Planax: A JAX-Accelerated High-Fidelity Reinforcement Learning Benchmark for Large-Scale Fixed-Wing Aerial Swarms
IEEE Robotics and Automation Letters · 2026-12 · 11 (12) · 14660-14667
DOI: 10.1109/lra.2026.3746054 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Robotics and Automation Letters, 2026-12. Verfasst von Chuanyi Xue und 8 weiteren Autor:innen. Die Arbeit nennt 16 Literaturangaben. Gefördert von: National Science and Technology Innovation 2030 - Major Project.
Abstract (Originalsprache)
Multi-agent reinforcement learning (MARL) for fixed-wing aerial robots requires both aerodynamic fidelity and high-throughput simulation, yet existing platforms typically provide either simplified kinematics or CPU-bound high-fidelity models. We present Planax, a JAX/XLA-accelerated, GPU-resident benchmark framework that combines nonlinear six-degree-of-freedom (6-DOF) dynamics with wind-tunnel-derived F-16 aerodynamic lookup tables using tensorized interpolation. On a single NVIDIA A100 GPU, Planax achieves $2.33 \times 10^{7}$ environment steps per second for independent heading/pitch/velocity rollouts and supports batch sizes of up to $2 \times 10^{7}$ independent 6-DOF instances. These figures characterize independent rollouts; interacting formation and pursuit-evasion tasks additionally incur relative-geometry costs. Under matched model, interpolation, and rollout settings, the XLA-compiled JAX Tensor-LUT backend achieves lower steady-state latency than the evaluated same-model PyTorch baseline across the tested components and batch sizes. Cross-fidelity multi-agent ablations, with all policies evaluated in the high-fidelity environment, show that high-fidelity training improves both task performance and safety. Across three seeds, the pursuit-evasion win rate increases from $0.534\pm 0.097$ to $1.000\pm 0.000$ , while formation slot RMSE decreases from $1493\pm 130$ m to $445\pm 211$ m and overload incidence decreases from $0.480\pm 0.049$ to $0.067\pm 0.071$ .
Förderung
- National Science and Technology Innovation 2030 - Major Project
Zitieren
APA
Xue, C., Dong, Q., Liu, Q., Qi, Y., Qin, X., Ren, J., Zhou, Z., Liang, B., & Yang, J. (2026). Planax: A JAX-Accelerated High-Fidelity Reinforcement Learning Benchmark for Large-Scale Fixed-Wing Aerial Swarms. IEEE Robotics and Automation Letters, 11(12), 14660-14667. https://doi.org/10.1109/lra.2026.3746054
Vancouver
Xue C, Dong Q, Liu Q, Qi Y, Qin X, Ren J, et al. Planax: A JAX-Accelerated High-Fidelity Reinforcement Learning Benchmark for Large-Scale Fixed-Wing Aerial Swarms. IEEE Robotics and Automation Letters. 2026;11(12):14660-14667. doi:10.1109/lra.2026.3746054
BibTeX
@article{Xue2026,
author = {Xue, Chuanyi and Dong, Qinyu and Liu, Qihan and Qi, Yang and Qin, Xinyao and Ren, Jinsheng and Zhou, Zhiyang and Liang, Bin and Yang, Jun},
title = {Planax: A JAX-Accelerated High-Fidelity Reinforcement Learning Benchmark for Large-Scale Fixed-Wing Aerial Swarms},
journal = {IEEE Robotics and Automation Letters},
year = {2026},
volume = {11},
number = {12},
pages = {14660--14667},
doi = {10.1109/lra.2026.3746054}
}
EUVIMED