ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace

ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace
Alexander Prutsch, David Schinagl, Horst Possegger Graz University of Technology
ICRA 2026

Getting Started

Create and Activate Virtual Environment

conda create -n ascent python=3.11
conda activate ascent

Install PyTorch

Our implementation was primarily developed and tested using PyTorch 2.8.0 and CUDA 12.8.
The codebase is flexible across different configurations; for example, it has been successfully tested on PyTorch 1.11.0+cu113 and PyTorch 2.9.1+cu126.

Install PyTorch e.g.

pip install torch==2.8.0 torchvision --index-url https://download.pytorch.org/whl/cu128

Install Dependencies

pip install -r ./requirements.txt

TrajAir Dataset Setup

Download the Trajair dataset from the official site.
Unzip the files in the dataset folder.
The first time a dataset is loaded with a specific configuration (history length, future length, and sampling step), a cached file is saved to dataset/_cache to accelerate future loading times.

Tartan Aviation Dataset Setup

** Instructions comming soon. **

Training

Get available parameters: python train.py -h

Run training on 7days1 split with default parameters: CUBLAS_WORKSPACE_CONFIG=:4096:8 && python train.py --dataset_name 7days1
For each run, a new directory is created in runs/ containing checkpoints, log files, model configurations, and a copy of the model implementation to ensure reproducibility.

Evaluation

Pretrained Models

We provide pretrained checkpoints for TrajAir split days1-4 with 16s history as input (lower group in paper Table 1) in the checkpoints folder.
Run evaluation:
python test.py --exp_folder checkpoints/model_7days1/ --epoch 11

Expected results:

minADE5	minFDE5
0.313	0.562

Custom Runs

To evaluate a custom model, select its experiment folder and choose an epoch. You can optionally select a dataset split; if left blank, the split defaults to the one defined in the training configuration.
python test.py --exp_folder runs/2026-XX-XX_XX-XX-XX/ --dataset_name 7days1 --epoch 10

Visualization

Visualize TrajAir prediction results:
python visualize.py --exp_folder checkpoints/model_7days1/ --epoch 11

Bibtex

@inproceedings{prutsch2026ascent,
    title={{ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace}},
    author={Prutsch, Alexander and Schinagl, David and Possegger, Horst},
    booktitle={In Proceedings of the IEEE International Conference on Robotics and Automation},
    year={2026}
}

Acknowledgements

This repository is based on TrajAirNet. We thank them for their work!