Quadrotor navigation: TRUCE
TRUCE combines persistent route guidance, recovery from measured vehicle state and predictive continuation checks. Polynomial references are checked using Bernstein bounds and subdivision, separately from predicted vehicle motion.
On 100 shared MetaUrban tasks under simulated quadrotor dynamics, TRUCE achieved 83% goal reaching with no recorded contacts; official ROS EGO v1 achieved 22%, and EGO v1 with Gatekeeper achieved 26%. A separate 240-run simulation study increased aggregate goal reaching from 104/120 to 118/120 and reduced recorded contacts from eight to zero, at increased computation and travel time. These observations do not establish real-world collision guarantees.
The first-author manuscript with David Boyle has been submitted to ICRA 2027.
Physics-constrained wildfire prediction
At Texas A&M University, with Yiwei Lyu, I work on learned spread-rate corrections and differentiable anisotropic Eikonal propagation. Current work includes FireBench and WRF-SFIRE evaluation pipelines, fireline geometry and arrival-time comparisons, and investigation of intensity and uncertainty modelling. Project
Teleoperation and manipulation
At Penn State’s ARI Lab, with Junyi Geng, I configure operator viewpoints and camera views, and develop robotic-arm control and software simulations for teleoperation experiments. Project
Connectome-inspired control: early exploration
MaleCNS v1.0 data have been verified. The next steps are circuit selection, controller and adapter design, simulator selection and matched baselines. Training has not started; cross-morphology transfer remains an objective to test. Project