AeroCast predicts drone/bird paths 50% more accurately in 0.1ms
New Transformer-MDN model cuts collision risk for autonomous aerial vehicles by half.
Autonomous aerial vehicles navigating shared airspace face a critical challenge: predicting the future positions of non-cooperative obstacles like birds, uncontrolled drones, or debris. These objects exhibit multi-modal motion that deterministic predictors cannot capture. Existing methods either rely on recurrent encoders that miss long-range kinematic precursors or produce point forecasts without distributional information for downstream planners.
AeroCast solves this with a Transformer encoder combined with a Mixture Density Network (MDN) output head, predicting per-timestep Gaussian mixture distributions over future 3D displacements. A translation-invariant consecutive displacement encoding and a calibration-oriented training objective address input design and mode-degradation challenges. On a hybrid dataset of real and synthetic quadrotor maneuvers across nine categories, AeroCast reduces Average Displacement Error (ADE) and Final Displacement Error (FDE) by roughly 50% over a five-second horizon versus baselines. It also achieves the lowest negative log-likelihood and Continuous Ranked Probability Score. Ablation studies show velocity input and model capacity drive quality, while positional encoding ensures long-horizon coherence. With inference at 0.1ms per sample, AeroCast is deployable onboard at 100Hz, enabling real-time evasive planning.
- Reduces Average and Final Displacement Error by ~50% over a 5-second prediction horizon vs. baselines
- Inference takes only 0.1ms per sample, enabling real-time deployment at 100Hz
- Handles non-cooperative obstacles like birds, uncontrolled drones, and debris with probabilistic Gaussian mixture outputs
Why It Matters
AeroCast makes autonomous drone collision avoidance safer and faster, critical for expanding UAV traffic in shared airspace.