Hacker News Reader

Comment by bodhiandphysics | Hacker News Reader

There’s one really important aspect of starship control that this article mentions... starship control is stochastic. Spacex needs to minimize the area of an ellipse that represents the likely touchdown area of the craft. To do this they use another trick... they continuously recompute the optimal trajectory, in real time. Lars Blackmore has a paper on the process!

View on HN

Yep, Lars learned GFOLD at JPL and took it to Spacex. GFOLD was invented by Behçet Açikmeşe https://www.aa.washington.edu/facultyfinder/behcet-acikmese

Though as I understand, spacex doesn’t actually use GFOLD, but a related (and secret!) development

Not to accuse spaceX of anything, but I've witnessed companies sell 'hyper-localised weather" where they take public weather information and average the nearest two datapoints to you.

So I always take 'secret' 'proprietary' data and algorythms with a bit of salt, as almost every time I dug into them, it was all spin on previous public stuff.

In this case, SpaceX is actually landing those Falcon boosters pretty reliably now. So maybe their secret proprietary stuff isn't as advanced as they make it out to be, but whatever they're doing it does seem effective.

The two have apparently co-authored the relevant papers. So perhaps both of them "learned GFOLD at JPL" in the same sense.

Cool, do you have a link for the paper?

EDIT: Found I think [http://www.larsblackmore.com/nae_bridge_2016.pdf]

This sounds like model predictive control. Does anyone know the difference between the methods?

You can implement it as MPC and minimize e.g. the trace of the covariance matrix that describes the state uncertainty. Wrote my thesis on something similar.

Reply on news.ycombinator.com