
MEAM Seminar: “Real-Time Safe and Energy-Efficient UAV Motion Planning in Windy Urban Environments”
February 19 at 10:15 AM - 11:15 AM
Recent advancements in hardware and software are bringing autonomous aerial vehicles closer than ever to finally delivering on futuristic visions of flying cars and package delivery drones. However, the safe deployment of autonomous aircraft at scale in urban environments poses significant challenges, one of which being uncertainties contributed by complex spatial and temporal winds. Clusters of buildings generate dangerous high speed wind flow patterns which increase both the costs and risks for aerial operations. Existing approaches for modeling and planning for urban winds rely on either distributed sensors, extensive exploration, or accurate global representations of the environment and expensive offline calculations.
In this talk, we first introduce a novel approach to predicting wind flow fields in real time, without GPS or prior knowledge of the environment, and only using on-board sensing with minimal compute requirements. We accomplish this by strategically reducing the problem to predicting the wind flow field on a local robot-centered domain, and by leveraging the predictability of the wind flow field induced by conservation of mass and momentum constraints from nearby topography. Using vast amounts of simulated winds through procedurally-generated urban environments, we train deep neural networks to synthesize navigational LiDAR scans with sparse in-situ measurements of the wind, achieving real-time prediction rates and surprising generalizability. In the latter half of the talk, we demonstrate how the local wind prediction can be incorporated into a receding horizon optimal control architecture to improve the accuracy of trajectory forecasting and cost functions associated with energy consumption. The result is a system that is capable of perceiving surrounding winds in real time using readily-available on-board sensors, and then adjusting the planned route accordingly to avoid environmental hazards and maximize energy efficiency along the way.

Spencer Folk
Ph.D. Candidate, Department of Mechanical Engineering & Applied Mechanics, University of Pennsylvania
Spencer Folk is advised by Mark Yim and Vijay Kumar.