Quadrotor Neural Dead Reckoning in Periodic Trajectories

TL;DR

This work tackles drift in pure inertial navigation for GNSS-denied quadrotor operation by introducing QuadPosNet, an end-to-end neural dead reckoning approach that regresses the 3D position change, $\Delta \mathbf{p}$, directly from inertial measurements during periodic trajectories. It presents two architectures, a single-head and a multi-head 1D-CNN/FC network, trained with the Adam optimizer to minimize mean-squared error, thereby eliminating the need for heading-based estimates. Evaluated on outdoor DJI Phantom 4 RTK data and indoor Crazyflie 2.1 data, QuadPosNet achieves substantial RMSE reductions over the QuadNet baseline, averaging $-27.1\%$ outdoors and $-79\%$ indoors, with indoor RMSE dropping from about $1.84$ m to below $0.2$ m in some cases. The approach enables robust, GNSS-denied navigation with only software changes, though the authors acknowledge a power consumption trade-off due to the emphasis on periodic trajectories and suggest future work to balance accuracy with mission duration.

Abstract

In real world scenarios, due to environmental or hardware constraints, the quadrotor is forced to navigate in pure inertial navigation mode while operating indoors or outdoors. To mitigate inertial drift, end-to-end neural network approaches combined with quadrotor periodic trajectories were suggested. There, the quadrotor distance is regressed and combined with inertial model-based heading estimation, the quadrotor position vector is estimated. To further enhance positioning performance, in this paper we propose a quadrotor neural dead reckoning approach for quadrotors flying on periodic trajectories. In this case, the inertial readings are fed into a simple and efficient network to directly estimate the quadrotor position vector. Our approach was evaluated on two different quadrotors, one operating indoors while the other outdoors. Our approach improves the positioning accuracy of other deep-learning approaches, achieving an average 27% reduction in error outdoors and an average 79% reduction indoors, while requiring only software modifications. With the improved positioning accuracy achieved by our method, the quadrotor can seamlessly perform its tasks.

Figures (8)

• Figure 1: Fig. Our QuadPosNet end-to-end neural network architecture. The network processes inertial measurements (accelerometer and gyroscope raw data) and predicts the change in the quadrotor's position vector.
• Figure 2: Fig. Single-head QuadPosNet architecture consisting of 1D-convolutional layers, used for feature extraction, and fully connected layers that output of quadrotor change in the position vector.
• Figure 3: Fig. Multi-head QuadPosNet architecture consisting has two input heads, each consists of 1D-convolutional layers, used for feature extraction, and fully connected layers that output of quadrotor change in the position vector.
• Figure 4: Fig. Outdoor experimental setup featuring a DJI Phantom 4 quadrotor equipped with four IMUs in our field experiment setup.
• Figure 5: Fig. Indoor experimental setup showing the Crazyflie quadrotor during data collection.