Blue: the world-model trunk (encoder, action encoder, GRU predictor). Orange: the JEPA target, used only in training. Green and gray: the readout heads. In flight only the trunk and one head run, restarted from the last measurement.
A small world model (TCN encoders, GRU predictor, about 11k parameters) that learns to predict how a simulated fixed-wing drone will fly from its recent states and the commands, so it can serve as a learned simulator. Independent project for Physics-Inspired Machine Learning, UC Berkeley, Fall 2026. Simulation only.
The project now runs the test suite of SkyJEPA (Rao et al., arXiv 2606.23444) on PX4's advanced_plane, the
fixed-wing model of the PX4 Gazebo simulation: the recipes of their Table III, the same metrics, and a closed-loop
test. The runs are in progress on Savio.
Ultra Stick study (archived) · Poster (PDF) · Implementation · Related: Fixed-Wing, GeneticWing
- Simulator:
wingjepa/sim/advanced_plane.py, a Python port of theadvanced_planeas Gazebo flies it (AdvancedLiftDrag with its post-stall blend, MulticopterMotorModel propeller, servos), checked against a C++ transliteration (tools/compare_advanced_cpp/). - Data:
configs/data_advanced.yaml, SkyJEPA's volume: 20,000 training flights of 10 s and 400 test flights of 60 s at 20 Hz, flown by a TECS autopilot, each on a random aircraft (mass, inertia, aerodynamics, motor, CG, servos); half of the test flights use aircraft outside the training ranges.data_advanced_turb.yamladds turbulence;data_advanced_upset.yamlhas 15 s flights, half of them starting in an upset (steep bank, nose high or low, slow or fast) that the autopilot recovers from, in gusty air. - Model: causal TCN encoders read 0.55 s of history, a GRU rolls the latent state forward with the commands, and a head turns it into the next state: an MLP, or a physics head that outputs aerodynamic coefficients and integrates Newton-Euler with RK4.
- Recipes (
scripts/train_a.py --mode):
| SkyJEPA, Table III | Here |
|---|---|
| Predictive | a0 --T 1 --T-max 1 |
| Predictive + physics regularization | a0 --T 1 --T-max 1 --lambda-phys 1 |
| Reconstruction + prober / physics prober | a0, then c_jepa --trunk <a0 run> |
| JEPA + prober / physics prober | jepa, then c_jepa --trunk <jepa run> |
| (extra) physics head trained end to end, plain MLP | c_direct, mlp_direct |
- Evaluation:
scripts/eval_skyjepa.py(position, velocity and attitude error, compounding ratio, straightening, observation noise) andscripts/closed_loop.py(MPPI with each model against the TECS autopilot).
Python 3.11, tested on Windows 10 (unit tests) and Linux (training):
pip install -e ".[dev]"
python -m pytest -q tests
python scripts/generate_flights.py --config configs/data_advanced.yaml --set train --shard 0 --n-shards 20
python scripts/check_dataset.py --dir data/data_advanced --set train --g 9.8
S="--H 11 --channels 8,8,16 --dilations 1,2,4 --z-dim 24 --aircraft-input 0 --action-encoder 1 --T 20"
P=configs/priors/advanced_plane_nul.json
python scripts/train_a.py --data-dir data/data_advanced --out runs/jepa_adv_s0 --mode jepa $S
python scripts/train_a.py --data-dir data/data_advanced --out runs/c_jepa_adv_s0 --mode c_jepa $S --prior $P \
--trunk runs/jepa_adv_s0
python scripts/eval_skyjepa.py --data-dir data/data_advanced --models runs/jepa_adv_s0 runs/c_jepa_adv_s0
python scripts/closed_loop.py --config configs/data_advanced.yaml --controllers tecs runs/c_jepa_adv_s0 \
--out runs/closed_loop/adv.jsonOn Savio: cluster/gen_data.sbatch, cluster/train_a.sbatch and cluster/smoke_advanced.sbatch, a quick check of
every recipe (cluster/README.md). scripts/idfds_to_npz.py converts real PX4 flights (IDF-DS)
to the same format.
The first study flew a simulated Ultra Stick 120 (a port of the University of Minnesota 6-DoF model with NASA
wind-tunnel tables). There, the physics head trained end to end predicted 5 s of flight within 5.8 m of position
error, against 8.1 m for SkyJEPA's frozen physics probe and 14.3 m for the same network without physics.
Poster, results, code and how to rerun it
(the tag ultrastick-final reproduces the poster).
- Simulation only, and the flights come from my autopilot, which shapes the states the models see.
- The physics head is given the true mass and inertia; the models without physics get no aircraft parameters.
- States are fully observed and the test noise is white.
Aircraft model: PX4's advanced_plane (PX4 Gazebo models). Ultra Stick study: University of Minnesota UAV Research
Group (UMNUAV_sim_v4, BSD) with NASA Langley wind-tunnel tables. Method from SkyJEPA (Rao et al., arXiv 2606.23444),
which I reproduced without its code (I am not one of its authors); SIGReg from LeJEPA (Balestriero and LeCun, arXiv
2511.08544). Full list in docs/IMPLEMENTATION.md.
Clément Dumeril, clement.dumeril.net. MIT license (LICENSE); citation metadata in CITATION.cff.
