Distilling Tiny and Ultra-fast Deep Neural Networks for Autonomous Navigation on Nano-UAVs

TL;DR

This work tackles onboard visual navigation for nano-UAVs under severe memory and compute constraints by distilling Tiny-PULP-Dronet v3 CNNs that dramatically shrink model size while boosting throughput to up to 139 fps. A new 66k-image unified dataset with joint collision and steering labels is introduced and released, enabling end-to-end learning for static obstacle avoidance and navigation. Through hardware-aware deployment on the GAP8 platform and 8-bit quantization, the authors demonstrate substantial memory savings (up to 168×) with competitive accuracy, and validate performance in-field against both the prior PULP-Dronet v2 and their new models. The results show that Tiny-PULP-Dronet v3 can safely navigate challenging narrow corridors with static and dynamic obstacles up to speeds of 1 m/s, signaling a practical path toward multi-task onboard perception on ultra-light nano-UAVs.

Abstract

Nano-sized unmanned aerial vehicles (UAVs) are ideal candidates for flying Internet-of-Things smart sensors to collect information in narrow spaces. This requires ultra-fast navigation under very tight memory/computation constraints. The PULP-Dronet convolutional neural network (CNN) enables autonomous navigation running aboard a nano-UAV at 19 frame/s, at the cost of a large memory footprint of 320 kB -- and with drone control in complex scenarios hindered by the disjoint training of collision avoidance and steering capabilities. In this work, we distill a novel family of CNNs with better capabilities than PULP-Dronet, but memory footprint reduced by up to 168x (down to 2.9 kB), achieving an inference rate of up to 139 frame/s; we collect a new open-source unified collision/steering 66 k images dataset for more robust navigation; and we perform a thorough in-field analysis of both PULP-Dronet and our tiny CNNs running on a commercially available nano-UAV. Our tiniest CNN, called Tiny-PULP-Dronet v3, navigates with a 100% success rate a challenging and never-seen-before path, composed of a narrow obstacle-populated corridor and a 180° turn, at a maximum target speed of 0.5 m/s. In the same scenario, the SoA PULP-Dronet consistently fails despite having 168x more parameters.

Figures (7)

• Figure 1: Our autonomous nano-UAV navigating an unknown environment.
• Figure 2: A) our dataset collection methodology, B) our dataset collector GUI, and C) a sample sequence from our collected dataset.
• Figure 3: Our CNN architecture exploration includes: i) three block types -- RB, D+P, IRLB; ii) an optional bypass connection (dashed line); iii) variations on the number of channels based on $\gamma$. Output feature map sizes are represented as ($Width\times Height \times Channels$).
• Figure 4: Distribution of the yaw-rate labels (normalized in [-1,+1]) of our testing dataset and classification/regression performances of three trivial predictors.
• Figure 5: A) our U-shaped path for the in-field experiments. B) T op-view 2D representation of the path, highlighting its division into three segments (S1, S2, and S3) and the position of the four obstacles (represented with black rectangles).