No. 003 · Systems engineering

The camera that costs forty liters of dirt

NASA's 2024 Lunabotics rules put an explicit price on every kilogram, every watt-hour and every camera a student-built robot carried. Read the scoring sheet closely and it stops being a rulebook. It becomes an objective function.

In May 2024, fifty-eight university teams applied to NASA's Lunabotics Challenge, forty-two were invited to the qualifying round at the University of Central Florida, and ten reached the final at the Kennedy Space Center. The task had changed that year. Previous competitions asked robots to mine as much regolith simulant as they could and deposit it in a collector. The 2024 competition asked them to build something: a berm, a raised embankment of soil, constructed inside a marked target zone 1.5 meters long and 0.9 meters wide, in a bed of Lunar Highlands Simulant roughly 8.1 by 4.6 meters and a meter deep. Each team got two fifteen-minute construction attempts. Operators sat where they could neither see nor hear the arena.

Berms matter on the Moon for unglamorous reasons. A landing engine firing over unconsolidated regolith throws particles at speeds that damage hardware kilometers away, and a wall of piled soil is the cheapest blast shield available, because the material is already there. So the competition was, in a narrow sense, a rehearsal. But the more interesting artifact the 2024 challenge produced was not any of the robots. It was the scoring sheet.

The rulebook is an objective function

Most engineering competitions score outcomes and gesture vaguely at efficiency. Lunabotics published exchange rates. The construction score began at zero, and then:

Teams earned 5,000 points for every cubic meter of berm built above grade inside the target zone, with no minimum threshold. Against that, they lost 8 points per kilogram of total robot mass, up to an 80-kilogram ceiling. They lost one point per watt-hour of energy consumed, logged by a commercial data logger. They lost one point for every 50 kilobits per second of average data bandwidth used. And they lost 200 points for each situational-awareness camera carried. Passing safety and communications inspection was worth a flat 1,000. Dust-tolerant design was worth up to 30 points and dust-free operation up to 70. Autonomy was tiered rather than additive: 75 points for automated excavation, 75 for automated dumping, 175 for automated travel, and, for a machine that completed a cycle unaided, 400 for one cycle or 500 for two or more.

Every one of those coefficients converts into the same unit. At 5,000 points per cubic meter, a point is worth 0.2 liters of dirt. Once you notice that, the rulebook can be read as a single sentence about what a design is allowed to cost.

What a kilogram is worth

Take the mass penalty. Eight points per kilogram, at 5 points per liter, means one kilogram of robot must earn back 1.6 liters of berm to break even. That is a small volume — about half a gallon of soil. So mass is cheap, and the intuition that a lighter robot always scores better is wrong. A heavier drivetrain that improves traction and lets the machine take deeper cuts will usually pay for itself several times over.

But the penalty is linear and the ceiling is hard. A team that arrives at the full 80 kilograms starts the run 640 points down. To break even on mass alone, that robot has to place 0.128 cubic meters of simulant — about 128 liters, which spread over the 1.35-square-meter target zone is a berm averaging roughly 9.5 centimeters high. At the loose bulk density of the highlands simulant, somewhere near 1.6 metric tons per cubic meter, that is on the order of 205 kilograms of soil moved by an 80-kilogram machine inside a fifteen-minute window. The mass penalty is not what makes heavy robots lose. It is that heavy robots tend to be heavy for reasons that also cost energy.

Energy is priced at one point per watt-hour, or 0.2 liters per watt-hour. A robot drawing 300 watts for the full fifteen minutes spends 75 watt-hours and gives back 15 liters. Bandwidth is the same currency at a different rate: 50 kilobits per second averaged over the run costs one point, so a team streaming two megabits per second of video for the full window is surrendering something like 40 points, or 8 liters. These are real numbers, but they are second-order.

The most expensive part on the robot

The camera penalty is the one that reorganizes a design. Two hundred points per situational-awareness camera is 40 liters of berm — twenty-five times the cost of a kilogram of structure, and comparable to an eighth of everything a competitive team built in an entire run. NASA was not being arbitrary. On the Moon, the expensive thing about a camera is not the sensor; it is the continuous, low-latency, human-in-the-loop link that makes the sensor useful, and the round-trip delay that makes teleoperation awkward even at lunar distance. The rule encodes an operational truth as a number.

That single coefficient forces the central trade study of the competition. A team can drive by camera, accept the 200-point hit per view, and rely on a human operator who is fast, adaptive and good at recovering from surprises. Or it can pursue autonomy, where automated travel alone is worth 175 points and a complete autonomous cycle is worth 400, or 500 for two cycles or more. The decision is not primarily about which is technically nicer. It is about which one the team can actually deliver in a nine-month build cycle with undergraduate labour, because an autonomy stack that fails during the run scores nothing while the cameras it replaced would at least have scored something.

This is the shape of nearly every real engineering decision, and it is the part that classroom problems almost never include: the option with the higher expected value under ideal execution is not always the option with the higher expected value under your execution.

Why the answer is not "minimize everything"

Students meeting a constrained problem for the first time tend to attack every penalty term at once — shave mass, cut power, drop cameras, minimize data. The scoring sheet punishes that instinct, because the penalties are small and linear while the reward term is large and depends on capability. A robot optimized down to 25 kilograms, one camera and 60 watt-hours has saved itself 440 penalty points against the mass ceiling and may well be unable to move enough soil to matter. The berm term dominates. Everything else is a correction.

There is also a caveat worth stating in front of students, because it is exactly the kind of thing that separates a model from the world it models. The competition runs in Earth gravity. Excavation force on a wheeled machine is limited by traction, and traction scales with weight, so on the Moon an 80-kilogram robot presses down with roughly the force of a 13-kilogram robot here. The mass–capability relationship that makes a heavy design pay off in a Florida sandpit does not transfer cleanly to the surface it is meant to rehearse. The scoring sheet is a good model. It is not the Moon.

The trade study, formalized

Lunabotics required teams to document their design using NASA systems engineering methods, which is to say: not to arrive at a configuration and defend it, but to enumerate alternatives, define the criteria, weight them, and show the arithmetic that selected one. A trade study is simply the honest version of a decision — the version where the rejected options are written down alongside the reason they were rejected, and where the reason is a number.

The results that year were unusually instructive. Iowa State University and the University of Alabama tied for the Artemis Grand Prize, the first tie in the competition's history; the University of Utah took third in robotic construction, third in autonomy and third in presentations. Two teams reached the same total by different routes through the same objective function, which is the clearest possible demonstration that a well-posed trade space has more than one good answer in it.

Whether students can find those answers depends on a skill that sits slightly to the side of the mathematics they are usually taught. Solving for x is a procedure. Deciding which x to solve for — reading a set of rules, noticing that four penalties and one reward all reduce to a common unit, and computing the break-even point where a design change starts paying for itself — is a different act, and it is the one that engineering actually consists of.