💻 computer science

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

The paper presents ASCENT, a lightweight transformer-based model that integrates domain-aware coordinate normalization and parameterized predictions to achieve accurate, low-latency multi-modal 3D aircraft trajectory forecasting in non-towered terminal airspace, outperforming existing baselines on TrajAir and TartanAviation datasets.

Original authors: Alexander Prutsch, David Schinagl, Horst Possegger

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alexander Prutsch, David Schinagl, Horst Possegger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy, uncontrolled intersection in the sky. Unlike a major airport with a tower and air traffic controllers telling every plane exactly where to go, this is a "non-towered" airspace. Here, pilots fly by the book and their own judgment, much like cars at a four-way stop. While this works most of the time, it's risky. General Aviation (small private planes, business jets, crop dusters) has a much higher accident rate than commercial airlines, mostly because there's no "traffic cop" to prevent collisions.

The authors of this paper, Alexander Prutsch and his team, have built a digital "crystal ball" called ASCENT to help solve this. It's an AI system designed to predict exactly where a plane will go next, giving pilots and safety systems a heads-up before a potential crash.

Here is how ASCENT works, explained through simple analogies:

1. The Problem: Reading the Road vs. Reading the Sky

In the world of self-driving cars, predicting where a vehicle goes is easier because cars are stuck on lanes. They can't just drive over a curb or turn 90 degrees in the middle of a block. They follow a strict map.

Planes, however, are like birds. They fly in 3D space. They can climb, dive, turn left, turn right, or circle around. They don't have painted lanes in the sky. Furthermore, while a car might only need to predict where another car will be in 3 seconds, a plane needs to know where it will be in the next 2 minutes.

Previous AI models tried to guess these paths using "generative" methods (like rolling dice to see what happens). The authors realized this was like trying to guess a plane's path by randomly throwing darts. Instead, they looked at how self-driving cars solve similar problems and adapted those techniques for the sky.

2. The Solution: ASCENT (The Smart Navigator)

ASCENT is built on a Transformer, a type of AI famous for understanding language (like the tech behind chatbots). But instead of reading words, ASCENT reads "flight paths."

Here are the three secret ingredients that make it special:

A. The "Head-Up" View (Coordinate Normalization)

Imagine you are driving a car. If you look at a map, you see your car as a dot moving across a giant grid. But if you are in the car, you don't care about the map; you care about "turning left at the next light" or "going straight."

Older AI models looked at the map (global coordinates). ASCENT puts the AI inside the pilot's seat. It takes the plane's history and rotates the world so the plane is always facing "forward" and "up." This helps the AI recognize patterns, like "a left turn" or "a climb," regardless of which direction the plane is actually flying over the ground.

B. The "Detective's Notebook" (The Encoder)

Once the data is rotated into the pilot's view, ASCENT uses a Transformer Encoder to read the history. Think of this like a detective reading a suspect's past behavior.

"Did they speed up?"
"Did they start banking left?"
"Are they getting closer to the runway?"

It doesn't just look at the last second; it looks at the whole story to understand the intent of the flight.

C. The "Crystal Ball" with Options (The Query-Based Decoder)

This is the most creative part. In the past, AI models would try to guess one path or generate a messy cloud of possibilities.

ASCENT uses something called "Learnable Mode Queries." Imagine a flight instructor asking the AI:

"Okay, what if the plane lands?"
"What if the plane turns right?"
"What if the plane circles around?"

The AI has five "mental slots" (queries) ready to answer these specific questions. Instead of guessing randomly, it actively generates five distinct, high-quality scenarios (e.g., "Landing," "Go-around," "Turn Left," "Turn Right," "Straight"). It then tells you the probability of each scenario happening.

3. Why It's a Big Deal

The authors tested ASCENT on real flight data from airports in Pittsburgh and found it was significantly better than all previous methods.

Accuracy: It predicted where planes would be much closer to reality than the old models.
Speed: It's incredibly fast. While other models take a long time to calculate (like a slow computer), ASCENT is so lightweight it could run on a small device in a plane or a pilot's tablet in real-time.
Safety: By predicting multiple possibilities (not just one), it gives a much better safety net. If the AI sees a plane might turn left or right, it can warn the other pilot to be ready for either.

The Bottom Line

Think of ASCENT as a super-smart co-pilot that never gets tired. It watches the sky, understands the "language" of flight maneuvers, and instantly whispers to the pilot: "Hey, that plane ahead looks like it might turn left in 30 seconds. Be ready."

By borrowing ideas from self-driving cars and applying them to the unique, 3D world of aviation, the authors have created a tool that could make flying in uncontrolled airspace much safer for everyone.

1. Problem Statement

The paper addresses the critical safety challenge in General Aviation (GA) within non-towered terminal airspace. Unlike commercial aviation, which operates under strict Air Traffic Control (ATC) in controlled airspace, GA flights (recreational, business, aerial work) rely on pilot judgment and standard traffic patterns. This environment has an accident rate approximately 30 times higher than commercial aviation, with most incidents occurring near airports.

The core technical problem is multi-modal 3D aircraft trajectory prediction. The system must forecast future 3D positions ( $T_f$ steps) based on historical motion ( $T_h$ steps). Key challenges include:

3D Dynamics: Aircraft move in three dimensions (pitch, yaw, roll) unlike 2D ground vehicles.
Lack of Lanes: Unlike autonomous driving where vehicles follow lane topologies, aircraft follow fluid but structured traffic patterns (e.g., standard traffic patterns, take-off/landing vectors).
Long Horizons: Aviation prediction requires forecasting over minutes (e.g., 120 seconds), whereas automotive prediction typically covers only a few seconds.
Multi-modality: A single historical trajectory can lead to multiple valid future maneuvers (e.g., landing, turning, going around).

2. Methodology: The ASCENT Architecture

The authors propose ASCENT (Aircraft Sequence Encoding Transformer), a lightweight, transformer-based model designed to bridge the gap between autonomous driving motion forecasting and aviation trajectory prediction.

A. Coordinate Modeling & Normalization

To effectively learn local motion patterns while retaining global context, ASCENT employs a novel normalization strategy:

Agent-Centric Normalization: Historical trajectories are shifted and rotated relative to the aircraft's current pose. This involves calculating the yaw angle ( $\gamma$ ) and pitch angle ( $\theta$ ) using the last two position steps.
3D Positional Embeddings: To prevent the loss of global context (e.g., distance to the runway) after normalization, the model injects global positional embeddings. These encode the absolute 3D position $(x, y, z)$ and orientation angles $(\gamma, \theta)$ using sine/cosine wrapping.

B. Motion Encoding (Encoder)

Input: The normalized local history $H_{local}$ is projected into a feature space and augmented with temporal encodings.
Architecture: A Transformer-based motion encoder using self-attention blocks processes the sequence.
Feature Fusion: The encoded motion features are combined with the global positional embeddings (processed via an MLP) to create a comprehensive agent feature vector $A$ .

C. Decoding with Learnable Mode Queries

Unlike previous aviation models that use Variational Autoencoders (VAE) or Diffusion models, ASCENT adopts a query-based decoder inspired by state-of-the-art autonomous driving methods (e.g., AgentFormer, SceneTransformer).

Mode Queries: The model uses $k$ learnable "mode queries" ( $Q \in \mathbb{R}^{k \times D}$ ) to represent distinct potential future maneuvers (e.g., landing, left turn, right turn).
Parameterized Prediction: Instead of directly predicting 3D coordinates, the decoder predicts flight parameters:
- Speed: Predicted via linear output.
- Angles: Yaw and pitch are predicted as sine/cosine pairs to handle angular periodicity and ensure stable loss, then converted back to angles via arctangent.
Trajectory Generation: The predicted parameters are integrated kinematically to generate future 3D positions, which are then transformed back to the global coordinate system.

3. Key Contributions

Novel Architecture: Introduction of ASCENT, the first transformer-based model specifically adapted for 3D aircraft trajectory prediction in non-towered airspace, achieving new state-of-the-art (SOTA) results.
Cross-Domain Transfer: Successfully adapts mode query-based decoding from autonomous driving to aviation, proving it superior to generative stochastic models (VAE/Diffusion) for structured aircraft traffic patterns.
Comprehensive Evaluation:
- Achieves SOTA on the TrajAir dataset across multiple input horizons (11s, 16s, 40s) and output modes.
- First-ever reported results on the TartanAviation dataset, including cross-dataset evaluations between TrajAir and TartanAviation.
Efficiency: Demonstrates significantly lower inference latency compared to existing baselines, making it suitable for real-time edge deployment.

4. Experimental Results

The model was evaluated on the TrajAir and TartanAviation datasets using standard metrics: minADE (Average Displacement Error) and minFDE (Final Displacement Error) over $k$ candidates.

Performance on TrajAir:
- 11s Input: ASCENT achieved a minADE5 of 0.35 km and minFDE5 of 0.58 km, significantly outperforming the previous SOTA (TrajAirNet: 0.78/1.55).
- 16s Input: Achieved minADE5 of 0.33 km and minFDE5 of 0.59 km, beating competitors like ATP and GooDFlight.
- 40s Input (20 modes): Achieved minADE20 of 0.19 km and minFDE20 of 0.26 km, surpassing GooDFlight (0.29/0.39).
Performance on TartanAviation:
- Showed strong generalization on both the non-towered KBTP airport and the towered KAGC airport, with minADE5 values around 0.32–0.50 km.
Ablation Studies:
- Decoder: Mode queries outperformed CVAE baselines, confirming better alignment with structured traffic patterns.
- Coordinates: The combination of local normalization + global positional embeddings yielded the best results, outperforming pure global or pure local coordinate systems.
- Prediction Format: Parameterized prediction (speed/angles) resulted in more stable long-horizon trajectories compared to direct 3D position decoding.
Latency: ASCENT demonstrated extremely low latency (e.g., 11–18 ms for up to 32 aircraft on a V100 GPU), roughly 3–6x faster than TrajAirNet, due to the absence of iterative sampling.

5. Significance and Conclusion

ASCENT represents a significant advancement in aviation safety systems. By leveraging insights from autonomous driving (transformers, mode queries) and adapting them to the unique 3D, long-horizon nature of flight, the model provides:

Higher Accuracy: Significantly reduced prediction errors compared to existing methods.
Real-Time Capability: Low latency enables potential integration into real-time conflict detection systems for General Aviation.
Diversity: The ability to generate diverse, multi-modal hypotheses helps pilots and safety systems anticipate a wider range of possible pilot actions.

The authors note that while the model implicitly learns runway relationships, generalization to entirely new airports without fine-tuning remains a challenge due to data scarcity. However, the architecture is designed to easily incorporate additional context (e.g., weather, ATC voice) in future iterations.