Landing Trajectory Prediction for UAS Based on Generative Adversarial Network

Jun Xiang, Drake Essick, Luiz Gonzalez Bautista, Junfei Xie, Jun Chen

TL;DR

Researchers at San Diego State University have developed an LSTM-based Generative Adversarial Network (GAN) to predict the landing trajectories of Unmanned Aircraft Systems (UAS). By training on a newly created dataset of over 2,600 real-world pilot-controlled flights, the model significantly outperforms traditional Gaussian Mixture Regression (GMR) in handling the "messy," non-linear behaviors of human pilots.

What Problem Does This Paper Address?

As the sky becomes increasingly crowded with drones—with the FAA forecasting nearly 2 million units by 2026—managing traffic near "vertiports" (UAS landing hubs) is becoming a major safety challenge. To prevent mid-air collisions, Conflict Detection and Resolution (CD&R) systems need to know where a drone is going before it gets there.

Existing trajectory prediction methods often fall into two traps: they are either too slow for real-time "online" use, or they rely on rigid physics-based models that fail to account for the unpredictable nature of human pilots or environmental turbulence. This paper seeks a middle ground: a model that is fast, accurate, and capable of learning the "subtle logic" of real-world flight paths.

Methods

The authors employ a Generative Adversarial Network (GAN) architecture, specifically utilizing Long Short-Term Memory (LSTM) units to handle the sequential nature of flight data.

  • The Generator: Inspired by "Social-GAN" architectures, the generator uses an LSTM encoder to process the last 10 points of a trajectory. A pooling module then combines these features with environmental factors before an LSTM decoder predicts the next 10 points.
  • The Discriminator: Instead of a mathematical formula, a second LSTM network acts as a "critic." It learns to distinguish between real flight paths and the "fake" ones produced by the generator, forcing the generator to produce increasingly realistic trajectories.
  • The Baseline: The model was compared against Gaussian Mixture Regression (GMR), a popular probabilistic method that models trajectories as a collection of Gaussian distributions.

Datasets

The study utilized three distinct datasets to test different flight behaviors:

  1. Vertical Landing (Simulated): Drones fly horizontally to a point above the target, then drop vertically.
  2. Linear Landing (Simulated): Drones fly in a direct, straight line from their start point to the destination.
  3. Real UAV Landing Dataset: This is a major contribution of the paper. The team collected over 2,600 trajectories using a customized DJI Tello drone. Real pilots manually controlled the drones, and their paths were recorded with high precision using a motion capture system.

Key Insights and Ideas

  • The "Real-World" Advantage: In simulated datasets (Vertical and Linear), the GAN only beat the baseline GMR for the first few points. Because simulated data is mathematically "clean," the Gaussian-based GMR handled it well. However, on the Real Dataset, the GAN outperformed GMR across the entire trajectory.
  • Mining Hidden Rules: The authors argue that real pilots control drones with a level of randomness and subtle logic that traditional models can't capture. The GAN’s strength lies in its ability to "mine" these hidden patterns from raw data without needing an explicit physical equation.
  • The Discriminator as an Evaluator: Beyond training the generator, the authors found that a well-trained discriminator can be used independently to "score" the safety or quality of any given trajectory prediction.

Key Results

The primary metric used was Average Displacement Error (ADE), measured in meters for simulated data and normalized units for real data.

  • Accuracy on Real Data: For the real-world dataset, the GAN consistently showed lower error than GMR. For example, at the first predicted point, the GAN had an ADE of 0.10 vs. GMR’s 0.15. By the tenth point, the GAN maintained a lead with 0.18 vs. 0.24.
  • Short-term vs. Long-term: The GAN is exceptionally strong at short-term prediction (the first 4 points). However, the authors noted that as the prediction window lengthens, the error accumulates, a common challenge in recursive neural networks.
  • Discriminator Success: The discriminator successfully assigned high positive scores (avg ~2.32) to real trajectories and significantly lower or negative scores to "fake" ones, proving it had learned the underlying distribution of safe landings.

Limitations and Future Work

The authors identify a clear "drift" in accuracy: the further into the future the model tries to predict, the less reliable it becomes. This suggests the current LSTM-GAN framework might not yet be suitable for long-term flight planning.

For future iterations, the team suggests exploring more advanced neural network encoders and decoders (such as Transformers) that can process larger amounts of contextual information to maintain accuracy over longer durations.

Summary updated

Abstract

Models for trajectory prediction are an essential component of many advanced air mobility studies. These models help aircraft detect conflict and plan avoidance maneuvers, which is especially important in Unmanned Aircraft systems (UAS) landing management due to the congested airspace near vertiports. In this paper, we propose a landing trajectory prediction model for UAS based on Generative Adversarial Network (GAN). The GAN is a prestigious neural network that has been developed for many years. In previous research, GAN has achieved many state-of-the-art results in many generation tasks. The GAN consists of one neural network generator and a neural network discriminator. Because of the learning capacity of the neural networks, the generator is capable to understand the features of the sample trajectory. The generator takes the previous trajectory as input and outputs some random status of a flight. According to the results of the experiences, the proposed model can output more accurate predictions than the baseline method(GMR) in various datasets. To evaluate the proposed model, we also create a real UAV landing dataset that includes more than 2600 trajectories of drone control manually by real pilots.
BibTeX
@misc{xiang2024landing,
  author = {Jun Xiang and Drake Essick and Luiz Gonzalez Bautista and Junfei Xie and Jun Chen},
  title = {{Landing Trajectory Prediction for UAS Based on Generative Adversarial Network}},
  year = {2024},
  eprint = {2411.14403},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.14403},
}
Figure