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

AeroCast presents a novel AI framework combining Transformer neural networks with Mixture Density Networks to predict probabilistic 3D trajectories of non-cooperative aerial obstacles. The system achieves 50% error reduction compared to existing methods while maintaining real-time performance at 100Hz, enabling safer autonomous aerial vehicle operations in shared airspace.

AeroCast addresses a critical safety challenge in autonomous aviation: predicting the unpredictable movement of non-cooperative obstacles like birds, uncontrolled drones, or debris. Traditional trajectory prediction systems rely on deterministic models that produce single-point forecasts, inadequate for capturing the inherent uncertainty in non-cooperative behavior. This research tackles the problem by outputting probabilistic distributions rather than point estimates, allowing downstream planners to account for multiple possible futures simultaneously.

The technical innovation combines two complementary approaches: Transformer encoders that capture long-range temporal dependencies in motion patterns, replacing slower recurrent architectures, and Mixture Density Networks that model multi-modal distributions reflecting real-world behavioral variability. The use of consecutive displacement encoding and calibration-focused training objectives specifically addresses trajectory prediction challenges rather than applying generic deep learning techniques. Experimental validation on a hybrid real-and-synthetic dataset spanning diverse motion categories demonstrates substantial improvements in both prediction accuracy and uncertainty quantification metrics.

For the autonomous aerial vehicle industry, this work directly impacts safety certification and real-world deployment feasibility. Regulators increasingly demand probabilistic safety guarantees rather than best-case scenarios. The framework's 0.1ms inference latency proves compatibility with onboard hardware constraints, removing a key deployment barrier. This enables autonomous systems to make collision-avoidance decisions based on distributional information rather than point predictions, fundamentally improving safety margins in shared airspace scenarios where humans, animals, and machines coexist.

• →AeroCast achieves 50% error reduction in trajectory prediction over five-second horizons compared to existing baselines.
• →Probabilistic output distributions provide downstream planners with uncertainty information critical for safety-critical autonomous systems.
• →Real-time 0.1ms inference enables onboard deployment on aerial platforms operating at 100Hz control rates.
• →Transformer architecture with Mixture Density Networks outperforms recurrent approaches at capturing long-horizon motion precursors.
• →Framework validated on real and synthetic quadrotor data spanning nine distinct motion categories with strong generalization.