Computationally Efficient Autonomous Racing of a 72-gram Drone



We present a tiny autonomous racing drone, weighing only 72 grams. This tiny drone uses very efficient algorithms for onboard vision, state estimation and control in order to fly a (quite narrow) drone racing track with on average 2 m/s (peaks of 2.7 m/s). This speed is on a par with much larger, state-of-the-art autonomous racing drones.

The key idea behind our approach is not to perform generic, but computationally expensive, visual inertial odometry. Instead, the drone relies on model predictions, which are corrected by means of visual localization with the help of gate detections.

The details are described in the article “Visual Model-predictive Localization for Computationally Efficient Autonomous Racing of a 72-gram Drone.” – submitted.

Article link: https://arxiv.org/abs/1905.10110

source

Recommended For You

About the Author: MAVLab TU Delft

3 Comments

  1. Nice, but seems like you need a better PID tune due to the constant oscillations. Would probably get even better results if the oscillations were tuned out.

  2. Technisch heel indrukwekkend maar qua snelheid komt 't nog niet in de buurt van door mensen bestuurde FPV race-drones.

    Zou het in de nabije toekomst mogelijk zijn een drone zo te programmeren dat je een tof filmpje van een auto-race kan maken o.i.d., onder variabele omstandigheden?

Comments are closed.