New center harnesses AI to advance autonomous exploration of outer space
A new center at the Stanford School of Engineering will leverage artificial intelligence in the service of space science, exploration, and business.
The Center for AEroSpace Autonomy Research, or CAESAR, aims to make these activities more efficient, safe, and sustainable. Researchers at the center say that AI could optimize navigation for spacecraft; deftly land space vehicles on planets or asteroids; allow unmanned rovers to make decisions about where to go, what to avoid, and what to analyze; keep tabs on all the space junk (some 60,000 pieces, at last count) whirling around Earth, threatening lives and equipment; and much more.
To enable such ambitions, Marco Pavone, CAESAR co-founder and associate professor of aeronautics and astronautics, announced that one of the center’s main projects will be to develop a foundation model for space pursuits. A foundation model is a kind of general-purpose AI that’s trained on huge amounts of data and can handle a variety of tasks, like generating text, images, and even videos. This “space foundation model,” Pavone said, will be designed to synthesize information across a range of modalities, including vision, text, remote sensing (such as multispectral imagery and radar), and space-object catalogs, and will be capable of addressing a variety of space-related tasks, including situational awareness, positioning, and navigation.
“We want to develop rigorous tools for the trusted deployment of AI for spacecraft systems – trusted in the sense that they can behave within bounds described by the user,” said Simone D’Amico, CAESAR co-founder and associate professor of aeronautics and astronautics.
D’Amico made it clear that he, Pavone, and their collaborators plan to proceed prudently, with eyes wide open to the potential pitfalls of pursuing AI for spacecraft and robots. He said that in some cases, AI components are not the most effective choice for space systems.
“We founded CAESAR with the objective to tackle unsolved problems in spaceborne autonomy through the judicious incorporation of artificial intelligence components,” D’Amico said.
D’Amico also noted the constraints space imposes on the fledgling technology. For example, space is a harsh, remote environment that’s not readily available for AI training, and powerful microprocessors needed for AI are still not resilient to space radiation.
Pavone and D’Amico spoke May 22 during a daylong symposium to mark the official launch of CAESAR, which is a collaboration between academia, industry, and government with D’Amico’s Space Rendezvous Lab and Pavone’s Autonomous Systems Lab at its core. The event featured presentations from Stanford faculty members, postdoctoral scholars, and students, as well as representatives from NASA, Aerospace Corp., and aerospace manufacturers Redwire Space, Blue Origin, and Lockheed Martin.
CAESAR projects underway
CAESAR’s initial focus has been on developing machine learning models for space rendezvous, proximity operations, and docking – a suite of maneuvers for bringing two or more spacecraft close enough to one another to interact or dock while in orbit. One of these models, D’Amico said, is the Spacecraft Pose estimation Network (SPN), which integrates machine learning with a classical navigation algorithm to robustly estimate a target spacecraft’s position and orientation from monocular images.
Another is the Autonomous Rendezvous Transformer (ART). ART aims to optimize spacecraft trajectories by allowing AI to provide a “smart initial guess” that is fine-tuned by traditional mathematical optimization, said Daniele Gammelli, a research fellow at CAESAR. This approach could be useful because optimizing spacecraft trajectories with only conventional numerical methods is likely beyond the capabilities of today’s space-grade microprocessors, Gammelli said. “We try to combine the best of both worlds between numerical optimization and learning-based methods,” he said.
In a different realm of space exploration, researchers at CAESAR are designing a small, autonomous robot called ReachBot, which can deploy extendable booms from its body for gripping the walls of, say, a lava tube or rock overhang. Its creators have the Martian landscape in mind. “ReachBot can descend into a Martian lava tube, scan for geologically interesting areas, and drill into the wall to extract and transport material for analysis,” said Daniel Morton, a graduate student in mechanical engineering and one of the project’s lead researchers.
Moon endeavors
CAESAR’s kickoff event was an opportunity to showcase faculty from the School of Engineering unaffiliated with the center but interested in AI-enabled space projects.
Manan Arya, assistant professor of aeronautics and astronautics, wants to anchor a 350-meter-diameter radio reflector in a crater on the far side of the moon. The goal is to shield the telescope from radio interference emanating from Earth. “The reason we want to do that is for looking at very, very early signals from the early part of the universe – a time referred to as the cosmic dark ages,” Arya said.
He explained how such a project could be achieved: A lander touches down at the bottom of a crater. The lander fires cables tipped by harpoons that anchor themselves in the regolith at the rim of the crater. Then, tiny, autonomous robots crawl up these tension cables, deploying a lightweight reflector as they go.
Grace Gao, assistant professor of aeronautics and astronautics, and her lab are also shooting for the moon with several AI-related projects, including one that could help autonomous lunar robots and humans with navigation: a “GPS” system for the moon. Noting that more than 100 missions are planned for the moon over the next decade, Gao said it’s vital to provide positioning, navigation, and timing services there. They would be enabled by a low-cost satellite system in orbit around the moon. “We want to have smaller satellites – as small as a shoebox,” she said. “In comparison, the terrestrial GPS satellites are as big as a truck.” She also said clocks on the lunar satellites could be “1,000-times cheaper” because they could rely on information relayed from atomic clocks on satellites orbiting Earth.
Gao’s lab is also working with NASA’s Jet Propulsion Laboratory to develop the Cooperative Autonomous Distributed Robotic Exploration, or CADRE, project, which aims to put three autonomous rovers on the moon next year. Gao’s team has developed technology to help the robots autonomously navigate while dealing with communication challenges they may experience.
Pavone and D’Amico said they hope to tap the expertise of many engineers and scientists both at Stanford and other organizations to catalyze space-related research projects at CAESAR. Addressing symposium attendees, Pavone circled back to the subject of the space foundation model, which he said will be “a game changer.” “We look forward to engaging with the many people that are here today and the space community in general to build such a capability,” he said.
ReachBot is a joint collaboration between Marco Pavone’s lab and the labs of Mark Cutkosky, the Fletcher Jones Professor in the School of Engineering and professor of mechanical engineering, and Mathieu Lapôtre, assistant professor of Earth and planetary sciences. Redwire Space and Blue Origin are co-sponsors of CAESAR.
