Research & Projects
Research & Projects
Eliminate two common causes of field failure in autonomous drone development: simulated environments that do not match the real operating site, and deployed code that differs from what was tested. The pipeline, built entirely on PX4, ROS 2, and Gazebo, separates the application layer from the vehicle layer so identical mission logic runs unchanged in simulation and on real hardware, and pairs it with a georeferenced site twin constructed purely from public LiDAR, aerial orthophotos, and OpenStreetMap building footprints, with no dedicated survey or site access required.
Ran the same application-layer code in simulation and on a physical quadrotor (Holybro X500 V2, Pixhawk 6C, Jetson Orin NX), with only stage-specific connection settings changing between the two
Validated the site twin against independent USGS LiDAR: 0.331 m terrain RMSE over 97% of the site, and a derived building height of 4.3 m matching in-person measurement
Completed 60 simulated and 15 real flights across 3 mission types; real-flight cross-track error never exceeded 2 m, with sub-meter RMSE in every mission
Reused the site twin as a live digital twin that mirrors the real vehicle's position and attitude inside Gazebo during field operations
Produce a seamless, high-resolution daily precipitation dataset for the Laurentian Great Lakes basin that remains accurate for heavy and extreme rainfall, where existing products fail. The approach fuses five gridded precipitation products (CaPA, MPE, ERA5, MERRA-2, NCEP-2) with more than 20,000 rain gauges using a residual U-Net with a factorized spatio-temporal transformer. A masked loss trains the network only at gauged grids, so the output inherits no interpolation artifacts and stays continuous across the US–Canada border and the open lakes.
Released a 10-year dataset (2009–2018) on a 640x800 grid at 2.5 x 2.1 km, distributed as CF-1.6-compliant NetCDF-4
On the held-out 2018 test year, reached a domain-mean RMSE of 2.30 mm/day and MAE of 1.15 mm/day, 40% and 44% below the best of five baselines, with a median per-grid correlation of 0.79 against 0.43 to 0.61 for the baselines
Achieved a critical success index of 0.574 for extreme rainfall (50 mm/day threshold) against at most 0.016 for any baseline, roughly a 35x margin, with detection skill nearly flat from 1 to 50 mm/day while every baseline collapses
Reproduced 86% of observed day-to-day variability at a typical grid, where coarse products shed the heavy tail entirely