Automated Tuning of End-to-end Neural Flight Controllers for Autonomous Nano-drones

Convolutional neural networks (CNNs) are fueling the advancement of autonomous palm-sized drones, i.e., nano-drones, despite their limited power envelope and onboard processing capabilities. Computationally lighter than traditional geometrical approaches, CNNs are the ideal candidates to predict end-to-end signals directly from the sensor inputs to feed to the onboard flight controller. However, these sophisticated CNNs require significant complexity reduction and fine-grained tuning to be successfully deployed aboard a flying nano-drone. To date, these optimizations are mostly hand-crafted and require error-prone, labor-intensive iterative development flows. This work discusses methodologies and software tools to streamline and automate all the deployment stages on a low-power commercial multicore System-on-Chip. We investigate both an industrial closed-source and an academic open-source tool-set with a field-proofed state-of-the-art CNN for autonomous driving. Our results show a $2 \times$ reduction of the memory footprint and a speedup of $1.6 \times$ in the inference time, compared to the original hand-crafted CNN, with the same prediction accuracy.