Learning to fly

Overview

PREVOT Marc — 2017-02-23T10:56:54Z

What is Learning to Fly ?

The importance of drones in our society is getting more and more important. Many companies are now using drones in order to realise specific missions like infrastructure inspection. The usage of drones and machines learning technologies combined is a real opportunity for all the actors of this market. This consideration is at the birth of the ISAE-Supaero's project : learning to fly.
Learning to Fly aims at increasing the autonomy of a fixed wing drone and improve energy efficiency by benefiting from thermals. For this, we use artificial intelligence, more precisely reinforcement learning. Hence, our glider will be able to stay as long as possible in the air.

Who's in charge of the project ?

Learning to Fly is as project that has involves many people over the years. Researchers and students have contributed to the progress of this project. People involved are from ISAE-Supaero a prestigious engineering school in Toulouse (FRANCE) that deals mainly with aeronautics. Therefore, they were able to highlight their scientific know-how not only in the field of flight mechanics but also in the field of machine learning.

Project progress

At this stage, all our results are simulations on computer, this allows to test different models of artificial intelligence and thermals. The simulations are running on a homemade software. Initially, the simulator software was coded under Matlab, but for performance reasons, it was decided to recode it under C++. A detailed description of this new software, named L2F, is available in the section « Software ». An exemple of a L2F simulation is disposable in the section “Demo”.

In this website we will present the project in different headings
People : people involved in the project (Researchers and students)
Publications : Contains scientific publications related to the project.
Software : Contains a detailed description of the demonstrator as well as an explanation to how the code can be downloaded and its documentation accessed.
Demo : Contains an example of simulation and its results.

Software

ELJAAFARI Nail — 2017-01-24T14:16:09Z

L2F has been developed by an engineering student team of ISAE-Supaero, one of the best European aerospace engineering school. The team designed L2F as an open-source software, well documented and easy to adapt to any needs of soaring simulations. For now, it is used by the LearningToFly project as a proof of concept for the autonomous glider controlled by a machine learning algorithm.

ADAPTATIVE

The L2F code is divided in four main classes. These classes represent the major elements of a soaring flight (the aircraft, the pilot and the flight zone) and the stepper (representation of time in the simulation). This construction of code allows any users to put his models with his parameters and run simulations adapted to his need.
For now, L2F includes at least one model for each class and can be run properly.
Here a quick description of the classes :

Class AIRCRAFT
Aircraft implements the aircraft models. An aircraft holds three important concepts : his current configuration, an observation of the current state and a dynamic function which defines the next state given a command instruction. For now, the Beeler Glider model is implemented. This model has been defined by Beeler, Moerder and Cox, in the A Flight Dynamics Model for a Small Glider paper. It defines all the flight mechanical equations of a small glider.More information about the Beeler Glider model : https://ntrs.nasa.gov/search.jsp?R=20040031358.
Class PILOT
Pilot implements the pilot models. A pilot holds two important concepts : an observation of the current state and a command instruction. Two models are implemented for the moment : a passive pilot defined by a set command and an autonomous pilot controlled with a Q-learning algorithm.
Class FLIGHT ZONE
Flight Zone defines the flight zone in which the aircraft and its pilot move. A flight zone holds two important concepts : the wind and the ground geography. For now, the implemented model is a flat zone with thermals. The implemented thermal models are the Allen's model, the Childress's model, the Lenschow's model and the Lawrance's model. The thermals are randomly generated in the flight zone during the simulation.
More information about the thermal models in the Publication page.
Class STEPPER
Stepper manages the temporal evolution of the system. Given a set delta_t, it moves the current state of the classes Aircraft, Pilot and Flight Zone to the next state. The Euler integrator method and the Runge Kutta method are implemented for now.

OPENSOURCE

L2F is available below for downloading. Any user is free to study, change the code for his purpose and share it. The source code is based on the C++ language and works with Linux, Windows and MacOS operating systems.

DOCUMENTATION

The L2F source code is documented with the Doxygen standard. It generates a HTML documentation of the code which provides information on all functions, variables and interfaces. A “readme.txt” file is also provided and informs the user of how to run the simulation.
More informations about Doxygen : www.doxygen.org/

DOWNLOAD

Here you can download the L2F software : SOON

Here you can download the L2F tutorial :

People

ELJAAFARI Nail — 2017-01-24T14:15:16Z

FACULTY

	Emmanuel Rachelson	Project Leader	emmanuel.rachelson@isae.fr
	Eric Poquillon	Flight Mechanics specialist	eric.poquillon@isae.fr

PHD STUDENT

Photo

Erwan Lecarpentier

PhD student

erwan.lecarpentier@isae.fr

MASTER STUDENT

Photo	Erwan Lecarpentier	PhD student	erwan.lecarpentier@isae.fr
	Ludovic Becq	Student	ludovic.becq@isae.fr
	Adrien Buffort	Student	adrien.buffort@isae.fr
	Houssam Akhmouch	Student	houssam akhmouch
	Tâm Le Minh	Student	tam.le-minh@supaero.isae.fr
	Etienne Herlaut	Student	etienne.herault@isae.fr
	Naïl Eljaafari	Student	nail.eljaafari@isae.fr
Photo	Sebastian Rapp	Student	mail
Photo	Marc Melo Olivier	Student	mail

Publication

ELJAAFARI Nail — 2017-01-24T14:04:54Z

- Publication

Demo

ELJAAFARI Nail — 2017-01-24T14:04:11Z

L2F Demonstration

In order to fly, a fixed wing drone needs to use thermals so that he can gain energy and altitude. That is why we had to implement this phenomenon inside L2F to get a realistic flight mechanics for our fixed wing drone. There are many mathematical models for thermals available in the scientific literature. We can, for example, use the Allen model like represented on the graph bellow.

After choosing the thermals' model, we implement them in a flight zone. This allows us to represent realistic flight conditions for the drone. With thermals in place, we will be able to represent the trajectory of our drone piloted by a really simplistic algorithm.

With the thermals represented above, we obtain the trajectory bellow. The pilot we coded was really simplistic, the command he had was to turn left as soon has he detects a thermal. We see that the fixed wing drone doesn't have any trouble to detect the thermal but since this pilot is not a machine learning algorithm he ends his course on the ground