Luke Hollis
I'm a research scientist in simulation and computational robotics. I'm interested in world model development and simulation environments for embodied AGI. I create simulation environments for the Harvard Computational Robotics Group and several research groups at MIT.
I previously bootstrapped a spatial computing startup and engineered large-scale enterprise 3D rendering pipelines.
Outside of my research, I volunteer and practice western-style boxing.
w1: World-state Frame Foundation Model
Run simulation with a foundation geospatial model and economic, geopolitical, logistics data to see counterfactual events and alternative histories
real2sim2real
Convert any 3d capture into a segmented scene graph with predictive physics properties. Part of real2sim pipeline using 3dgs.
Autonomous Vehicle Orchestrator
Control groups of robotics and other agents from multiple vendors to collaborate in problem solving with an LLM, world model, and VLA. Evaluated on GDM Melting Pot, SC2, CtF, and similar problems. This project specifically was important to many people.
Overwatch: Multimodal Remote Sensing World Model
Foundation geospatial model trained on multispectral remote sensing data to rapid evaluate change, land usage, flood tracks, and data imputation against noise.
Data Scientist Agent: Automating Exploration and Generation of Knowledge
Agent skills that complete experiment design, data source selection and preprocessing, running regressions and other analysis, and interpreting results across a wide range of fields.
Text-to-3D geospatial simulation with RL+LLM agents
Create a 3D world simulation with live geospatial data. Used for counterfactual analysis, training, and risk assessment.
Causal-ST
Estimate causal effects from diverse multimodal spatiotemporal data streams
AS SEEN IN
Reuters 
Wired 
Reddit 
Hacker News 
BBC 
CNN 
National Geographic 
The Guardian 
Smithsonian 
PBS 
Google 
OpenAI 
Harvard Gazette 
The Harvard Crimson 
University of Chicago 
Lonely Planet 
Matterport 
Polycam 
Replicate 
LiveScience 
Sharp 
Nerdist 
Washington Post 
Boston Globe 
My Modern Met IN PROGRESS
Predictive Real2Sim2Real Digital Twins: A World Model Approach to Deformable Mesh and Soft-Body Simulation Physics
2026Applying world model architectures to predictive physics simulation for deformable objects.
Conformal Inference in Language Model Based Survey Simulation
2026Simulating human behavior with a conformal inference pipeline and LLMs
Spatiotemporal Representations of Urban Mobility and 3DGS with Remote Sensing
2026Novel approach to represent urban mobility data with spatiotemporal transformers and language-grounded 3D gaussian splatting.
EXPERIENCE
EDUCATION
PATENTS
Patent-pending systems and methods for probabilistic world-state estimation and forecasting from heterogeneous real-world observations.