Amelia: A Large Dataset and Model for Airport Surface Movement Forecasting

Abstract: The growing demand for air travel necessitates advancements in air traffic management technologies to ensure safe and efficient operations. Predictive models for terminal airspace can help anticipate future movements and traffic flows, enabling proactive planning for efficient coordination, collision risk assessment, taxi-out time prediction, departure metering, and emission estimations. Although data-driven predictive models have shown promise in tackling some of these challenges, the absence of large-scale curated surface movement datasets in the public domain has hindered the development of scalable and generalizable approaches. In this context, we propose the Amelia framework, which consists of four key contributions. First, Amelia-48, a large dataset of airport surface movement collected through the FAA's System Wide Information Management (SWIM) Program. This dataset includes over two years' worth of trajectory data (~70TB) across 48 US airports and map data. Second, we develop AmeliaTF, a large transformer-based baseline for multi-agent, multi-airport trajectory forecasting. Third, we propose Amelia-10, a training and evaluation benchmark consisting of 292 days of post-processed data from 10 different airports and a series of experiments to promote the development of foundation models in aviation. We provide baseline results across our benchmark using AmeliaTF. Finally, we release our framework and tools to encourage further aviation research in the forecasting domain and beyond at https://ameliacmu.github.io

• The paper presents the Amelia dataset and a transformer-based model for accurately forecasting airport trajectories across 48 US airports.
• It leverages an ego-centric scene representation with multiple attention layers to extract nuanced temporal and interaction features.
• Experimental results demonstrate that increased training diversity significantly improves generalization, setting a new benchmark in surface movement prediction.

Overview of "Amelia: A Large Model and Dataset for Airport Surface Movement Forecasting"

The paper "Amelia: A Large Model and Dataset for Airport Surface Movement Forecasting" presents significant contributions to the domain of airport surface movement prediction. Authored by researchers at The Robotics Institute, Carnegie Mellon University, the study addresses the escalating complexities within terminal airspaces due to increasing air travel demands. The researchers tackle these issues by introducing two primary contributions: a large-scale dataset, named Amelia, and a transformer-based trajectory forecasting model, referred to as the Amelia model.

Dataset Description

The Amelia dataset is curated from the FAA's System Wide Information Management (SWIM) Surface Movement Event Service (SMES), encompassing data beginning in December 2022. This Phase 1 dataset includes over a year's worth of data, amounting to approximately 30TB, and covers 48 airports within the US National Airspace System. The dataset contains raw position reports, processed interpolated trajectories, and detailed airport map information derived from OpenStreetMap (OSM). The processed data aims to be compatible with modern machine learning libraries, facilitating ease of use for researchers across various domains.

The dataset is meticulously validated through statistical analyses to ensure high accuracy and representativeness. The study highlights the diversity within the dataset by selecting ten specific airports for detailed analysis, showcasing varying traffic levels and topological complexities. This validation ensures the Amelia dataset's robustness and suitability for training large-scale predictive models.

Model Architecture

The Amelia model leverages transformer-based architecture, inspired by transformer applications in natural language processing and computer vision. It employs a next-token-prediction approach to forecast aircraft trajectories across multiple airports. Key aspects of the model include:

1. Scene Representation: The model follows an ego-centric representation, transforming the scene with respect to a pre-selected ego-agent. This transformation is fundamental for capturing relevant kinematic and interaction profiles of agents within the scene. The model prioritizes agents based on their criticality, leveraging automated scoring functions to select high-relevance scenes.
2. Scene Encoder: The scene encoder uses multiple Attention mechanisms, specifically Temporal, Interaction, and Cross-attention layers, to model temporal relationships, agent-to-agent interactions, and agent-to-context relationships. This hierarchical encoding strategy ensures rich feature extraction and efficient representation of complex multi-agent scenarios.
3. Trajectory Decoder: The model utilizes a Gaussian Mixture Model (GMM) for predicting a distribution of potential future trajectories. Each mode within the GMM specifies a potential future state with associated confidence scores, accommodating the inherent uncertainty and multimodality of future trajectories in airport operations.

Experimental Evaluation

The paper conducts two primary experiments to evaluate the model's performance:

1. Scene Representation Strategy: This experiment compares the proposed scene selection strategy against a baseline random selection approach. By analyzing metrics such as mADE and mFDE for short-term and long-term predictions, the study demonstrates the superiority of the proposed strategy in selecting more dynamically relevant agents, albeit making the prediction task inherently more challenging.
2. Generalization Across Airports: The generalization capabilities of the model are tested by training on varying subsets of airports and evaluating on unseen airports. The results indicate that increasing the diversity and amount of training data enhances the model's generalization performance significantly. Particularly, the multi-airport experiment with the highest diversity of training data (7 airports) displays the best generalization performance, outperforming the single-airport baseline in several unseen settings.

Implications and Future Work

The implications of this research are profound both practically and theoretically. Practically, the open-sourcing of the Amelia dataset paves the way for more accessible and scalable research in airport surface movement forecasting. The transformer-based Amelia model, with its robust multi-future prediction capabilities, can be leveraged for various downstream tasks such as collision risk assessment, anomaly detection, and taxi-out time prediction. Theoretical advancements also emerge from investigating new scene representation and encoding strategies in predictive modeling.

Future research directions include further scaling the dataset and improving model efficiency and performance. The continued collection of SWIM SMES data and exploration of more sophisticated scene encoding strategies are crucial for enhancing long-term generalization and operational robustness. Moreover, applying the Amelia model to more specific downstream applications will validate its utility and foster advancements in autonomous and remotely-operated Aerial Mobility (AAM) systems.

Conclusion

In conclusion, the paper "Amelia: A Large Model and Dataset for Airport Surface Movement Forecasting" presents a significant leap forward in airport operational research. With the introduction of a large-scale dataset and a sophisticated predictive model, the research sets a new benchmark for studying airport surface movements. The open-sourcing of these resources invites further exploration and innovation in the field, holding potential to address the growing complexities of air traffic management and ensure safer, more efficient airport operations.