pulp-dronet

Why we included this project

Autonomous nano-drones have to make navigation decisions in milliseconds on hardware with almost no power budget, and this repository shows one proven way to do it. PULP-Dronet pairs a convolutional network trained end-to-end for visual obstacle avoidance with a deployment pipeline built for the PULP family of ultra-low-power RISC-V processors, so the same model that dodges obstacles in simulation can fly a pocket-size quadrotor in real time. The repo holds the whole stack: training code, the distilled network architectures, and the firmware and closed-loop control glue needed to run on actual hardware. Several peer-reviewed papers document the design choices, so you can trace the decisions and reproduce the reported results instead of trusting a black box. Teams working on embedded low-power vision can use it as a practical starting point, and researchers get a reference baseline for nano-UAV autonomy.