Language Conditioning Improves Accuracy of Aircraft Goal Prediction in Non-Towered Airspace

TL;DR

A multimodal framework for aircraft goal prediction that integrates natural language understanding with spatial reasoning to improve autonomous decision-making in non-towered airspace and demonstrates that language-conditioned prediction increases prediction accuracy.

Abstract

Autonomous aircraft must safely operate in non-towered airspace, where coordination relies on voice-based communication among human pilots. Safe operation requires an aircraft to predict the intent, and corresponding goal location, of other aircraft. This paper introduces a multimodal framework for aircraft goal prediction that integrates natural language understanding with spatial reasoning to improve autonomous decision-making in such environments. We leverage automatic speech recognition and large language models to transcribe and interpret pilot radio calls, identify aircraft, and extract discrete intent labels. These intent labels are fused with observed trajectories to condition a temporal convolutional network and Gaussian mixture model for probabilistic goal prediction. Our method significantly reduces goal prediction error compared to baselines that rely solely on motion history, demonstrating that language-conditioned prediction increases prediction accuracy. Experiments on a real-world dataset from a non-towered airport validate the approach and highlight its potential to enable socially aware, language-conditioned robotic motion planning.

Figures (6)

• Figure 1: Example of a pilot radio call on the CTAF in non-towered airspace. An aircraft (blue in inset) makes a radio call announcing its intent with respect to the "Runway 8" left traffic pattern. Multiple goals are possible depending on the radio call. In this work, we show that conditioning on such radio calls along with the aircraft's recent trajectory (blue line on downwind leg) improves accuracy of goal prediction, which is critical for autonomous operation in non-towered airspace.
• Figure 2: An overview of our model architecture as described in \ref{['subsec: goal prediction']}. Our key finding is that conditioning on natural language radio calls improves autonomous aircraft's ability to predict the intent of other aircraft in non-towered airspace.
• Figure 3: Example trajectory prediction comparing our method to TrajAirNet patrikar2022predicting, showing the improvement from using radio call info (in speech bubble).
• Figure 4: Final displacement error (FDE) of predicted aircraft goals with respect to observation time horizon. Longer observation windows do not improve goal prediction accuracy. The observation horizons used in TrajAirNet patrikar2022predicting, GooDFlight yang2025goodflight and Social-PatteRNN-ATT navarro2022social are shown as vertical lines. IQR is plotted as a measure of dispersion.
• Figure 5: Final displacement error (FDE) of predicted aircraft goals vs. prediction time horizon length. Error increases with longer horizons, highlighting the challenge of long-term forecasting. However, language-conditioned goal prediction is better suited to address this challenge, showing a smaller increase in error. The prediction horizons used in TrajAirNet patrikar2022predicting, GooDFlight yang2025goodflight and Social-PatteRNN-ATT navarro2022social are shown as vertical lines. IQR is plotted as a measure of dispersion.