A robot taught itself football in a 140-year virtual World Cup

Yifan Su@yifansu312

When I was a kid waking up at 3am to watch Messi, I definitely didn’t think I’d somehow end up working on robot soccer years later.

We originally wanted to get this video out around the World Cup final to cheer for Argentina and Messi. We’re a little late now, and Messi has already said goodbye to the national team.

So I guess this can be my slightly late small tribute to the player who made me fall in love with football in the first place.

I’ll save most of the technical details for later — there’s a lot more we want to share. For now, just a few thoughts on why I’m excited about this direction.

AlphaGo and AlphaStar showed how far self-play can go in digital worlds.

One question I’ve always found really interesting is:

What happens if we bring the idea self-play into the physical world?

With robots, you can’t just come up with a good strategy. It also has to actually work through balance, contact, noisy perception, hardware limitations, and an opponent that doesn’t behave exactly the way you expect.

That’s what makes self-play especially interesting to me in robotics.

It’s not just about getting better at a game. It gives us a way to study how embodied agents can keep adapting through interaction and discover more robust behaviors on their own.

And the bigger vision here is definitely more than a robot soccer demo.

I like to think of it as Real-world FIFA:

robots actually learning skills, tactics, competition, and eventually cooperation — instead of just replaying motions we designed for them.

To me, soccer is a really fun testbed for embodied intelligence because everything comes together in one place: perception, control, decision-making, and multi-agent learning.

We’ve shown robot-vs-human. Expect to see 1v1, and then more players.

Really excited to see where this goes.

More to come. 🔵🔴🇦🇷⚽️🤖