Interstellar objects are among the last unexplored classes of solar system objects, holding tantalizing information about primitive materials from exoplanetary star systems. They pass through our solar system only once in their lifetime at speeds of tens of kilometers per second, making them elusive. Hiroyasu Tsukamoto, a faculty member in the Department of Aerospace Engineering in The Grainger College of Engineering, University of Illinois Urbana-Champaign, developed Neural-Rendezvous—a deep learning-driven guidance and control framework to autonomously encounter these extremely fast-moving objects.

“A human brain has many capabilities: talking, writing, etcetera,” Tsukamoto said. “Deep learning creates a brain specialized for one of these capabilities with a domain-specific knowledge. In this case, Neural-Rendezvous learns all the information it needs to encounter an ISO, while also considering the safety-critical, high-cost nature of space exploration.”

Tsukamoto said Neural-Rendezvous is based on contraction theory for data-driven nonlinear control systems, which he developed for his Ph.D. at Caltech, while this project was a collaboration with NASA’s Jet Propulsion Laboratory, where he spent his time as a post-doctoral research affiliate.

“Our key contribution is not just in designing the specialized brain, but in proving mathematically that it works. For example, with a human brain we learn from experience how to navigate safely while driving. But what are the mathematics behind it? How do we know and how can we make sure we won’t hit anyone?”

In space, Neural-Rendezvous autonomously predicts a spacecraft’s best action, based on data, but with a formal probabilistic bound on its distance to the target ISO.

Tsukamoto said there are two main challenges: the interstellar object is a high-energy, high-speed target, and its trajectory is always poorly constrained due to the unpredictable nature of its visit.

“We’re trying to encounter an astronomical object that streaks through our solar system just once and we don’t want to miss the opportunity. Even though we can approximate the dynamics of ISOs ahead of time, they still come with large state uncertainty because we cannot predict the timing of their visit. That's a challenge.”

The speed and uncertainty of ISO encounters are also why the spacecraft must be able to think on its own.

“Unlike traditional approaches in which you design almost everything before you launch a spacecraft, to encounter an ISO, a spacecraft has to have something like a human brain, specifically designed for this mission, to fully respond to data onboard in real time.”

Tsukamoto also demonstrated Neural-Rendezvous using multi-spacecraft simulators called M-STAR and tiny drones called Crazyflies. While he was at JPL, two Illinois aerospace undergraduate students, Arna Bhardwaj and Shishir Bhatta, contacted him to work on a research project using Neural-Rendezvous.

“Because of the speed and uncertainty, it's challenging to obtain a clear view of an ISO during a flyby with 100 percent accuracy, even with Neural-Rendezvous. Arna and Shishir wanted to show that Neural-Rendezvous could benefit from a multi-spacecraft concept.”

To theoretically justify the empirical observations from the M-STAR and Crazyfly demonstrations, their research looked at how to mathematically maximize the information gathered from the ISO encounter using a swarm of spacecraft.

“Now we have an additional layer of decision-making during the ISO encounter,” Tsukamoto said. “How do you optimally position multiple spacecraft to maximize the information you can get out of it? Their solution was to distribute the spacecraft to visually cover the highly probable region of the ISO’s position, which is driven by Neural-Rendezvous.”

Tsukamoto said he was impressed with the level of dedication and academic potential demonstrated by Bhardwaj and Bhatta.

“The topics explored in Neural-Rendezvous can be advanced even for Ph.D. students. Arna and Shishir were very productive and worked hard, and I was surprised to see them publish a paper, given this field initially was entirely new to them. They did a great job.

“And while the Neural-Rendezvous is more of a theoretical concept, their work is our first attempt to make it much more useful, more practical.”

The study, “Neural-Rendezvous: Provably Robust Guidance and Control to Encounter Interstellar Objects,” by Hiroyasu Tsukamoto, Soon-Jo Chung at California Institute of Technology, Yashwanth Kumar Nakka at Georgia Institute of Technology, and Benjamin Donitz, Declan Mages, and Michel Ingham at NASA’s Jet Propulsion Laboratory, is published in the Journal of Guidance, Control, and Dynamics. DOI:10.2514/1.G007671

The study, “Information-Optimal Multi-Spacecraft Positioning for Interstellar Object Exploration,” by Arna Bhardwaj, Shishir Bhatta, and Hiroyasu Tsukamoto, was presented today at the IEEE Aerospace Conference. The abstract is available on the conference site, where the full paper will be posted at a later date. The full paper is available now atDOI: 10.48550/arXiv.2411.09110