There's a research goal floating around robotics that sounds, on first hearing, like a joke: build a team of humanoid robots that can beat the human World Cup champions at soccer — on a real field, under real rules — by the year 2050.
I've come to regard that goal as one of the smartest pieces of research engineering of the past few decades. Not because I'm confident it will be achieved (nobody serious is), but because of what the <em>pursuit</em> of it has produced. There's a lesson in it for anyone who funds, leads, or does technical work: the right impossible goal is worth more than a hundred achievable ones.
Why soccer, of all things?
It seems frivolous until you enumerate what the game actually demands. Start at the bottom of the stack: hardware that can move quickly and agilely enough to control a ball. Then perception — sensing a fast-moving object and a fast-changing environment reliably. Then individual decision-making in real time: when to intercept, how to kick, how to keep your footing. Then, above all of that, the multi-agent layer: coordinating with teammates, holding positions instead of swarming the ball like five-year-olds, organizing a team strategically, and adapting your play to the specific opponent in front of you.
No single benchmark exercises that whole stack at once. And the game has a property that makes it uniquely valuable as a testbed for intelligence: it is <em>simultaneously collaborative and adversarial</em>. You cooperate intimately with teammates while competing against adversaries, in continuous real time, with imperfect information. That dual structure is exactly the texture of the messy real world — markets, traffic, logistics — where autonomous systems are neither purely allies nor purely opponents but self-interested agents sharing a space.
That last point deserves a beat. A view I've absorbed deeply over the years is that a single autonomous agent, however capable, is never complete. Almost nothing of consequence gets done by an intelligence acting alone; real intelligence is substantially <em>social</em> — reading other agents, predicting what they'll do, deciding whom to cooperate with. If that's true, then any serious pursuit of artificial intelligence has to grapple with multi-agent interaction, and a challenge problem that bakes it in from the start is aimed at the right target.
The spin-off engine
Here's the part that changed how I think about research strategy. Over decades of the robot soccer effort, dozens of PhD theses have come out of labs that participate. And by all accounts, essentially <em>none</em> of them says "my goal was to solve robot soccer." What happens instead is more interesting: a student works on the game, collides with some maddening sub-problem, and that sub-problem becomes the thesis.
A classic example: early robot soccer vision systems had to be painstakingly calibrated to the lighting in the room — a burnt-out bulb and suddenly the robots couldn't see the ball. One researcher's frustration with that fragility turned into deep foundational work on robust robot perception, which ended up having nothing in particular to do with soccer and everything to do with making real-world robots work. Multiply that by hundreds of labs and thirty years: locomotion, localization, coordination protocols, opponent modeling, real-time planning — the challenge problem functions as a <em>generator</em> of well-posed research questions that nobody would have thought to ask in the abstract.
This is what a good moonshot actually does. It doesn't get solved; it <em>sheds</em>. The 2050 goal is a forcing function that keeps producing honest, concrete sub-problems, each hard enough to matter and grounded enough to evaluate.
The annual reckoning
The other underrated ingredient: competition as a benchmarking ritual. The robot soccer community meets every year and plays. Your approach either works on the field or it doesn't, in front of everyone, against opponents who spent the year trying to beat you.
Contrast that with how most research fields measure progress: static benchmark datasets that saturate and get gamed, or papers evaluated against baselines chosen by the authors. An annual head-to-head competition is a crueler and healthier instrument. It resists overfitting because the opposition adapts. It forces integration — your beautiful planning algorithm must coexist with your teammate's flaky vision stack on a physical robot with a dying battery. And it converts a global research community from a set of rival paper-publishers into something closer to colleagues with a shared scoreboard: competing fiercely at the annual meeting point, and collectively inching toward the same horizon.
I'd argue the AI field's current benchmark malaise — leaderboards that saturate within months of publication — is partly the absence of this structure. A leaderboard is a number; a competition is an <em>ecosystem</em>. One saturates. The other fights back.
Choosing your own 2050 problem
Most of us don't run research labs, but the pattern transfers to any technical organization. A good challenge problem, as I've come to understand it, has a particular shape. It's concrete enough that you know what winning looks like — beat the champions, on a real field, by a date. It's layered enough that every level of your stack, from hardware to strategy, gets stressed. It's genuinely beyond current capability, so it can't be closed by grinding. And it's <em>evocative</em> enough that people opt in for decades — because the sustaining fuel of a thirty-year effort isn't funding, it's fascination.
Then you hold the annual reckoning, and you treat the spin-offs — not the goal — as the actual product. The goal is the lighthouse; the research is the shipping lane that forms around it.
Will robots beat the World Cup champions by 2050? I honestly don't know, and I notice the people closest to the problem don't claim to either. But I've stopped thinking that's the right question. The right question is what the attempt keeps generating — and by that measure, the "joke" goal has outperformed nearly every sober, achievable research agenda of its era. Pick your impossible game. Then let it quietly organize thirty years of your best work.
