Making Autonomous Nano-drones Smarter to Scale New Heights

TII ARRC Web Announcement Nano Drones EN Final


Our ARRC team has recently marked a new signpost in the autonomous navigation of a nano drone (9cm x 9cm x 3cm) in an unknown environment with static and mobile obstacles – using only a small camera onboard and an Artificial Intelligence deck, on which a Convolutional Neural Network (CNN) has been deployed. The drone is trained to recognize obstacles and avoid them.

The project is unique because it features a small drone, with limited computational, energy and memory capabilities, performing a complicated task with the help of only cost-efficient sensors. The current version 3 of the CNN, used to make the nano-drones smarter and faster, was designed and deployed collaboratively by our ARRC researchers and ARRC’s partner entity, the University of Bologna, and is supported by ETH Zürich.

The Crazyflie 2.1 (the nano-drone used) is a versatile open-source flying development platform and punches way above its weight – a mere 27g! Comfortably fitting in the palm of your hand, the nano-drone is a good flyer, and is equipped with low-latency, long-range radio as well as Bluetooth LE.

The demo highlighted two state-of-the-art technologies at work. The first included a nano-drone autonomously performing vision-based navigation and collision avoidance - enabled by the newly designed onboard CNN. The novelty of this version is that our researchers have found the way to embed intelligence into the neural network and make the nano-drones capable of carrying out the same operations as they did with previous versions of the CNN –10 times faster and using 10 times less memory! The impressive results will allow nano-drones to learn other skills and perform them parallelly with autonomous navigation. The second demo featured a swarm of nano-drones flying in formation using ultra-wideband communication. The demonstration concluded with a hackathon, during which researchers tackled the problem of random exploration of unknown environments.

Robotics primarily relies on vision as the key component in solving classic mobile robotics problems such as "obstacle detection" and "obstacle avoidance" – these are the main building blocks in defining autonomous navigation algorithms.

Such pioneering initiatives will boost TII’s efforts in positioning ARRC among an elite group of research centers worldwide that are working on making nano-drones smarter. After leveraging AI techniques to detect and avoid obstacles, the nano-drones learn how to find specific objects or people. Prof. Enrico Natalizio, Senior Director, ARRC, Prof. Luca Benini, head of the PULP team in Bologna and Zurich and their teams believe the results of this research will help first-aid responders in disaster management scenarios, where a quick intervention is often challenging.