Machines need a lot of information to operate autonomously. But, in the same way that your eyes and ears work together to understand what’s around you, multi-sensor systems must be able to combine information into a cohesive picture — and determine which sensor data it can trust.

Keith LeGrand, assistant professor of aeronautics and astronautics, uses mathematical approaches to data and perception challenges. His work in the SCOPE Lab is contributing to research in space domain awareness, multi-object tracking and intelligent sensor control through Purdue’s Cislunar Initiative.

From industry to academia

LeGrand heard his calling to research while he was a senior member of the technical staff at Sandia National Laboratories. He also recruited and mentored interns there, which is part of what inspired him to follow that call.

"I found that mentoring was a really rewarding part of the job because it allowed me to sort of broaden my research reach, and I really liked interacting with the students," he says.

After six years at Sandia, LeGrand and his wife agreed to leave Albuquerque in favor of Ithaca, New York, so he could pursue a PhD in aerospace engineering at Cornell University.

Keith LeGrand, Assistant professor of Aeronautics and Astronautics

"It was a pretty tough decision to give up the really good job, and that security. We sold our car and a lot of our belongings just to kind of prepare for the student salary," he says. "I guess I was pretty driven. I’m very fortunate that my spouse was willing to make the sacrifice also."

Using math to merge data and quash rumors

His work in spacecraft proximity operations at Sandia is what led him to pursue multi-object tracking research. "I was looking at how to autonomously inspect malfunctioning satellites using only local information. I started thinking about a swarm of satellites and estimating their relative trajectories," he says. "At Sandia, there were a million national security applications in mind."

Now running the Sensing, Controls and Probabilistic Estimation (SCOPE) Lab at Purdue, LeGrand is looking at the challenges of proliferated systems that contain hundreds or thousands of small satellites.

"It’s unreasonable for a human to task all 1,000 satellites. You need automation. But when you have large systems, how do we meaningfully control them so they’re always gathering the best information?" he says.

These satellites need to reference data from many different sensors to arrive at an answer. They must also be able to recognize and ignore bad data, and stop it from perpetually biasing results. These functions run into the limited communications and computation capability in aerospace systems.

One way to improve automation is by estimating what there is to gain from new data. This was part of his doctoral research at Cornell, which involved controlling sensors in multi-object tracking systems.

"We have to be really smart about what information we exchange and when, and how much we trust the information we receive. For a satellite, for example, do I get more information if I look at the same place again and compare, or if I look somewhere else?" he says.

"You can boil down these complicated sensing tasks into what’s known as information value, and create an information framework that can autonomously control your air-and spacecraft so you can get the measurements that are most informative."

All these approaches are handled mathematically, informed by statistical models of how each sensor perceives the real world. He believes his approach will be more broadly applicable as well.

"It’s a lot of theory," he says, "but always with application in mind. There’s always at least a simulation of a real system going into my models."