AeroCast: Probabilistic 3D Trajectory Prediction for Non-Cooperative Aerial Obstacles via Transformer-MDN Architecture

As autonomous aerial vehicles become more common, they must safely navigate shared airspace alongside non-cooperative obstacles—objects like birds, uncontrolled drones, or debris that do not communicate their intent. Existing methods often rely on recurrent neural networks that struggle to capture long-range kinematic patterns or produce deterministic point forecasts, which fail to communicate the inherent uncertainty of future motion. This paper asks: can a Transformer-based architecture provide accurate, multi-modal, and probabilistic 3D trajectory predictions that are fast enough for real-time onboard deployment?

The authors introduce AeroCast, a framework that combines a Pre-Layer Normalization (Pre-LN) Transformer encoder with a Mixture Density Network (MDN) output head. The model processes a 2-second history of 3D positions and velocities, encoded as translation-invariant consecutive displacements. By using a Transformer encoder, the model captures global temporal dependencies across the observation window. The MDN output head predicts a Gaussian mixture distribution for each future timestep, allowing the system to represent multiple potential future trajectories (e.g., an obstacle turning left versus continuing straight) and quantify uncertainty through the variance of these components.

AeroCast significantly outperforms existing recurrent baselines on a hybrid real-and-synthetic quadrotor dataset. It achieves approximately a 50% reduction in both Average Displacement Error (ADE) and Final Displacement Error (FDE) over a 5-second prediction horizon. Furthermore, the model demonstrates superior probabilistic performance, yielding the lowest negative log-likelihood and Continuous Ranked Probability Score. The architecture is computationally efficient, with inference completing in 0.1ms per sample, making it suitable for real-time deployment at 100Hz.