2015-07-27: Autonomy Incubator Intern Duo Tackles Kalman Filtering

Josh Eddy and Mike Esswein: the few and the proud

 This week marks the penultimate week for most Autonomy Incubator (AI) interns, and the last five days of work for Michael Esswein and Lauren Howell, the Autonomy Incubator's two Aero Scholars. On Thursday, both will present their research before parting ways to opposite ends of the country—Lauren to Alabama, Michael to the University of Buffalo.  We covered Lauren's incredible achievements for the summer last week, but what about Michael? Where has he been for the past nine weeks?

Michael ("Mike" to his friends, which are everyone) has been working on a team with fellow intern and soon-to-be Master's student Josh Eddy, investigating sensor fusion using a genre of algorithms called Kalman filters.  Remember those from Loc's work? They filter out noise from sensor data, and can also fuse data from different sensors (hence, sensor fusion) into one estimate of where the vehicle is in space. Loc, as we remember, is researching applications for extended Kalman filters (EKF) in sensor fusion for autonomous machines, while PI Jim Neilan is doing similar research with a more recent iteration called an unscented Kalman filter (UKF). For those curious, the UKF takes its name for a mathematical concept it uses called the "unscented transform," which in turn was named after a stick of unscented deodorant that its creator spotted on his labmate's desk. Science is amazing.

Josh and Mike's mission this summer has been to support Loc and Jim's research; they've spent the past two months reading jargon-riddled papers and teaching themselves out of textbooks from AI Head Danette Allen's office library—which, conveniently, is full of Kalman filtering books that were carefully read and annotated for her PhD thesis. It's been a Herculean endeavor that's kept both of them in the AI's intern cave (the colloquial term for the warm, dimly-lit room where the interns collaborate) for long hours each week. To keep morale up, they even sectioned off part of their whiteboard for the public to leave their Kalman-related jokes:

Josh explains Kalman filters as King Tuten-Kalman looks on.

This is the part of the blog where we explain how Kalman filters work, which is going to be difficult because there's math involved and the person who writes our blogs is an English major at a liberal arts college. Luckily for all of us, Josh and Mike are here to help everyone understand Kalman filters.

"First, read this," Josh said, flopping a massive C++ textbook onto the table.

"Then this," Mike said, holding up a Kalman filter book from the pile on his desk.

"Then go to grad school," Josh concurred. Finally, with a little more plying, they agreed to at least attempt to explain Kalman filtering to the American public.

Mike's book of the month.

"What Kalman filtering is, is a form of state estimation. We're answering a question, and that question is, 'What is the state of our system right now?'" Josh said. "It's an extremely powerful method of estimating the position of a vehicle." Basically, it provides a snapshot of what the vehicle is doing at a point in time, compares it to previous snapshots, and calculates where the vehicle is in space based on the differences.

Mike, meanwhile, emphasized the ability of the Kalman filter to seemingly create order in a chaotic universe of data.

"One of the things you have is random noise or drift, and one of the cool things about that is even though it's random, it always forms a bell curve," Mike said. That's because the filter uses normally distributed (and white) noise to account for uncertainty in its estimates.

Interestingly, both of them have dealt with Kalman filtering before: Josh at his internship at the National Institute of Aeronautics last year, and Mike through his his work on the University of Buffalo's joint satellite project with the Air Force and NASA.

"Once it's on your resume, you become the Kalman filter guy wherever you go," Mike said, to a knowing laugh from Josh.