ETH Zurich's AI Boosts Nano-Drone Speed 4x with Automated CNN Optimization

A team led by Vlad Niculescu and Luca Benini at ETH Zurich has developed an automated end-to-end optimization pipeline for deploying deep neural networks (DNNs) on nano-drones—sub-10 cm UAVs with severe computational and memory constraints. Their work, published in IEEE JETCAS, focuses on the PULP-Dronet CNN, which enables autonomous navigation without a human pilot. The manual tuning of such CNNs is error-prone and labor-intensive, so the researchers created a tool that automates the entire process from training to closed-loop deployment on a ULP multicore system-on-chip aboard the Crazyflie 2.1 nano-UAV.

The results are striking: compared to the hand-crafted baseline, the automated pipeline produced a 2x reduction in memory footprint and a 1.6x speedup in inference time while maintaining identical prediction accuracy. In real-world tests, the optimized nano-drone achieved obstacle avoidance with a peak braking speed of 1.65 m/s and improved the speed/braking-space ratio. In free flight within a familiar environment, it reached 1.96 m/s—nearly four times faster than the baseline's 0.5 m/s. The drone also successfully followed a lane with a 90-degree turn. Crucially, all computation consumed less than 1.6% of the drone's total power budget, leaving ample energy for flight and sensing.

The team has open-sourced the entire software design as a ready-to-run project compatible with the Crazyflie 2.1. This work demonstrates that automated, end-to-end DNN optimization can dramatically improve nano-drone performance without hardware upgrades, paving the way for practical applications in surveillance, inspection, and environmental monitoring.