GeneticWing: soaring trajectory planning for a fixed-wing UAV with genetic search in a learned latent space

Phase-by-phase results · Implementation · Results slides, in French · 3D flight viewers (HTML, open locally) · Related: Fixed-Wing, WingJEPA

A simulated SB-XC sailplane plans its thermal climbs with a flow matching generator distilled from a genetic search over a learned command space. On 64 cross-country tasks of 20 km it climbs at 2.23 m/s against 1.29 m/s for a circling rule with the same tactician (MacCready 0.5), at 0.24 s per plan; one run per method, thermals known.

Method

Started October 2026, simulation only. It extends the flatness-based glide planning of Morando et al. at ARPL (ICRA 2025, arXiv 2605.15619) from efficient gliding to soaring. A slow offline teacher (reactive law + SLSQP) solves thousands of single-thermal situations. A VAE learns the teacher's commands (C_L, 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 GA into a 0.24 s planner, and one DAgger round retrains it on the states met in flight. A MacCready tactician picks the thermals; the planner flies every climb.

Results

One 20 km task, same tactician: the circling rule (grey) and climbs planned by the latent GA (green), motor off in the climbs. Top: path and the drifting thermals; bottom: altitude against the distance flown toward the goal.

Climbs (64 tasks of 20 km, MacCready 0.5 / 1)	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, about 13 s per plan	2.08 / 2.07	28.2 / 29.5	2.9 / 4.7
Generator distilled on single-thermal situations only	1.50 / 1.17	25.1 / 28.8	4.2 / 20.2
Generator after one DAgger round, 0.24 s per plan	2.23 / 2.20	28.5 / 30.2	2.9 / 4.7

One cluster run per row; source: docs/RESULTS.md, phases 5 and 6 (raw outputs not committed). Every pilot flies the same tasks with the same tactician and autopilot. At MacCready 1 the DAgger generator is as fast as flying on the motor (30.2 against 29.2 km/h) with 13 times less battery, with no wing-section stall. The single-thermal comparison with the teacher and how the chain was found: docs/IMPLEMENTATION.md.

Install and quick start

Tested with Python 3.11 on Windows 10 and on Linux. Dependencies are not pinned.

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

The phase 3 to 6 studies run on a Slurm cluster: set the account and partitions in slurm/config.sh, then bash slurm/submit_all.sh (all commands: docs/RESULTS.md).

Code structure

Path	Content
src/geneticwing/	flat planning core, aircraft and strip wing, wind models, simulators, GA and NLP planners, VAEs, students, cross-country pilot
scenarios/	one script per phase (s0_glide.py to phase6_distill.py), figures, 3D viewers
slurm/	cluster jobs, chained by submit_all.sh
tests/	69 tests
docs/	results, implementation, figures, slides, viewers

Limitations

• Simulation only: point-mass dynamics with a strip-theory wing in turns, not 6-DOF, and no flight test.
• Airframes are estimates (SB-XC polar, C_L max, mass); the gain over the teacher shrinks to about 0.1 when the stall margin is reduced (mass or C_L max off by 20 %).
• One run per method on the cluster studies: no spread over seeds is reported.
• On the cross-country tasks the thermals are known exactly; the planner's belief is never updated in flight.
• The raw outputs of the cluster runs are not in the repository, so phases 3 to 6 can be checked only by rerunning.

Credits and citation

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.