Simulation study · SB-XC sailplane · since October 2026

GeneticWing: a soaring planner for a fixed-wing UAV

A simulated sailplane plans each thermal climb with a generator distilled from a genetic search in a learned command space. It climbs 70 % faster than a classic circling rule on tasks it never trained on, in 0.24 s per plan, and now flies in thermals built from a real weather forecast.

The first 8 km of a simulated 20 km cross-country flight in 3D: a red sailplane spirals up drifting thermal columns under cumulus clouds, its trail coloured by climb rate, racing a grey sailplane that uses a fixed circling rule — The first 8 km of a 20 km flight, 22 minutes in 49 s. The red sailplane plans every climb with the latent GA, its trail turning green where it climbs; the grey one flies the same task with the same tactician and a fixed circling rule. It leaves the second thermal 100 s ahead and arrives 4 minutes earlier over the task (26.7 against 24.4 km/h), without using its battery. Aircraft and thermal cores are drawn larger than life.

2.23 m/smean climb of the distilled generator on 64 tasks of 20 km, against 1.29 m/s for the circling rule

0.24 sper climb plan, against 13.4 s for the genetic search it imitates

2.21 m/son 64 new tasks that no training step has seen (circling rule: 1.30 m/s)

Six stages, from a slow teacher to a fast planner

It extends the flatness-based glide planning of Morando et al. (ICRA 2025) from efficient gliding to soaring. A slow offline teacher solves thousands of single-thermal climbs. A variational autoencoder learns its commands (lift coefficient, bank, thrust) and its decoder flies them through a point-mass simulator, so every sample respects the aircraft's limits. A genetic algorithm searching that latent space beats the teacher; a flow matching model distills the search into a 0.24 s generator, retrained once on the states the search met in flight (DAgger-like). A MacCready tactician picks the thermals; the planner flies every climb.

The GeneticWing chain in six stages: a teacher solves 20,923 climbs; a command VAE writes each climb as 16 numbers; a GA searches those numbers in the simulator; a flow-matching generator learns the GA's answers; it is retrained on the states met in flight; in a 20 km flight it proposes 64 climbs, flies each in the simulator and keeps the best

Same tasks, same tactician, different climbs

64 cross-country tasks of 20 km in synthetic drifting thermals, at two MacCready settings (0.5 / 1). Every pilot flies the same tasks with the same autopilot.

Climbs	Mean climb, m/s	Speed, km/h	Battery, kJ
Motor all the way		29.2	61.9
Circling rule	1.29 / 1.22	24.9 / 28.3	2.9 / 19.3
Latent GA, 1,000 flights per 60 s block	2.08 / 2.07	28.2 / 29.5	2.9 / 4.7
Generator distilled on single thermals only	1.50 / 1.17	25.1 / 28.8	4.2 / 20.2
Generator after one DAgger-like round, 0.24 s	2.23 / 2.20	28.5 / 30.2	2.9 / 4.7

At MacCready 1 the generator is as fast as the motor (30.2 against 29.2 km/h) with 13 times less battery, and no wing-section stall. One run per row; thermals are known exactly in these tasks.

One 20 km task flown with the circling rule and with planned climbs: path through the drifting thermals and altitude against the distance flown — One task with both pilots: path and drifting thermals (top), altitude against the distance flown toward the goal (bottom).

New · real forecast

Flying in the thermals of a real day over Issoudun

The synthetic thermals are replaced by ones built from the archived DWD ICON-D2 forecast (2.2 km) around Issoudun-Le Fay, for three days of 2025. The forecast surface heat flux and convective layer depth give the thermal strength w* point by point and hour by hour; following Allen's model (NASA), thermals are spaced by the layer depth, live 5 to 30 minutes, drift with the forecast wind, and the air sinks between them.

3D animation of a 20 km task over Issoudun on 12 July 2025: a blue sailplane climbs in translucent orange thermal columns that appear, drift and fade over ground coloured by the forecast thermal strength, while a grey sailplane flies on its motor — 12 July 2025 from 12:35, task 1. The blue sailplane uses the wind-aware circling rule (MacCready 1) and flies the 20 km in 53 minutes without its battery; the grey one uses the motor all the way, 49 minutes and 114 kJ. The ground is coloured by the forecast w*; thermals grow, drift with the 5 m/s wind and fade.

Left: the forecast w* over Issoudun on 12 July 2025 with the thermals alive at mid-flight and one flight path from start to goal. Right: w* and the convective layer depth at the site through the day for the three days — Left: forecast w* at mid-flight, the thermals alive then, and one path. Right: w* and the convective layer depth at the site through each day (parcel method; GFS dotted).

48 tasks per day	Arrived	Battery, kJ	km/h
12 Jul · strong blue thermals
Motor all the way	100 %	120.7	26.7
Circling rule, MacCready 1	100 %	1.3	24.0
2 Aug · cumulus day
Motor all the way	100 %	89.3	29.3
Circling rule, MacCready 1	100 %	0.5	28.5
22 Jun · windy, shallow layer
Motor all the way	100 %	94.7	36.7
Circling rule, MacCready 1	54 %	65.4	26.6

On the two good days the circling rule flies nearly every task without the motor, 1 to 4 km/h slower than the motor. On the windy day, with about 7 m/s of wind for an aircraft whose best-glide speed is 9 m/s, it climbs at 0.8 m/s and drifts a kilometre per climb: a day for the motor.

The tactician learned to count the wind: glides over the ground, the drift during a climb, a final glide only when it is really safe. The genetic search and the generator fly these days next.

Open this task in 3D: the forecast map hour by hour, every thermal, three pilots and a barogram.

GeneticWing: a soaring planner for a fixed-wing UAV

Six stages, from a slow teacher to a fast planner

Same tasks, same tactician, different climbs

Flying in the thermals of a real day over Issoudun

Fly it yourself

Issoudun, 12 July 2025

A 20 km cross-country flight

A 60 s climb

The whole chain, animated