GHOST: A GPS-Free Inertial Navigation System for Autonomous Vehicles

Recent learned inertial odometry (IO) methods such as AirIO [26] and AI-IO [2] have substantially improved per-window displacement accuracy on UAV datasets, yet deployment of these models on production autopilot hardware has remained largely unaddressed: existing work runs on offline GPUs or companion computers and uses research-grade EKFs rather than a real flight controller’s state estimator. We present GHOST, a systems contribution that closes this gap: a complete GPS-free inertialnavigation stack that runs a learned 3-D displacement network inside the production ArduPilot EKF3 sensor-fusion pipeline on commodity STM32H7 hardware. The runtime stack consists of (i) a custom TFLM-deployable v2.5 convolutional network — multi-scale + dilated convolutions + squeeze-and-excitation attention, trained with a heteroscedastic negative-log-likelihood loss — that achieves 0.148 m mean 3-D error on the public tlio_golden benchmark (67,323 windows); (ii) a C++ AP_GHOST library that schedules TFLM inference at 200 Hz, applies ZUPT and an IIR vibration filter, and publishes VISION_POSITION_ESTIMATE messages with adaptive covariance derived from the network’s predicted variance to EKF3 via the standard ExternalNav source. We validate the full pipeline in ArduPilot SITL and verify that the firmware boots correctly on a real CubeOrangePlus flight controller. Although demonstrated on a multirotor autopilot, the same network and EKF integration apply to ground vehicles whose motion is predominantly planar. All code, pre-trained weights, and notebooks are released to support reproducibility.