RESEARCH

AI that perceives, reasons about and acts in physical space

RESEARCH DIRECTION

Structured Spatial World Models

Drones and robots sense the world, build accurate 3D world models, and use those models to predict, decide and act. What they learn from acting flows back to update the model.

Structured Spatial World Models Flagship direction: Autonomous mapping agents

Sensing · Autonomy See the world now
- Image–LiDAR co-serialization
- Resident space object (RSO) tiny-object detection
- Pedestrian trajectory sensing
- Quality-aware drone mapping
3D World Model Know what the world looks like
- LoD1 to LoD2 building reconstruction
- Semantic-first BIM/IFC digital twins
- Single-view 3D reconstruction with diffusion priors (GeoPriorPC)
- Diffusion models for RSO imagery
- Gaussian splatting for 3D city models
Prediction · Action Predict what happens before acting
- Spatial graph + LLM agents
- Pedestrian and CFD simulation
- Vision-language models for railways
- Autonomous mapping agents
- 3D-grounded vision-language-action models

Results of action feed back into sensing, closing the loop.

Infrastructure · Digital twinsMaps, BIM, power lines and rail
Mobility · Public healthDelivery robots, pedestrian flow, disease modelling
SpaceSpace situational awareness (K-SSA)

NSERC Discovery: Embodied Spatial IntelligenceThe program that supports all of the lab's research

Q-Drone UWB positioning system and test sites

PILLAR 01 · SEE THE WORLD NOW

Sensing · Autonomy

Drones, mobile mappers and satellites collect LiDAR, imagery and orbital observations. We fuse these sensors and detect what matters in them, from tiny objects in orbit to people on a sidewalk.

TOPICS

Image–LiDAR co-serialization
OCI C2C · Aethon Jacob Yoo, Youssef Korny

Resident space object (RSO) tiny-object detection
K-SSA, with Prof. Regina Lee's lab Youssef Beshir

Pedestrian trajectory sensing
OCI C2C, with Hanyang University Mohammadjavad Ghorbanalivakili

Quality-aware drone mapping Looking for students

PILLAR 02 · KNOW WHAT THE WORLD LOOKS LIKE

3D World Model

We turn sensor data into accurate, structured 3D models of cities and infrastructure, from building geometry to semantic BIM, so machines can reason about space.

TOPICS

LoD1 to LoD2 building reconstruction Mohammad Moein Sheikholeslami

Semantic-first BIM/IFC digital twins
with Yonsei University Md. Ehsan Shahmi Chowdhury

Single-view 3D reconstruction with diffusion priors (GeoPriorPC)
NSERC Discovery Youssef Korny

Diffusion models for RSO imagery
K-SSA, with Prof. Regina Lee's lab Merna Tamer Youssef

Gaussian splatting for 3D city models Looking for students

Simulating a sidewalk delivery robot in a campus digital twin

PILLAR 03 · PREDICT WHAT HAPPENS BEFORE ACTING

Prediction · Action

With a world model in hand, agents can plan, simulate and act. We build agents that map on their own, connect spatial graphs to language models, and simulate how people and vehicles move.

TOPICS

Spatial graph + LLM agents
ORF-RE Md. Ehsan Shahmi Chowdhury

Pedestrian and CFD simulation
ORF-RE · NSERC CREATE SMART Mohammadjavad Ghorbanalivakili

Vision-language models for railways Mohammadjavad Ghorbanalivakili, Ashley Varghese

Autonomous mapping agents Looking for students

3D-grounded vision-language-action models Looking for students