Physicists train robotic gliders to soar like birds

On Tuesday, September 25, 2018 at 10:06:14 AM UTC-4, Bret Hess wrote:
From the summary this paper used accelerations (bumps) and torques as the method to find and center thermals, as opposed to sensing vertical air speed. Has that been done before?

Sort of.

Folks at the naval postgraduate school have used the time derivative of the vario signal in a controller, this should be pretty highly correlated with the surge acceleration. Their work is here if you have an aiaa subscription (https://arc.aiaa.org/doi/abs/10.2514/1.51691).

The Atlantiksolar team at ETH Zurich has used rolling moment to help figure out where thermals are. (https://onlinelibrary.wiley.com/doi/...1002/rob.21765)

The paper combines the two ideas but takes a different approach, using machine learning to figure out how to react to the surge and rolling moments instead of designing control laws and thermal estimators.

For those equipped with a biological computer trained in thermal soaring, the results will be familiar. As they note, you could summarize the results on how to respond to accelerations as:

* As climb improves, flatten the circle (approx. 15-20 deg)
* As climb deteriorates, steepen the circle (approx. 15-20 deg)
* If climb remains constant, keep constant bank (approx. 25-30 deg)

Which is page 10 in my copy of Reichmann.

It is interesting to note though that they suggest that rolling moment should generally override surges (i.e. almost always turn slightly into a raised wing, regardless of whether you get a surge or not).

I think UH is right on. Thermalling is relatively "easy," a moderately skilled human doesn't really leave a lot on the table, to say nothing of contest winners. How to find that thermal, and where to go afterward is a much harder problem and has a bigger impact on your speed.