Overview
Predicts seconds until a UAV first enters a 1 km observation sphere, trained only on synthetic motion. Two preregistered sim-to-real experiments (72 tuning trials, 36 frozen models) on MASC-3, ALFA, and AMOVFLY ask what makes synthetic training data useful on real flights.
Technologies
Features
- Synthetic-Only Training - RK4 point-mass generator with 30 fixed-wing / multicopter profiles; deterministic across worker counts.
- Finding: Diversity ≠ Transfer - Mixed profiles cut synthetic error 24% but raised MASC-3 GRU error 53%; no arm improved AMOVFLY mean.
- Finding: Matching Helps - Sampling matched to 3 calibration dates cut GRU error 50.9 → 14.4 s and LSTM 39.5 → 16.0 s on 69 held-out flights (all 3 seeds); MLP stayed near 16 s.
- Frozen Protocols - Model hashes and zone geometry locked before prediction; 144 combinations independently recomputed from saved arrays.