Skip to content
Back to the collection

Autonomous navigation of stratospheric balloons using reinforcement learning

Marc G. Bellemare, Salvatore Candido, Pablo Samuel Castro

Show all 8 authors

Jun Gong, Marlos C. Machado, Subhodeep Moitra, Sameera S. Ponda, Ziyu Wang

2020 · Nature

Read the paper
HN score
0
Mentions
3
Readers
3
Months
3
What’s in a score?

In the conversation

Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels

3 selected excerpts · 3 total mentions

alsodumb
Neutral mention · rule
A lot of people don't realize that a relatively simple balloon can be designed to take off from anywhere in the world, reach any other point in the world, and hover for multiple days sending back data with very little deviation. All the balloon needs is a way to change its altitude, that's it, it doesn't need any actuators to move in horizontal direction. Google's project Loon (https://x.company/projects/loon/) did this and I thought it was pretty darn cool. The way it works is this: different layers of stratosphere have different wind directions. If you can accurately predict the wind directions in different layers, you can in theory move in horizontal direction just by…
Read the full comment
changing your altitude. Loon had pretty good models and simulated the wind directions well. Their balloon used to take off from Nevada and reach Kenya (I believe) and hover there for multiple days to provide comms. They use a cool reinforcement learning based controller in their loop: https://www.nature.com/articles/s41586-020-2939-8 I remember some Chinese researchers writing papers on this topic back then and them saying they'd have good results if they had access to data from Loon: https://www.nature.com/articles/srep45976 I don't think Loon ever shared any data publicly but Loon did get shut down (like a lot of other Alphabet's moonshots) and I remember some of their tech and patents were sold to SoftBank and others. I was surprised US did not try to stop this in the interest of national security.
activatedgeek
Neutral mention · rule
This is quite a surprise. Only recently, Loon collaborated with seasoned reinforcement learning researchers to build more efficient navigation controllers for the balloons, titled "Autonomous navigation of stratospheric balloons using reinforcement learning" [1]. Although, it would appear that much of the gains come from a simulator that they've been developing for a while. [1]: https://www.nature.com/articles/s41586-020-2939-8

Keep following the thread.

More in Robotics

More from the same topics.