FALCON-S: Fixed-wing ground-effect Aerodynamics Simulator and Flight Control Learning Suite

Matteo El Hariry; Pedro Lima; Andrej Orsula; Matthieu Geist; Miguel Olivares-Mendez

cs.ROSep 5, 2026

FALCON-S: Fixed-wing ground-effect Aerodynamics Simulator and Flight Control Learning Suite

Authors: Matteo El Hariry, Pedro Lima, Andrej Orsula, Matthieu Geist, Miguel Olivares-Mendez

Organizations: Interdisciplinary Centre for Security, Reliability and Trust (SnT), University of Luxembourg, L-1855 Luxembourg · Earth Speacies Project

Abstract

We present a modular, high-fidelity simulation framework for the development and benchmarking of flight control strategies in fixed-wing aerial robots operating near the ground. Unlike existing simulators that rely on simplified or hover-oriented dynamics, our framework models full 6DoF rigid-body physics, semi-empirical ground-effect aerodynamics, actuator dynamics, sensor noise, and environmental disturbances. This physical realism, combined with modular component design, enables systematic analysis of low-altitude flight behavior under realistic conditions. The simulator supports both CPU and GPU backends via Torch and NVIDIA Warp, enabling high-throughput parallel execution suitable for large-scale reinforcement learning training and optimal control rollouts. A unified interface accommodates a range of controllers (both RL and optical control algorithms) across tasks such as altitude regulation and trajectory tracking. Cross-validation with X-Plane and JSBSim is also supported to facilitate engineering integration and visual fidelity.

Figures & tables

Fig. 1 : Airship flying in ground effect, rendered in X-Plane.

Feature	Ours	NeuralPlane	QPlane	JSBSim	XPlane
Open-source	✓	✓	✓	✓	–
Physics-based FDM	✓(WIG, 6DoF)	✓(fixed-wing only)	✓(JSBSim/XPlane)	✓	✓
Ground Effect Model	✓(semi-empirical)	–	* (depends on JSBSim)		✓
GPU Acceleration	✓(Warp)	✓(PyTorch)	–	–	–
Multi-agent Support	–	✓	✓		* (via UDP)
Multiple Flight Tasks	✓	✓	✓

TABLE I: Comparison with existing aircraft simulation frameworks. Our platform combines near-ground aerodynamics, modular tasks, advanced controllers, rich sensor and actuator models. [✓] full, [*] partial/optional, [–] not supported.

Fig. 2: Ground-effect fixed-wing simulation architecture. The modular environment models aerodynamics, actuators, ground effect, turbulence, and sensors, while the agent module supports classical and learning-based controllers. Tasks, metrics, and visualization enable extensible benchmarking and training.

	Longitudinal	Lateral	Vertical
Filter	$\displaystyle G_{u}(s)=\dfrac{\sigma_{u}K_{u}}{(1+T_{u}s)^{2}}$	$\displaystyle G_{v}(s)=\dfrac{\sigma_{v}K_{v}(1+\sqrt{3}T_{v}s)}{(1+T_{v}s)^{2}}$	$\displaystyle G_{w}(s)=\dfrac{\sigma_{w}K_{w}(1+\sqrt{3}T_{w}s)}{(1+T_{w}s)^{2}}$
Constants	$K_{u}=\sqrt{\dfrac{2L_{u}}{\pi U_{0}}},\ T_{u}=\dfrac{L_{u}}{U_{0}}$	$K_{v}=\sqrt{\dfrac{L_{v}}{\pi U_{0}}},\ T_{v}=\dfrac{L_{v}}{U_{0}}$	$K_{w}=\sqrt{\dfrac{L_{w}}{\pi U_{0}}},\ T_{w}=\dfrac{L_{w}}{U_{0}}$

TABLE II : Dryden turbulence velocity spectral filters.

Panel of Figure 5 — Fig. 3 : Illustration of sensor imperfections supported by the simulator, including noise, bias and scaling.

Backend	Envs	Env-steps/s	Speedup vs. CPU
Python/SciPy (CPU)	1	589	$1\times$
PyTorch (GPU)	16,384	2,586,610	$4{,}390\times$
Warp (GPU)	16,384	19,461,717	$\mathbf{33{,}029\times}$
Warp (GPU)	262,144	73,640,000	peak
Warp (GPU)	1,048,576	67,070,000	saturated

TABLE III : Simulation throughput by backend and parallel-environment count (mean over 10 trials).

Fig. 4 : Throughput vs. parallel environments for the Warp, PyTorch, and CPU backends (log-log).

Condition	Altitude	Pitch rate	Airspeed
$h_{0}=1$ m, GE on	0.9999	0.9989	0.9480
$h_{0}=1$ m, GE off	0.9999	0.9999	0.9999
$h_{0}=10$ m	1.0000	0.9885	0.9997
$h_{0}=100$ m	0.9996	0.9975	0.9973

TABLE IV : Pearson correlations between our simulator and JSBSim under identical zero-input free-flight conditions. GE: ground effect.

Metric	Formula / description
Survival	fraction of episodes reaching the horizon without
	crash or stall termination
Acquisition	fraction additionally reaching and holding the
	target to within $\delta$
RMSE	$\sqrt{\frac{1}{\|W\|}\sum_{t\in W}e_{t}^{2}}$ over the settled window $W$
Time-in-band	fraction of scored time with $\|e_{t}\|$ inside the

TABLE V : Evaluation metrics computed from trajectory and control logs.

e_{t}

is the tracking error at step

t

, and

\delta

the capture band,

1

m on the altitude task.

Panel of Figure 10 — Fig. 5 : PPO holding a sequence of altitude targets, Airship V7 (left) and Volantex Ranger (right). Each panel set shows the flight paths, commanded versus achieved altitude for ten episodes, the tracking error over 256 Monte Carlo episodes with its 10–90% band, a side view, and the elevator and throttle traces of one episode. A new target is released only once the current one is held, so each error spike is one acquisition.

	Alg.	$n$	Surv.	Acq.	RMSE (m)	$t_{s}$ (s)	Over. (m)
Airship V7	PPO	200	1.0 $\pm$ 0.0	1.0 $\pm$ 0.0	0.3 $\pm$ 0.0	7.8 $\pm$ 0.4	0.8 $\pm$ 0.0
	SAC	200	0.8 $\pm$ 0.2	0.5 $\pm$ 0.4	3.5 $\pm$ 3.6	6.9 $\pm$ 0.7	0.2 $\pm$ 0.1
	TD3	200	0.9 $\pm$ 0.2	0.9 $\pm$ 0.2	0.1 $\pm$ 0.1	8.3 $\pm$ 1.1	0.6 $\pm$ 0.4
	LQR	50	1.0	1.0	$<0.1$	10.0	1.9
	MPPI	50	0.9 $\pm$ 0.0	0.0 $\pm$ 0.0	30.7 $\pm$ 0.3	—	41.5 $\pm$ 0.4
Volantex	PPO	200	1.0 $\pm$ 0.0	1.0 $\pm$ 0.0	0.0 $\pm$ 0.0	7.3 $\pm$ 0.0	0.7 $\pm$ 0.0

TABLE VI : Altitude acquisition and hold. Surv. is survival rate, Acq. is successful acquisition, RMSE is altitude error,

t_{s}

is settling time, and Over. is overshoot.

Fig. 6 : Attitude-command interface for Airship V7 (left pair) and Volantex Ranger (right pair). Left: ground tracks across the commanded bank envelopes, coloured by achieved bank. Right: commanded (dashed) and achieved (solid) channels for the steepest sustained turn, normalised by their respective ranges.

Fig. 7 : Bank tracking on Airship V7. Rows show the four maneuvers; columns show the five executors. Dashed lines show commanded bank, with the $\pm 8.6^{\circ}$ tolerance shaded; panels report RMSE and time-in-band. Faceting reveals LQR departures, SAC tremor, and the TD3 limit cycle, which overlap other traces on shared axes.

Fig. 8 : Completed maneuvers flown by the attitude executor, rendered from the simulator’s 3D visualisation. Rows are the two airframes, columns the four maneuvers, and the track is coloured by bank angle over $\pm 50^{\circ}$ , blue left and red right. Note the axis scales: a maneuver name denotes a comparable difficulty, not a comparable geometry (Table VII ).

Airframe	Maneuver	$\phi_{\max}$ ( ∘ )	$R$ (m)	$n$
Airship V7	circle	36.0	110	1.23
	figure-8	35.8	111	1.23
	helix	21.6	202	1.07
	s-turn	25.2	170	1.10
Volantex Ranger	circle	28.0	43	1.13
	figure-8	27.9	43	1.13

TABLE VII : Maneuver definitions. Each stream uses the airframe’s own envelope, so

\phi_{\max}

, turn radius

R

, and load factor

n=1/\cos\phi_{\max}

vary between airframes. Commands are identical for all five controllers.

	Alg.	Circle	Fig.-8	Helix	S-turn
Airship V7	PPO	1.8 $\pm$ 0.2 (100)	2.2 $\pm$ 0.1 (100)	0.8 $\pm$ 0.2 (100)	2.3 $\pm$ 0.1 (100)
	SAC	13.0 $\pm$ 1.4 (49)	20.9 $\pm$ 12.8 (47)	7.1 $\pm$ 2.4 (83)	12.7 $\pm$ 4.9 (58)
	TD3	23.6 $\pm$ 6.4 (18)	18.0 $\pm$ 5.6 (33)	11.1 $\pm$ 3.2 (49)	12.1 $\pm$ 2.1 (49)
	LQR	10.6 (76)	9.1 (77)	2.0 (100)	5.2 (100)
	MPPI	4.8 $\pm$ 6.9 (88)	0.7 $\pm$ 0.3 (100)	0.3 $\pm$ 0.2 (100)	0.2 $\pm$ 0.1 (100)
Volantex	PPO	1.0 $\pm$ 0.0 (100)	1.8 $\pm$ 0.0 (100)	0.4 $\pm$ 0.0 (100)	2.0 $\pm$ 0.0 (100)

TABLE VIII : Bank-angle RMSE (deg), mean

\pm

standard deviation, with time-in-band (%) in brackets. Out-of-envelope runs are excluded.

Fig. 9 : Ground effect versus $h/b$ . Left: thrust change for trim (dash-dot), policy (solid), and theory (dotted). Right: settled altitude error with correction on/off (solid/dashed). Colours denote airframes. A0S has no policy curve below $h/b=2$ .

	$h/b$	$h$ (m)	$T_{\mathrm{GE}}$ (N)	$T_{\mathrm{no\,GE}}$ (N)	$\Delta T$	$\Delta T_{\pi}$	Theory
Airship V7	0.25	1.25	54.39	72.89	$-25.4$ %	$-29.1$ %	$-20.6$ %
	0.50	2.50	66.32	72.89	$-9.0$ %	$-11.8$ %	$-6.9$ %
	1.00	5.00	70.92	72.89	$-2.7$ %	$-3.6$ %	$-2.1$ %
	2.00	10.00	72.49	72.89	$-0.6$ %	$-0.8$ %	$-0.5$ %
	3.00	15.00	72.81	72.89	$-0.1$ %	$-0.1$ %	$-0.1$ %
	6.00	30.00	72.90	72.90	$+0.0$ %	$+0.0$ %	$+0.0$ %

TABLE IX : Ground-effect impact on thrust.

\Delta T

is the thrust change;

\Delta T_{\pi}

is the policy-controlled change; Theory is the lifting-line prediction.

Airframe	Disturbance	Nom.	Mild	Med.	Sev.
Airship V7	Sensor noise	1.00/2.0	1.00/2.0	1.00/2.0	1.00/2.0
	Sensor delay	1.00/2.0	1.00/7.0	1.00/12.0	1.00/20.6
	Action delay	1.00/2.0	1.00/7.2	1.00/14.5	1.00/21.7
	Turbulence	1.00/2.0	1.00/7.7	0.75/18.7	1.00/22.0
Volantex Ranger	Sensor noise	1.00/1.3	1.00/1.3	1.00/1.3	1.00/1.4
	Sensor delay	1.00/1.3	1.00/1.3	1.00/1.3	1.00/9.7

TABLE X : PPO attitude-executor robustness. Entries are Surv./

\phi

(bank RMSE, ∘ ); values use the true simulator state.

Fig. 10 : Bank RMSE of the PPO executor against disturbance severity, one line per family, for the two airframes. Survival is annotated only where it leaves 1.00.

Robotics Simulation

Robotics simulation builds virtual robots and environments to model, train, and evaluate autonomous systems before real-world deployment.

See all 320 papers

UWB Meets Crazyflow: Simulating Degraded Feedback at Scale for Aerial Robotics
Oct 6, 2026Martin Schuck, Marcel P. Rath, Yufei Hua +3Aerial Robotics Robotics Simulation
Learning Systems and Robotics Lab, Technical University of Munich, Germany
CARLA-Air: Fly Drones Inside a CARLA World -- A Unified Infrastructure for Air-Ground Embodied Intelligence
Mar 30, 2026Tianle Zeng, Yanci Wen, Hong ZhangRobotics Simulation Autonomous Driving Simulation
1Shenzhen Key Laboratory of Robotics and Computer Vision, Southern University of Science and Technology.
Dynamic System Emulation: Fixed Wing Dynamics on a Multicopter
Sep 6, 2026Abdelhakim Amer, Andriy SarabakhaAerial Robotics Robotics Simulation
Department of Electrical and Computer Engineering, Aarhus University, Denmark

Oct 6, 2026J/K move · Enter open · S save

UWB Meets Crazyflow: Simulating Degraded Feedback at Scale for Aerial Robotics

Martin Schuck, Marcel P. Rath, Yufei Hua, Abhishek Goudar+2

Learning Systems and Robotics Lab, Technical University of Munich, Germany

Aerial Robotics Robotics Simulation Differentiable Physics

Open paper AskPDF arXiv

In this work, we introduce Crazyflow, an accurate, differentiable simulator built on JAX. By leveraging jit compilation via XLA, Crazyflow unifies physics and control into a single differentiable computation graph, enabling massive parallelization on accelerated hardware without sacrificing modeling accuracy. This architecture achieves order-of-magnitude speedups over existing baselines, capable of training deployable reinforcement learning agents in seconds. To highlight its highly modular design, we demonstrate how easily Crazyflow can be extended by integrating a complete, high-fidelity Ultra-Wideband (UWB) and Inertial Measurement Unit (IMU) simulation pipeline coupled with a full-state Extended Kalman Filter (EKF). This capability allows for massive parallel controller evaluation under realistic, degraded state feedback with minimal impact on GPU throughput. By combining speed, accuracy, and extensibility, Crazyflow serves as a foundational tool for the next generation of aerial robotics research.

Abstract

Robotics Simulation

Explore similar work

UWB Meets Crazyflow: Simulating Degraded Feedback at Scale for Aerial Robotics

CARLA-Air: Fly Drones Inside a CARLA World -- A Unified Infrastructure for Air-Ground Embodied Intelligence

Dynamic System Emulation: Fixed Wing Dynamics on a Multicopter