I don't think so. The article says the race track was controlled by the race organisers but not that it wasn't known to the participants before the race.
Anyway given the state of the art, flying autonomously at great speeds and beating human champions without pre-training, i.e. on an unknown race track, would be a much bigger breakthrough than just beating some human champions (which has already happened except in a less official environment). You can rest assured that if that was what the team achieved, the article would be telling us all about it.
Shoot, you're totally right. They had no prior knowledge before the event, but I don't know how they teach it the course. There's more than one gate visible at a time, so they must do something to fine-tune it.
That being said, I'm sure they have a base model too, so I'm right back to wondering about the parent question: would it work if you set it down in front of a few fresh gates?
Probably not. RL is really bad at generalising to unseen environments. There was a paper about an ... otter?
Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial Observability
https://arxiv.org/abs/2107.06277
OK, it's a robotic zookeeper looking for the otter cage.
Where does it say they had no prior knowledge before the event? I can't find that in the text. Is it in the video?
I guess there's no paper yet.
> Where does it say they had no prior knowledge before the event? I can't find that in the text. Is it in the video?
Reading back through it, I'm synthesizing this statement. It's never said explicitly, and I could very well be wrong.
I'm combining the knowledge that the novel development here is that the event supervisors control the track with the fact they're showing off a training run in their video.[0] The video also links a few papers from the teams past that have some additional clues.[1][2]
> The reason for this mostly lies in the real-world aspects of the competitions. They take place in environments previously unknown by the teams, with no opportunity for benign, solution-specific changes, and little time for adapting the developed solution to the environment in situ. Moreover, competitions often pose a more challenging environment, with gates located slightly differently than on the precommunicated maps or even moving during the race, unforeseen lighting effects optimized for spectators rather than for drones, and large crowds of moving people around the flight arena.
This makes it sound like they're at least given the layout.
Note this was from a different competition (Artificial Intelligence Robotic Racing by Lockheed, with DRL) back in 2019. The other paper is from 2024, but I don't have access.
https://www.youtube.com/watch?v=yz2in2eFATE
>> This makes it sound like they're at least given the layout.
Yeah, I read it as saying they were given a short amount of time to train before the race, where they wouldn't have time restrictions otherwise. Which makes the result more impressive, of course.