#thermal-soaring.
Posts on the Alpisto engineering blog filed under #thermal-soaring.
- 01#flybeeper #fanet-vario #fanet
The lift does not fade: where 11,046 competition thermals end, and how to know before you get there
The usual advice is to leave a thermal when it weakens. In 11,046 thermals from competition tracks it hardly weakens at all: pilots climb at full rate to within a hundred metres of the top, and the last fifty metres still give most of it. The top is a ceiling, and your own vario cannot see a ceiling coming. The pilots around you can: the ones who have already left, and the tops of the day's earlier thermals, tell you how much is left to within about a hundred metres, a minute before you get there.
9 min read - 02#flybeeper #fanet-vario #fanet
The thermal leans, then it drifts: 8,409 thermals rebuilt from competition tracks
Every pilot knows a thermal drifts with the wind. How much, and in what way, a single pilot cannot tell: climbing and drifting happen at the same time. A competition gaggle can — a hundred pilots in one thermal at different heights in the same minute. From the published tracks of 28 competitions on 18 sites I rebuilt the axis of 8,409 thermals. In the mountains, near the ground, the column stays put and the wind lays it over at 45°; over flat land, and higher up everywhere, it stands nearly straight and the whole thing drifts at two thirds to three quarters of the wind. And the radio a FANET vario already listens to is enough to see it from the air.
21 min read - 03#flybeeper #machine-learning #reinforcement-learning
Machine Learning for Thermal-Soaring Optimisation: a Survey of Approaches, Architectures and Open Problems
A survey of deep learning and reinforcement learning for autonomous thermal soaring — LSTM/GRU/TCN/Transformer architectures, POMDP and PPO/TD3 formulations, Kalman + ML hybrid filtering, and a concrete five-stage plan for an on-board AI co-pilot for paragliders.
17 min read