A Real-Time Algorithm for Non-Convex Powered Descent Guidance [pdf]

depts.washington.edu

46 points·nill0··7 comments

I must be half asleep, I was expecting some novel machine learning algorithm. I was pleasantly surprised to find a spaceflight algorthim instead.

I am completely out of my depth, so have nothing constructive to add. But it sounds interesting.

Funny, I fell into that one too... I guess "guidance" has a few letters in common with "gradient" so our brain triggers...

See: https://medium.com/@stephenwatathi/the-brains-power-to-read-...

This is the algorithm SpaceX uses to catch rockets. If you watch any of the landings/chopstick catches, you’ll notice that they don’t seem to oscillate back and forth like a lot of control systems we’re “used to” seeing.

This algorithm is why! Super cool.

so as soon as the rocket starts to fall @ the minimum controll speed, they draw a virtual strait line between the rocket and the landing spot, and follow it, sort of an inverse geodetic line, compensate for wind with the "flaps" and vectored thrust on the rockets, giving enough horizontal thrust, and gravity providing a very steady and reliable pull makes a lot of sense, as it is the most fuel efficient and fastest return, and any issues with guidance are imediately clear, making it part of the abort system as well

Totally. It’s the guaranteed convexity that is the cool and awesome part.

Reminds me of L. Blackmore’s https://www.semanticscholar.org/paper/Lossless-Convexificati...

Would love to see this control a classic Lunar Lander type game such as http://moonlander.seb.ly/

Reply on news.ycombinator.com