GateIO: Yaw-Rate-Gated Learned Dead Reckoning for UAV GPS-Outage Bridging, Tested Across Held-Out Flights

GateIO bridges GPS outages on small UAVs by predicting GPS velocity from inertial data and integrating it to a position. It is a formulation rather than a single network: a yaw-rate-gated residual head that holds the last known velocity while the aircraft flies straight and reacts during turns. We test it with two backbones, a two-layer LSTM (GateIO-LSTM) and a convolution-plus-attention network (GateIO-TCN). The main lesson of the paper is about evaluation. A within-flight split, where training and test windows come from the same flights, leaks information and makes the model look far more accurate than it is. Under leave-one-flight-out cross-validation, where whole flights are held out, GateIO-LSTM reaches 11.2 m mean endpoint drift over a 10-second outage across 1,390 held-out sequences. That is about 7.5 times lower than a tuned extended Kalman filter and a constant-velocity baseline, and roughly ten times lower on turns. The higher-capacity convolutional backbone does worse (23.4 m), so on this small five-flight dataset the recurrent model generalises better and more stably. We report the leakage-free numbers, include a two-backbone ablation and held-out trajectories, and state the limits plainly: five flights from one platform, and a point estimate with no uncertainty. All code and preprocessing scripts are public. Dataset derived from MARS-LVIG (Li et al., 2024)Code: https://github.com/nandini1612/gateio