FALCON-S: Fixed-wing Aerodynamics and Learning Control Suite

ICLR 2026 Conference Submission18705 Authors

Keywords: Flight Control, Optimal Control, Reinforcement Learning, ground-effect aerodynamics

TL;DR: FALCON is a high-fidelity, GPU-accelerated simulation benchmark for fixed-wing aircraft in ground effect, supporting both RL and classical control with modular physics, scalable training, and detailed flight control analysis.

Abstract: We introduce FALCON-S, a modular and high-fidelity framework for learning and control of fixed-wing aerial vehicles operating in ground effect. In contrast to existing aerial platforms with simplified dynamics, FALCON-S incorporates full 6DoF simulation alongside detailed modeling of ground-effect aerodynamics, actuator dynamics, and environmental disturbances. It offers a level of physical fidelity and modular component design that enables fine-grained manipulation and systematic analysis of low-altitude flight phenomena, capabilities rarely found in open-source or state-of-the-art simulation platforms. The framework includes both CPU and GPU simulation backends via Python and NVIDIA Warp, supporting high-throughput simulation across up to millions of parallel environments, which makes it suitable for reinforcement learning, sampling-based control algorithms, and large-scale evaluation. FALCON-S features a flexible architecture with interchangeable controllers, supporting optimal control, model-free and model-based RL, as well as a suite of flight control tasks such as altitude regulation and trajectory tracking. We include optional interfaces for validation and comparison through MATLAB/Simulink and XPlane, making it compatible with both engineering workflows and commercial simulators. The framework is released as open-source to facilitate reproducibility and enable controlled benchmarking in realistic flight scenarios.

Primary Area: datasets and benchmarks

Submission Number: 18705