Flow-Informed Path-Planning With Realistic Geometries for Safe Autonomous Flight in Cities

Autonomous flight research usually focuses on the inner-loop, "local'' control of a flier. Relying entirely on on-board sensing and control means the flier can only respond to disturbances after being hit. We propose incorporating the surrounding flow in path-planning, such that wind disturbances can be rejected without the flier ever encountering them. This requires an efficient way to predict flight-relevant wind conditions, possibly on-board and during flight.

In this work, we present a path-planning approach built on a surrogate model of the mean velocity and turbulence intensity fields of wind around realistic city geometries. Our deep-learning based surrogate model is trained on data from a large-scale simulation campaign covering 3D models of cities across the world, in which Lattice-Boltzmann simulations were resolved in time to statistical convergence to produce samples for supervised learning. Our trained model is capable of inference on out-of-sample city geometries orders of magnitude faster than it would take to converge the corresponding simulation, and efficiently enough to run in a laptop or on-board a drone. Our path-planner uses optimal control to balance speed and safety given flow estimates from the surrogate model. We test tracking performance experimentally using a fan-array wind tunnel to generate representative wind conditions in an obstacle course.