System Identification (SysID) for UAVs

This subteam focuses on System Identification (SysID) for UAVs: learning the parameters of the ordinary differential equations that describe how an aircraft behaves, from noisy input and output data such as velocity, flight path angle, angle of attack, and pitch rate. These parameters are never measured directly. They have to be inferred from sensor logs,which is exactly the setting where a single best guess is not enough.

We take a Bayesian approach: instead of returning one number, these methods return a probability distribution over the parameters. The peak is the best estimate and the width is how confident the method is, given the data it was shown. This matters most during abnormal flight conditions such as aggressive maneuvers. We also deal with known and unknown external forcings on the system because real systems rarely evolve in isolation.

The obstacle is cost. Scoring one candidate set of parameters normally means simulating the entire system and comparing it to the data, and Bayesian inference needs to do this thousands of times, which rules out real-time use. Our current research uses Gaussian processes to remove that bottleneck. A Gaussian process places a probability distribution over the aircraft’s trajectory itself rather than over a fixed set of parameters, and because it has an exact closed-form expression for its own derivative, the differential equation can be evaluated directly. The numerical solver leaves the inference loop entirely.

Team members start by reading foundational literature and implementing methods on simulated benchmark systems where the true answer is known. From there the work will scal to 3-DOF aircraft models and then to real 6-DOF flight data. The final SysID results will be used for the digital twin simulator, used in the fixed-wing drone competition and also for SysID of the industry partner Windracers US LLC. Potential collaboration with other industry partners is also possible (Skydio etc.).

Activities: 1) Literature review and mathematical formulation 2) Derivation and extension of existing methods 3) Simulation, data collection and preprocessing 4) Model implementation in PyTorch 5) Benchmarking accuracy, uncertainty quality, and runtime 6) Log-to-parameter pipeline and sensitivity analysis.

No prior background in Bayesian methods or machine learning is required. Researchers also have the opportunity to learn to fly fixed-wing drones and collect flight data through our collaboration with PURT and SATT.