PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.06046v1cs.ROcs.LGcs.SYeess.SY

TL;DR

Near-ground fixed-wing control is hard to benchmark because existing simulators do not readily combine realistic aerodynamics and subsystem models with scalable execution. FALCON-S provides a modular simulation and benchmarking framework for this regime; its controller comparisons vary by task, while its Warp backend supports high-throughput simulation.

  • Problem

    Existing simulators make it difficult to study ground effect, realistic actuator and sensor behavior, and controlled physical variations together, limiting comparisons across modeling assumptions, tasks, and control paradigms.

    Source: §I · p. 1, §II · p. 2

  • Method

    The framework combines 6DoF fixed-wing dynamics, semi-empirical ground-effect aerodynamics, configurable physical components, CPU/GPU backends, and tasks for classical and learning-based controllers.

    Source: §I · p. 1, §II · p. 2, B. Environment Module · p. 5

  • Results

    No controller ranks consistently across tasks: MPPI is most accurate where it converges but does not acquire commanded altitude, while PPO survives and acquires every reported altitude episode on both airframes.

    Source: C. Altitude Acquisition and Hold · p. 7, §VI · p. 12

  • Takeaways & Limitations

    At 16,384 environments, Warp reaches 19.5 million environment-steps per second, with simulation accounting for under 10% of training wall-clock time.

    Source: C. GPU Backend Performance · pp. 5–6

Abstract

from arXiv · show

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.

I. INTRODUCTION

Near-ground fixed-wing control is difficult to benchmark because existing simulators trade aerodynamic and subsystem realism against scalable execution. The paper introduces a modular fixed-wing benchmark combining these capabilities for systematic controller evaluation.

I. INTRODUCTION

  • Existing simulators often trade high-fidelity physics for scalability, leaving ground effect, realistic actuators and sensors, and controlled physical variations difficult to study together.This limits comparisons across modeling assumptions, task complexity, and control paradigms.Source: §I · p. 1, §II · p. 2
  • The framework supports systematic evaluation across fixed-wing airframes, low-altitude tasks, and model-based and learning-based controllers.Its task library includes altitude regulation, trajectory tracking, and attitude-command tracking.Source: §I · p. 1
  • Warp sustains 19.5 million environment-steps per second at 16,384 parallel environments, a measured 13.4× speed-up over the cited state of the art.The PyTorch implementation is reported to have single-step numerical parity with Warp.Source: §I · p. 1
  • Table I compares aircraft simulation frameworks on near-ground aerodynamics, modular tasks, controllers, and sensor and actuator models.Its symbols denote full, partial or optional, and unsupported features.Source: Table I · p. 2

II. RELATED WORK

  • The benchmark combines full 6DoF dynamics, semi-empirical ground-effect aerodynamics, actuator and sensor dynamics, wind turbulence, and CPU/GPU execution for classical and learning-based control.It also supports JSBSim validation and X-Plane closed-loop testing with high-fidelity rendering.Source: §I · p. 1, §II · p. 2

III. PRELIMINARIES

The preliminaries formulate fixed-wing flight as coupled 6DoF rigid-body dynamics and describe aerodynamic forces and moments using semi-empirical coefficients. The framework architecture separates the environment from the controller agent.

III. PRELIMINARIES

  • The vehicle state includes position, orientation, linear velocity, and angular velocity, with throttle and control-surface deflections as inputs.Orientation can be represented by a unit quaternion or Euler angles.Source: §III · p. 2
  • Newton–Euler equations describe translational and rotational motion under gravity, aerodynamics, and propulsion, assuming constant mass and neglecting gyroscopic effects.The body-fixed frame is attached at the center of mass, with motion referenced to an inertial NED frame.Source: §III · p. 2
  • Figure 2 separates the simulation environment, which models flight components, from an agent module for classical and learning-based controllers.Tasks, metrics, and visualization support extensible benchmarking and training.Source: Figure 2 · p. 3

A. Aerodynamic Model

  • Semi-empirical aerodynamic forces and moments depend on angle of attack, sideslip, Reynolds number, and control-surface deflections.Lift, drag, and side-force coefficients come from lookup tables or geometry-derived parametric expressions.Source: A. Aerodynamic Model · p. 3

B. Ground Effect Model

The ground-effect model applies semi-empirical coefficient corrections that capture near-surface lift and drag trends. The broader framework combines configurable physics with interchangeable backends and controller implementations.

B. Ground Effect Model

  • Semi-empirical ground-effect corrections parameterized by height-over-span ratio capture increased lift and reduced induced drag near the surface.The closed-form correction is GPU-compatible and can be toggled per environment for controlled ablations.Source: B. Ground Effect Model · p. 3
  • The simulator exposes time-continuous dynamics through configurable integration schemes, including Euler or RK4, with CPU and GPU implementations.These dynamics form the basis for controller-training environments.Source: B. Ground Effect Model · p. 3

IV. GROUND-EFFECT FIXED-WING FLIGHT SIMULATION FRAMEWORK

  • The environment combines a 6DoF flight model with actuator dynamics, sensor imperfections, Dryden turbulence, and optional ground effect across SciPy, PyTorch, and Warp backends.Physical components can be toggled independently for controlled ablations.Source: §IV · p. 3

A. Agent Module

  • The agent module supports LQR, MPPI, PPO, SAC, and TD3 for evaluation or large-scale training on CPU or GPU.MPPI samples control sequences and weights propagated trajectories by cumulative task-specific costs.Source: A. Agent Module · pp. 3–4
  • LQR uses a precomputed gain from a trim-condition linearization; its applicability depends on whether the airframe has a stabilizable linearization around trim.LQRI augments the state with integrator states before solving the Riccati equation.Source: A. Agent Module · pp. 3–4
  • Gymnasium-compatible CPU environments and batched Warp and Torch variants support vectorized PPO, SAC, and TD3 training and evaluation.The CPU environment also remains compatible with off-the-shelf libraries such as Stable-Baselines3.Source: A. Agent Module · p. 4

B. Environment Module

The environment module combines configurable physics backends with a modular 6DoF flight model, aerodynamic and actuator dynamics, and configurable environmental and sensor effects. It also supports multiple flight tasks and aircraft configurations.

B. Environment Module

  • Three interchangeable backends—SciPy, PyTorch, and Warp—support prototyping, GPU execution, and large-scale accelerated simulation.The PyTorch and Warp backends are validated to single-step numerical parity.Source: B. Environment Module · p. 4
  • The core physics module simulates full 6DoF rigid-body aircraft dynamics for low-altitude scenarios.Its focus is on settings where ground proximity and turbulence matter.Source: B. Environment Module · p. 4
  • Aerodynamic forces and moments vary with airspeed, attitude, control-surface deflection, and height above ground, with semi-empirical ground-effect corrections.The aerodynamic model uses precomputed coefficients from OpenVSP.Source: B. Environment Module · p. 4
  • Configurable first- or second-order actuator dynamics reproduce latency, saturation, and bandwidth-limited slew behavior.Servo surfaces use a second-order form; propulsion uses a first-order form, with throttle lag depending on powerplant class.Source: B. Environment Module · p. 4
  • Wind and Dryden turbulence perturb airspeed and thereby affect aerodynamic quantities; turbulence parameters are recalculated with altitude and airspeed.Wind and turbulence are disabled by default and can be enabled for disturbance-rejection evaluation.Source: B. Environment Module · p. 5
  • The sensor model can add noise, bias, scaling error, clipping, quantization, reduced sampling, and transport delay to flight measurements.Sensor routing is opt-in; attitude measurements are renormalized to remain valid rotations.Source: B. Environment Module · p. 5
  • Modular tasks include fixed-altitude keeping, path following, trajectory tracking, and attitude-command tracking, while aircraft configurations support switching among airframes.Physical components can be independently toggled or simplified for controlled ablations.Source: B. Environment Module · p. 5

C. GPU Backend Performance

The backend benchmark compares simulation throughput across parallel environment counts. Warp delivers the highest reported throughput, while simulation accounts for a small share of PPO training time.

C. GPU Backend Performance

  • Backend throughput was measured on identical Airship V7 altitude-task dynamics, averaging 10 CUDA-fenced trials after warm-up exclusion.Relative standard deviation was at most 0.3%.Source: C. GPU Backend Performance · p. 5, Table III · p. 6
  • Simulation accounts for under 10% of training wall-clock, and a PPO altitude policy completes 150 million environment steps in under five minutes.The remaining training wall-clock is set by gradient updates.Source: C. GPU Backend Performance · p. 6
  • Warp reaches 19.5M env-steps/s at 16k environments, three orders of magnitude above the saturated CPU machine and 7.5× the parity-validated PyTorch backend.Source: C. GPU Backend Performance · p. 5

D. Cross-Simulator Validation

Cross-validation compares the simulator with JSBSim under zero-input free flight and tests controller transfer to X-Plane. Agreement with JSBSim is near-perfect when ground effect is disabled, while the ground-effect condition diverges as expected.

D. Cross-Simulator Validation

  • X-Plane validation tests whether a controller tuned in the simulator retains stable tracking under an independent physics engine.The authors treat this as qualitative robustness evidence, not numerical validation.Source: D. Cross-Simulator Validation · p. 6
  • With ground effect disabled, altitude and pitch-rate trajectories agree with JSBSim at Pearson r ≥0.999 throughout.The comparison uses identical zero-input free-flight conditions.Source: Table IV · p. 6, D. Cross-Simulator Validation · p. 6
  • At h0 = 1 m with ground effect enabled, airspeed correlation with JSBSim is r = 0.948, consistent with its lack of ground-effect modeling.The authors describe the resulting divergence as expected and as isolating the ground-effect correction.Source: Table IV · p. 6, D. Cross-Simulator Validation · p. 6

E. Visualization Tools

The paper describes visualization tools for inspecting simulation outputs and experiment comparisons that vary modeling choices and task conditions. The experiments compare controllers on altitude-holding and attitude-command tracking tasks under shared conditions.

E. Visualization Tools

  • Python tools plot model states, aerodynamic quantities, actuator commands, and ground-effect indicators, and render animated 3D flight paths.The X-Plane interface also provides photorealistic renderings.Source: E. Visualization Tools · p. 6

V. EXPERIMENTS & RESULTS

  • Experiments compare five controllers across altitude holding and attitude-command tracking on Airship V7 and Volantex Ranger airframes.The five controllers are LQR, MPPI, PPO, SAC, and TD3.Source: §V · p. 6
  • Comparisons vary ground effect, altitude, sensor noise, and turbulence while holding airframe, actuator and sensor models, and reference stream identical within each comparison..Source: §V · p. 6

A. Task Definitions

Both tasks use a 6DoF plant with 10 ms integration, normalized and clipped observations, and actuator commands passed through servo and motor dynamics. The benchmark defines altitude-holding and attitude-based tracking tasks with different command and control interfaces.

A. Task Definitions

  • Each task runs on the 6DoF plant at 10 ms steps, with normalized, clipped observations and actions mapped to actuator ranges and passed through servo and motor dynamics.Actions are bounded in [−1, 1].Source: A. Task Definitions · p. 6
  • Altitude keeping regulates height to a commanded target while holding trim airspeed, using elevator and throttle actions.The reward combines altitude tracking with airspeed, action-smoothness, and safety shaping; episodes end on ground contact, stall, or the horizon.Source: A. Task Definitions · p. 7
  • Attitude-based tracking executes bank, climb-rate, and airspeed commands directly through all four actuators, without an intervening autopilot.The reward combines tracking terms for the three command channels with airspeed-margin and action-rate penalties.Source: A. Task Definitions · p. 7
  • Climb rate, rather than pitch attitude, is commanded because pitch does not determine altitude rate when airspeed varies.A pitch-tracking executor therefore cannot be composed into an altitude-holding outer loop.Source: A. Task Definitions · p. 7

B. Metrics

The section defines a shared scoring protocol and evaluates altitude acquisition separately from survival and steady-state accuracy. PPO survives and acquires every tested episode on both airframes, while other methods expose distinct trade-offs between acquisition, precision, and survival.

B. Metrics

  • The shared metrics separate survival, acquisition, steady-state accuracy, and transient behavior, with altitude acquisition defined by a 1 m capture band.Errors average only runs that remain inside the flight envelope; acquisition, settling time, and overshoot are meaningful because episodes begin off-reference.Source: B. Metrics · p. 7, Table V · p. 7

C. Altitude Acquisition and Hold

  • LQR is the most precise altitude controller when settled, but its acquisition is incomplete on the Volantex Ranger, showing that accuracy and acquisition rank methods differently.Its settled error is 2.8 × 10−7 m on the Airship V7 and 1.4×10−4 m on the Volantex Ranger; it acquires only half of the Ranger episodes.Source: C. Altitude Acquisition and Hold · p. 7
  • PPO survives and acquires every altitude episode on both airframes, whereas MPPI acquires almost none despite surviving 93% on the Airship V7.MPPI’s settled error there is 30.7 m and overshoot 41.5 m against a commanded change of at most 18 m; on the Volantex Ranger it survives 18% and acquires none.Source: C. Altitude Acquisition and Hold · p. 7, Table VI · p. 8
  • On Airship V7, SAC acquires 0.50 ± 0.41 with 3.52 ± 3.55 m settled error, versus PPO’s 1.00 ± 0.00 acquisition and 0.288 ± 0.022 m error.PPO’s tracking error collapses inside the ±1 m capture band after every target change and remains there.Source: C. Altitude Acquisition and Hold · p. 8

D. Attitude-based Maneuvers Tracking

The attitude executor is tested on four command-stream maneuvers generated for each airframe’s own envelope, and controllers are scored on achieved attitude and energy states rather than waypoint tracking. PPO meets the bank tolerance across all eight airframe–maneuver cells, while other methods trade off bank tracking, repeatability, and energy-channel performance.

D. Attitude-based Maneuvers Tracking

  • The four maneuvers are a sustained circle, figure-of-eight, climbing helix, and weaving s-turn, evaluated as command streams rather than position-space paths.Each stream uses its airframe’s own envelope, so maneuver names indicate comparable difficulty rather than identical geometry.Source: D. Attitude-based Maneuvers Tracking · p. 8, Table VII · p. 8
  • PPO stays inside the bank tolerance on all eight airframe–maneuver cells, with 0.42–2.28° RMSE and seed spread no greater than 0.21°.MPPI is more accurate on three V7 maneuvers where it converges, but fails to converge on other cells; out-of-envelope runs are excluded from Table VIII.Source: Table VIII · p. 8, D. Attitude-based Maneuvers Tracking · p. 9
  • Figure 6 presents ground tracks colored by achieved bank and commanded-versus-achieved channels for the steepest sustained turn.The comparison illustrates the attitude-command interface across bank envelopes and a sustained turn.Source: Figure 6 · p. 9
  • SAC and TD3’s bank tremor also degrades climb tracking, with 0.61–1.36 m/s error against 0.10–0.33 m/s for PPO and MPPI.Airspeed errors reverse by airframe: PPO has the largest error on V7 at 3.5 m/s but the smallest on Ranger at 1.1 m/s.Source: D. Attitude-based Maneuvers Tracking · p. 9

E. Ground Effect and Thrust demand

Ground effect changes the thrust required to hold a trajectory, while robustness varies by disturbance family and airframe. The results distinguish trim capability from policy-controlled performance and show that noise, latency, and turbulence have different effects.

E. Ground Effect and Thrust demand

  • At h/b = 0.25, trim thrust falls 25.4% on the Airship V7, 22.5% on the Airship A0S, and 16.6% on the Volantex Ranger, versus predictions of 20.6%, 15.2%, and 15.4%.Retrimming raises lift, lowers angle of attack, and reduces drag further than the fixed-lift prediction; settled tracking error remains below 0.19 m on the V7 and 0.03 m on the Ranger.Source: E. Ground Effect and Thrust demand · pp. 9–10, Table IX · p. 10
  • Policy-controlled thrust change is 29.1% on the Airship V7 and 12.1% on the Volantex Ranger, while each policy holds its altitude band.The Airship A0S offers a 22.5% trim reduction at h/b = 0.25, but its altitude policy does not hold that band; Fig. 9 has no A0S policy curve below h/b = 2.Source: Table IX · p. 10, E. Ground Effect and Thrust demand · p. 10, Figure 9 · p. 11

F. Robustness to Sensing, Latency and Turbulence

VI. CONCLUSIONS

The paper presents a fixed-wing simulation framework for evaluating learned and classical controllers across airframes and flight regimes. Its comparisons show no consistently best method, motivating further work on guidance and sim-to-real transfer.

VI. CONCLUSIONS

  • The framework lets learned and classical controllers drive different fixed-wing airframe models across varied flight regimes.Source: §VI · p. 12
  • Command-stream tasks evaluate controllers on the states they must continuously maintain, and learned methods are reported across three training seeds on both tracking tasks.Source: §VI · p. 12
  • Ground-effect reporting separates thrust measured for a freely flying airframe from thrust under policy-driven flight, while robustness is tested across four disturbance parameters and severities.Source: §VI · p. 12
  • No method ranks consistently: MPPI is most accurate when it converges but does not acquire commanded altitude, LQR is the most precise regulator but acquires half the episodes on one airframe, and PPO is most consistent.Future work adds a guidance layer above the attitude executor for sim-to-real transfer and considers ground-effect flight over waves.Source: §VI · p. 12

Related papers

Loading 2609.06046v1…