GeoWind2Plan: seconds-scale 3D urban wind prediction for energy-efficient drones

AnimaAnandkumar · x · 2026-10-24

Anima Anandkumar's group shares GeoWind2Plan, a NeurIPS 2026 Spotlight paper with an interactive demo. Key points:

Problem: wind-aware drone planners assume the wind field is known, but CFD simulation around buildings takes hours — impractical for short missions with changing wind
Method: rotating the city layout to align with a fixed reference wind turns wind-direction generalization into building-geometry generalization; a localized neural operator predicts 3D wind in the flight corridor from multi-directional distance features, then a planner optimizes path and speed in continuous 3D space to cut energy use
Results: drones generally fly higher and faster in tailwinds, with routes validated under high-fidelity CFD wind

The lab also released ScaleSplit-NO for memory- and data-efficient 3D turbulence prediction.

More from Embodied