GeneticWing: a generative planner for a sailplane's climbs in thermals

Simulated flight, first 8 km of 20. Red: climbs planned by the genetic search in the learned space. Grey: the standard circling rule. Same tactician, same thermals. Fly it in the browser (3D): 20 km flight · 60 s climb · real forecast over Issoudun

GeneticWing plans every thermal climb of a simulated sailplane flying cross-country.

• It climbs 1.7 times faster than the standard circling rule: 2.23 against 1.29 m/s on 64 simulated 20 km flights, with the same tactician choosing the thermals.
• It keeps the gain on 64 tasks no training step has seen: 2.21 against 1.30 m/s.
• It plans a climb in 0.24 s, 56 times faster than the genetic search it was distilled from (13.4 s per climb).

Chain: teacher, command VAE, genetic search in the VAE's latent space, flow-matching generator, one DAgger-like round. It extends ARPL's flatness-based glide planning (Morando et al., ICRA 2025) from gliding to soaring. Simulation only; the thermals are known.

Project site · Detailed results · Implementation · Slides, in French · Related: Fixed-Wing, WingJEPA

How it works

• A slow offline teacher (a circling law tuned by a genetic algorithm, then refined with SLSQP) solves 20,923 single-thermal climbs.
• A VAE learns the teacher's commands (lift coefficient, bank, thrust) and writes each climb as 16 numbers. Its decoder flies them through a point-mass simulator, so every sample respects the aircraft's limits.
• A genetic algorithm searching those 16 numbers beats the teacher, but takes about 13 s per climb.
• A flow-matching generator learns the search's answers. In flight it proposes 64 climbs, flies each in the simulator and keeps the best, in 0.24 s. One DAgger-like round retrains it on 1,095 states the genetic search met in flight.
• A MacCready tactician picks the thermals; the planner flies every climb. The autopilot keeps every wing section below 90 % of its stall.

In the thermals of a real day

Issoudun, 12 July 2025. Blue: the circling rule, 20 km in 53 minutes without the battery. Grey: the motor all the way, 49 minutes and 114 kJ. Ground coloured by the forecast thermal strength. Open it in 3D.

• The thermals come from the archived DWD ICON-D2 forecast over Issoudun-Le Fay for three days of 2025: the forecast heat flux and convective layer give their strength point by point and hour by hour (Allen's model). They grow, drift with the forecast wind and fade.
• On the two good days the circling rule flies 48 of 48 tasks at MacCready 1 almost without the motor (1.3 and 0.5 kJ on average, against 121 and 89 kJ on the motor), 1 to 3 km/h slower.
• On the windy day (about 7 m/s of wind for a 9 m/s best-glide speed) it arrives on half the tasks: a day for the motor. The genetic search and the generator have not flown these days yet.

Run it

Python 3.11, Windows 10 or Linux.

pip install -e ".[dev]"                                # numpy, scipy, matplotlib, torch, pytest
python -m pytest -q                                    # 69 tests, about 2 min
python scenarios/s0_glide.py --out results             # replays the ARPL straight glide, about 4 min
python scenarios/phase2_teacher.py --out results/phase2 --workers 8

Phases 3 to 6 run on a Slurm cluster: set slurm/config.sh, then bash slurm/submit_all.sh. All commands: docs/RESULTS.md. Code layout: docs/IMPLEMENTATION.md.

Limitations

• Simulation only: point-mass dynamics with a strip-theory wing in turns, no flight test.
• The thermals are known exactly; the planner never updates its belief in flight.
• One run per method, no spread over seeds.
• The airframe is estimated: with 20 % more mass or 20 % less maximum lift, the search's lead over the teacher almost vanishes.

Credits and license

Flatness-based planning, glide data and energy definitions: Morando et al., ARPL. Baseline GA and reactive law: T. Zohdi's course. Thermal model: M. J. Allen. Full list in docs/IMPLEMENTATION.md.

Clément Dumeril, clement.dumeril.net. MIT license (LICENSE); citation metadata in CITATION.cff.