Illustration: Laura H. Azran

There’s never been an aqueous robot quite like DEPTHX. Most autonomous underwater vehicles look the same, Stone says. “Some have fat midsections, some are more elongated, but they pretty much all look like weird torpedoes.” We’re sitting in a cramped, airless room that’s become his temporary headquarters at the ranch. The temperature outside is climbing past 32 °C, and it’s even hotter in here, but Stone doesn’t seem to notice. He’s fired up, and he could go on for hours.

“Their design is dictated by their mission: traveling in straight lines at relatively high speed to survey the ocean floor or gather bathymetry data,” he continues. But for exploring uncharted territory, that shape can get you in trouble. You can back yourself into a tight spot where you can’t turn around.

Two years ago, a team at the University of Southampton, England, learned that lesson the hard way. Their autonomous robot, named Autosub, was exploring the ice shelf below Antarctica. “It was 17 kilometers in, under 200 meters of ice, when it got itself into a position where it couldn’t figure out which way was home,” Stone says. And so the $10 million robot parked itself, sent out an emergency signal, and waited. And waited. “The scientists back on the surface could still pick up its signal, but there was no way they could go in and rescue it.” It’s presumably still there, somewhere below the ice.

DEPTHX, by contrast, is designed not for high speed but for complicated maneuvering in unfamiliar environments. Hence its shape: a squashed sphere with no protruding parts to catch on things. “We used to have a Wi-Fi antenna on top,” Kerr says. “But during one run we surfaced under the chase boat and it snapped off.” And if one of the robot’s six thrusters goes out, it can just rotate around and use the others.

With its top half encased in pebbly orange syntactic foam, for buoyancy, the robot looks kind of like a giant tangerine. The vehicle’s shape was also dictated by its mapping software, known as SLAM, for simultaneous localization and mapping.

SLAM “is designed to solve a chicken-and-egg problem,” says David Wettergreen, an associate research professor at Carnegie Mellon’s Field Robotics Institute, in Pittsburgh, which was responsible for DEPTHX’s software—all 100 000 lines of it. “To build a map, you have to know where you are, but to know where you are, you need a map.” SLAM does both things simultaneously, creating a 3-D map as it moves along and then positioning itself within the map.

Variations of SLAM are commonly used in robotics, and they typically rely on identifying distinct features in the surroundings, like a doorway or a tree, viewing those features from many points, and then triangulating the robot’s relative position. But underwater environments have few recognizable features—a slime-covered wall looks very different when viewed from the front and from the sides.

So Nathaniel Fairfield, a Ph.D. student at Carnegie Mellon who wrote the SLAM algorithm for DEPTHX, designed the software to look not at discrete objects but at the shape of the environment as a whole. To do that, the robot uses 56 sonars, mounted on two circular steel frames that intersect at the top and bottom of the vehicle. As the robot descends through the cenote at the leisurely pace of 1 meter per second, it also spins around about once per minute; each sonar fires up to four times per second, allowing the beams to “paint in” the surroundings.

Controlling the sonars and the 19 other subsystems on the machine are 36 onboard computers. Power is supplied by a pair of lithium-ion battery packs that will run for up to 5 hours between recharges. Much of the hardware is enclosed in pressure-resistant aluminum housings, to protect the contents from being crushed by the external water pressure. Other components are built into oil-filled housings, which balance the outside pressure while keeping water out.

DEPTHX’s other key piece of software gives the robot autonomy, allowing it to make decisions about when and where to move. The Carnegie Mellon programmers paid particular attention to how the robot deals with faults. “We have contingency plans for all kinds of failures, like all the sonars turning off at once, or one battery giving out, or the robot losing its way,” Wettergreen explains. “If something goes wrong, and it’s at the bottom of the cenote, with its batteries running low, it can’t just stop and wait. It has to do something sensible”—initiate a controlled ascent, for example.

Autonomy also means the robot has to decide on the fly where and whether to gather biological samples. The machine starts by characterizing its surroundings. Sensors continuously measure the water’s salinity, temperature, pressure, and chemistry. “Changes in any of these conditions are where we’d expect to find biological activity,” says Ernest Franke, an engineer from Southwest Research Institute, where the robot’s sampling arm and science autonomy system were designed and built.

The robot then “trains” itself by taking a baseline water sample. The liquid is inspected under an onboard microscope, and a subroutine counts any moving objects (likely micro-organisms), tracks their paths, and measures their speed. Another subroutine tells the robot’s video camera to take a baseline reading of the cenote’s slime-covered walls, measuring their color, intensity or saturation, and texture.

The result of each subroutine is a statistical classifier, a value that averages all the parameters, much as your credit score is computed from details such as your age, salary, and mortgage payment history. These two classifiers—one for water, the other for the walls—are then compared with subsequent readings the robot takes. When it spots a significant difference, suggesting a rise in biological activity, it takes a water or solid sample. For the latter, the robotic arm extends about 2 meters and punches a pinky-size chunk from the cenote wall. To gather liquid, a water sipper on the arm fills up plastic bags. After the robot surfaces, the team will remove the specimens and freeze them in liquid nitrogen for later analysis.

Any robotic probe that swims on Europa will have to do all that—and much more. Stone has no doubt that machines will eventually have the requisite smarts, though. Europa will still be there in 10 or 20 years, he notes. “We have time.”