Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Tiny-PULP-Dronets: Squeezing Neural Networks for Faster and Lighter Inference on Multi-Tasking Autonomous Nano-Drones

Lorenzo Lamberti, Vlad Niculescu, Michał Barcis, Lorenzo Bellone, Enrico Natalizio, Luca Benini, Daniele Palossi

TL;DR

This work addresses enabling high-level onboard perception for nano-UAVs under tight power and memory constraints. It introduces Tiny-PULP-Dronet, a squeezing strategy that scales channels via $\gamma$ and removes by-passes from the PULP-Dronet CNN, guided by sparsity and overfitting analysis. The approach delivers up to $50\times$ fewer parameters and $27\times$ fewer MACs with minimal impact on regression/classification performance (minimum validation MSE $=0.01$, accuracy up to $0.88$). Power and latency experiments on the GAP8 MCU demonstrate substantial gains in frame-rate and energy efficiency, enabling practical multi-task onboard intelligence for nano-UAVs.

Abstract

Pocket-sized autonomous nano-drones can revolutionize many robotic use cases, such as visual inspection in narrow, constrained spaces, and ensure safer human-robot interaction due to their tiny form factor and weight -- i.e., tens of grams. This compelling vision is challenged by the high level of intelligence needed aboard, which clashes against the limited computational and storage resources available on PULP (parallel-ultra-low-power) MCU class navigation and mission controllers that can be hosted aboard. This work moves from PULP-Dronet, a State-of-the-Art convolutional neural network for autonomous navigation on nano-drones. We introduce Tiny-PULP-Dronet: a novel methodology to squeeze by more than one order of magnitude model size (50x fewer parameters), and number of operations (27x less multiply-and-accumulate) required to run inference with similar flight performance as PULP-Dronet. This massive reduction paves the way towards affordable multi-tasking on nano-drones, a fundamental requirement for achieving high-level intelligence.

Tiny-PULP-Dronets: Squeezing Neural Networks for Faster and Lighter Inference on Multi-Tasking Autonomous Nano-Drones

TL;DR

This work addresses enabling high-level onboard perception for nano-UAVs under tight power and memory constraints. It introduces Tiny-PULP-Dronet, a squeezing strategy that scales channels via and removes by-passes from the PULP-Dronet CNN, guided by sparsity and overfitting analysis. The approach delivers up to fewer parameters and fewer MACs with minimal impact on regression/classification performance (minimum validation MSE , accuracy up to ). Power and latency experiments on the GAP8 MCU demonstrate substantial gains in frame-rate and energy efficiency, enabling practical multi-task onboard intelligence for nano-UAVs.

Abstract

Pocket-sized autonomous nano-drones can revolutionize many robotic use cases, such as visual inspection in narrow, constrained spaces, and ensure safer human-robot interaction due to their tiny form factor and weight -- i.e., tens of grams. This compelling vision is challenged by the high level of intelligence needed aboard, which clashes against the limited computational and storage resources available on PULP (parallel-ultra-low-power) MCU class navigation and mission controllers that can be hosted aboard. This work moves from PULP-Dronet, a State-of-the-Art convolutional neural network for autonomous navigation on nano-drones. We introduce Tiny-PULP-Dronet: a novel methodology to squeeze by more than one order of magnitude model size (50x fewer parameters), and number of operations (27x less multiply-and-accumulate) required to run inference with similar flight performance as PULP-Dronet. This massive reduction paves the way towards affordable multi-tasking on nano-drones, a fundamental requirement for achieving high-level intelligence.
Paper Structure (6 sections, 3 figures, 1 table)

This paper contains 6 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Results
  4. Models evaluation
  5. Power analysis
  6. Conclusion

Figures (3)

  • Figure 1: Training/validation loss' components -- BCE (A) and MSE (B) -- of PULP-Dronet, comparing the the baseline model against the tiny one.
  • Figure 2: Comparing PULP-Dronet vs. its Tiny variants, in terms of size, MAC, Accuracy, and RMSE. We span $\gamma$ in the $[0.125, 0.250, 0.5, 1.0]$ range.
  • Figure 3: GAP8 power waveforms executing the smallest Tiny-PULP-Dronet ($\gamma=0.125$). (A) most energy-efficient SoC's configuration -- FC@50M Hz, CL@100M Hz, VDD@1.0V -- and (B) the maximum performance one -- FC@250M Hz, CL@175M Hz, VDD@1.2V.