PX4 Developer Summit 2021

The primary estimator in PX4 is the Extended Kalman Filter (EKF). EKF first predicts the states using the Inertial Measurement Unit (IMU), then corrects them using the barometer, magnetometer, and GPS. When GPS lock is lost, horizontal position and velocity estimates deteriorate quickly. This talk presents a Recurrent Neural Network (RNN) that fuses the raw IMU, baro, and mag measurements to predict the local position without aiding from any external sensor. The RNN was trained on 465 public logs from the PX4 flight review database and validated on another 83 logs. The maximum position error (MPE) between the RNN predictions (without GPS) and EKF estimates (with GPS) in a single flight is used for accuracy assessment. The median MPE in the 83 validation flights is 35 meters. MPE values as low as 2.7 meters in a 5-minutes flight were achieved using the presented system. The RNN works on the ground station laptop in real-time, the sensors measurement are fed to the network via MAVROS and the position and velocity predictions are published to dedicated topics. Lighter versions of the network were tested on Jetson Nano and raspberry Pi 3 B+ with slightly decreased accuracy.