The Balloon Learning Environment
ai.googleblog.comFor a bit of background, go to: https://x.company/projects/loon/
I was working on the Air Traffic Control side of unmanned free balloons in 2020, and Loon was an exciting project. While it's a shame it wasn't financially viable, they had it working with up to 60-ish balloons in the stratosphere at any given time. This was an amazing accomplishment.
The navigation was equally amazing. It was Loon who did some real research on stratospheric weather. While many meteorological entities throughout the world knew and understood tropospheric weather, the stratosphere was mostly unknown. In addition to AI, Google really did expand the knowledge and science of stratospheric weather.
I thought this project was pretty impressive. I feel like this is a thing that is going to come back around once costs come down and a killer app is found for it.
Balloon retrieval might be a problem that would have to be solved while keeping operation costs low.
I knew some friends over at Loon and one of their challenges was it wasn’t highly deterministic where the balloon would land at the end of life, and they had a rescue team with speedboats, helicopters and land rovers to go recover balloons. They could pick a big enough radius and have it land somewhere within that to avoid cities, and human settlements.
Also, Alphabet burned considerable money on this for 9 years before shutting down. Not sure if it’ll come back, at least not by Alphabet.
Yeah I don't buy the baseline comparison of the "expert designed controller" that cannot stay within bounds.
We can control systems that have not only fluid-mechanical, but also heat transfer and chemical reactions, all happening at the same time with huge bars of uncertainty.
Next.
The environment has been released. If you or anyone else wishes to give it a go you can develop a controller any whichever way you fancy.
The baseline comparison is not meant to be some final word on any given controller. In this case they clearly cite it to a particular paper. [1]
1: https://github.com/google/balloon-learning-environment/blob/...
Then their wording is misleading.
They compared against another bad RL system and they are better than that.
That is far from an "expert control system"
Ah! I see where the miscommunication happens. They use the word in a particular context.
Here "expert control system" means that it has a decision logic designed by some humans. They sit down wrote an algorithm and said "I think this is a good idea based on what I know about the problem." Compare it with a reinforcement learning system, where they let loose a generic learning algorithm on the problem and then they see how it performs.
In this context "expert control system" doesn't mean that it was expertly designed, or even that it is any good. It just means that humans designed it for this particular problem.
This makes sense. Thanks for the clarification :)
The word expert is not being used in the colloquial sense, but in the technical sense of an "Expert System": https://en.wikipedia.org/wiki/Expert_system
Between prevailing winds, power and superpressure limitations, etc., it is often the case that a balloon cannot be regulated within bounds regardless of the quality of controller. Nonetheless I’m certain there are better control laws than this solution.