The Geologists of the Future
Can a robot map a planet as well as a human can?
The government SUV is a white dot on the vast volcanic landscape. Beneath the open rear hatch, the geologists Jim Skinner and Alexandra Huff are bent over a map, glancing up at corresponding landmarks. To the west looms the giant lip of a volcano that flooded the area with scorching liquified rock tens of thousands of years ago. To the south, the triangular points of the San Francisco Peaks and, beyond them, the city of Flagstaff, Arizona. Grasses blanket the hills. If you squint, it looks a lot like the surface of the moon.
Skinner, a U.S. Geological Survey scientist with a salt-and-pepper beard and a North Carolina twang, waves me over to the map. He points to the center of a two-and-a-half-mile-wide circle, near where we’re parked. “This is the landing site,” he says.
Skinner and Huff are here in the San Francisco Volcanic Field getting ready for the fifth iteration of the Joint Extravehicular Activities and Human Surface Mobility Test Team, or JETT, one of a raft of exercises to prepare astronauts to once again conduct field science on the moon. NASA’s Artemis III mission, tentatively scheduled for later this decade, will put the first human beings on the lunar surface since the Apollo program ended more than 50 years ago.
[From the May 1910 issue: Through the eyes of the geologist]
Huff, a doctoral student at Arizona State’s School of Earth and Space Exploration, folds up the map and slides it onto a clipboard. She and Skinner, along with a team of others at NASA, have been toiling over maps like this for months, trying to glean the boundaries of the area’s geologic formations from the same type of satellite observations—imagery, elevation data, and radar—that are available from moon-orbiting spacecraft. Following a convention established for this site, they’ve given flat areas the names of desert animals: Javelina Plains, Bighorn Plains. Streambeds and valleys are named after Lord of the Rings characters.
Skinner and Huff are “ground truthing” the satellite observations, walking the site to see how well the maps match up with reality. They don’t expect perfection; the maps are just first drafts, made from a distance. They are like our current maps of the moon—approximate and, inevitably, wrong. They carry hallucinatory artifacts. The seams where orbital images are stitched together, for example, can look like rills on the surface, evidence of ancient erosion, though they represent nothing more than the limits of technology.
In a few months, prospective Artemis astronauts will come here to do the same ground truthing, except they’ll be wearing mock spacesuits, and Skinner and Huff will assess their performance. The idea is to practice field geology using the tools and methods they’ll eventually use on the moon, where their mission will be to close the gaps in our knowledge of the lunar landscape.
Since Apollo, extraterrestrial fieldwork has been done exclusively by robots—wheeled rovers, orbital sensor arrays, even a flying drone. But a person’s ability to gather useful data in the field is “leaps and bounds” better than a rover’s, Huff says. Robots are painfully slow and offer only a narrow field of vision to Earth scientists crowded around computer monitors, whereas astronauts can quickly absorb huge amounts of information themselves and home in on what’s special or interesting. They can piece together the story of a landscape in real time. That ability, as much as anything, is what scientists hope to ship off-world with Artemis, and one day with a crewed mission to Mars.
Yet not everyone believes that this is the dawn of humanity’s solar-system age; some argue that it is, rather, the last gasp of a human-centric sense of what it means to explore the cosmos. “Our emotional preference for human rather than robotic explorers rests on sentiments that each of us formed before we ever attempted to use reason as a guide,” wrote the astrophysicists Martin Rees and Donald Goldsmith in their 2022 book, The End of Astronauts.
A crewed spacecraft could take six years or more to reach orbit around Jupiter, at least a couple of decades to reach Pluto. Humans would be fussy passengers on a trip like that. We need a lot of oxygen and water, and can’t eat sunlight. And unlike humans, robots will keep getting better at everything they do until they’re better at pretty much everything than we are. P. Michael Furlong, a former NASA roboticist who now works at the Computational Neuroscience Research Group at the University of Waterloo, told me there’s “nothing magical about humans … Any capacity we have, given the time and resources, can be automated.”
I came to Arizona because I wanted to understand how the mind of a field scientist works. Could AI-equipped robots imitate our ability to make discoveries far from home? If so, what might we learn about our drive to explore, and about why humans do science in the first place?
As we walk from the truck across a plain labeled Legolas Playa, Huff is already noticing discrepancies between the map and the ground beneath our feet. Where satellite imagery had indicated a dry streambed, there’s only flat sand. Crossing a stretch of dark soil, she clocks it as a formation that hadn’t shown up at all from above. There’s no substitute, she tells me, for “hand on rock.”
It’s meticulous work. For long periods, she and Skinner don’t talk, their eyes on the ground or pressed to a hand lens as they loom over a chunk of rock. But they also argue, amicably, about what they’re seeing. Is this olivine or just moss? Is this ridge a continuation of that one over there?
The astronauts are never far from their minds. Two members of Artemis III’s crew will land in the highlands near the moon’s south pole, where the low angle of sunlight on craters creates permanent shadows that harbor ancient ice. One of the primary goals of the mission will be to determine what form the ice takes—is it a frozen pond? A thin layer of frost? Buried beneath the lunar soil?—which will help determine whether we could someday use it to establish a permanent base there.
Even in person, the scene will be challenging to interpret. That low sun throws long shadows from even the tiniest features on the lunar surface. Distances will be hard to judge—because there’s no atmosphere on the moon, light doesn’t scatter, which means giant mountains in the distance appear just as clear as small ones in the foreground. There are no trees for scale.
All geologic maps are a puzzle of cause and effect. They link events through space and time; this eruption led to this lava flow, which later eroded to form this basin. Huff leads us to a mass of jagged rock jutting out of the plain like a crumbled anvil. She had mapped it as lava, but couldn’t determine its origin from the satellite images. Up close, we can see canted horizontal striations in the face of the outcrop, which could mean the whole chunk broke off the flank of the volcano millennia ago and floated downslope on a river of molten rock. Much more likely, Skinner and Huff explain, the striations are purely coincidental.
[Tyler Austin Harper: The big AI risk not enough people are seeing]
An experienced field geologist can sift the signal from the noise, instinctively discarding explanations that don’t make sense and focusing on the observations most likely to result in meaningful discoveries, sometimes without knowing exactly why. There’s an old saying: “The best geologists are the ones who see the most rocks.” Many of the geologists I spoke with defined that knowledge in terms that are idiosyncratically human—the smell of a mineral warmed by a rock hammer’s strike, or the sink or crunch of the soil underfoot. (One stretch of soil, labeled Scorpion Plains on the map, feels spongier underfoot than a nearby swath; Huff says this means it’s older.)
Robots already outpace humans in their ability to observe many kinds of details. Outfitted with any number of instruments, they can see in a wider spectrum of light or sense objects hidden underground. In the not-too-distant future, robots—most likely groups of robots working together—could certainly create a first-draft map like Huff’s both more accurately and more quickly. But could they do field geology, not just as an extension of human scientists’ senses but on their own? Could they meld what they observe on the ground with what they understand about the processes that put it there?
Huff doesn’t think so. She explains that machine learning, however sophisticated, is still the province of equations. In other words, robots are bound by rules. If human intuition has rules, we don’t yet know them. “Nothing matches the computational power of our brain,” she says.
In 2009, a team led by the British computer scientist Ross King developed a “robot scientist” named Adam. The room-size machine had its own centrifuge and freezer, which it used to grow cultures of common baker’s yeast. By scouring existing knowledge and then mass-testing hypotheses, Adam identified three genes that encoded one of the yeast’s key enzymes, something human scientists had not yet done. A few years later, the team built another robot named Eve that could test new drugs faster, and more cheaply, than a labful of scientists.
Building artificially intelligent field scientists will be harder. The farther they travel, the more in-the-moment decisions robot explorers will have to make on their own. The communications lag time from here to Jupiter’s moon Europa, for example, can be nearly an hour. To learn the discipline of astro-geology well enough to practice it autonomously, an AI would need to ingest untold reservoirs of information about varied landscapes. Then it would need some sense of what to do with it all.
To make matters more complicated, many of the extraterrestrial landscapes robots will have to navigate won’t be familiar, even to their human teachers. Creating an AI scientist that can operate in an open environment that is not only uncontrolled but also poorly understood will require some epic coding. Rick Stevens, a director of the U.S. Department of Energy’s Argonne National Laboratory and one of the world’s leading researchers on automating science, explained that humans learn and explore by constantly processing the balance between what they expect to happen and what actually happens. Robots sent to Pluto may have only a vague idea of what they’ll encounter there; it will be hard for them to focus on unexpected details when virtually everything they see is unexpected.
There’s also the question of what exactly to program the robot to do once it finds something new. When a human discovers something they don’t recognize, Stevens said, “we get excited, right? Our blood pressure goes up; our endorphins kick off … That causes us to kick into another kind of stage.”
The only geologist to have walked on the moon to date is the Apollo 17 astronaut Harrison “Jack” Schmitt, who practically had an aneurysm when he noticed that some of the lunar soil was orange—“It's all over!! Orange!!!” He collected some on the spot, and it turned out to be tiny beads of glass thrown from an explosive volcanic eruption 3.6 billion years ago. At the time, Schmitt didn’t appear concerned about the importance of the discovery in the greater context of lunar science. It was just cool.
“Will a machine ever get pleasure out of figuring out a problem?” Gregory Feist, a psychologist at San Jose State who studies scientific talent, asked me. “The joy is not trivial.”
The excitement of learning something new, purely for its own sake, may be the product of an almost unfathomable latticework of cognition and knowledge converging on an often fleeting observation. “We’ve got lots of different things going on in our head, lots of different parts of the brain attending to different things with different capabilities and different functions,” David Wettergreen, a research professor at Carnegie Mellon’s Robotics Institute, told me. While field scientists are traipsing around the landscape, they’re carrying around their whole life’s experience, and also thinking about the weather, a memory from childhood, or lunch. “Maybe out of that soup is where we start to get the diversity of ideas that we’re able to deploy all at once,” Wettergreen said.
In other words, it’s hard to tell which of those layers is essential to doing science, or whether all of them are.
The sun is setting in a melon-colored band on the horizon, and a chilly wind picks up. While we eat our rehydrated dinners with plastic spoons, Skinner tells me he grew up mapping the bike trails in the woods near his house, digging in the dirt, and reading and rereading the “Space” entry in his family’s Encyclopedia Britannica. To put himself through graduate school for geology, he waited tables in a shirt with little peppers on it and donated plasma.
Huff was a competitive swimmer who planned to serve in the military until she got her hands on some rocks. She wants to leave Earth one day, with what she hopes will be a wave of interplanetary field geologists. Skinner is content with a walk-on part in this grand push back to the moon, which he describes as the most important work of his career. The astronauts, he says, are an extension of a much larger group of people. In a sense, they’re all going.
[From the January/February 2023 issue: Seeing earth from space will change you]
Someday, the robotic descendents of Adam and Eve may have their own cares, their own questions, based on their own experience of living in the universe. “AI could just as well stand for ‘alien intelligence,’” wrote Kevin Kelly, a co-founder of Wired, in 2016. “An AI will think about science like an alien, vastly different than any human scientist, thereby provoking us humans to think about science differently.”
It’s possible that AI will not only discover as much as people could in space but also learn what we’re incapable of learning or even understanding. Perhaps AI will know and love the universe in its own way. What will it name the places we’ll never see?
When the sun is good and down, we drive up a rutted dirt road to another mock landing site—Huff wants to try walking it in the dark to give her a sense of how hard the process will be for astronauts on the moon. The night has turned still, and recent rain has brought out mosquitos that crowd around our headlamps. Huff struggles up a slope, gesturing around her at a jumble of rocks and complaining that the orbital images didn’t register it as a hazard, especially for the wheeled cart the astronauts will use to carry samples.
Skinner is counting out their paces while Huff keeps a running commentary of what’s on the ground for their notes. This rock is the size of a soccer ball, she says, that one a baseball.
Then the sky clears, and their true field site, both familiar and impossibly strange, emerges overhead amid a wilderness of stars.
What's Your Reaction?