserve identity, anatomy, and consistency while representing uncertainty.
How can we reconstruct both sides of an interaction?
A human and an animal should not be reconstructed as unrelated subjects. Their relative motion, contact, occlusion, and responses define a shared event. Modeling that relationship is the longer-term goal.
Reading threads
06 / Intellectual background
Selected technical lineages behind the map—not my contributions or a ranked bibliography.
