Intention-Guided Airport Surface Trajectory Forecasting Using a Vision Language Model

The Transformer-GMM trajectory predictor represents the state of the art in many prediction tasks. It offers two main advantages. First, the attention mechanism in the transformer architecture learns which environmental features are most relevant for forecasting future motion. Second, the Gaussian mixture model captures highly multimodal trajectory outcomes by allowing each maneuver mode to correspond to a separate Gaussian component, producing multiple meaningful probability peaks in the predicted distribution. However, those advantages may not be essential. People with little driving or operational training naturally recognize which information matters most and which can be ignored, and they also can constrain future motion to a small set of likely patterns, since real trajectories tend to follow only a limited number of repeated behaviors. But we cannot rely on a human operator to assist the prediction model, so we look for something that can provide similar guidance: a pretrained vision–language model (VLM). In this paper, we propose an intention-guided airport surface trajectory forecasting framework that exploits a pretrained VLM’s reasoning ability for semantic interpretation of the airport map. The forecast result of the proposed framework is much more robust and useful. We also introduce a new metric, the Aberration Rate, which evaluates whether a predicted trajectory set is usable by an autonomous system.