#Learn to fly 3 codes for learn to fly 2 how to
#Learn to fly 3 codes for learn to fly 2 series

We observe a cruise speed of 231m/s and a mean cruise altitude of 10500. In order to measure the max speed our aircraft can reach, we will set the thrust to its maximum value and observe the speed at which the plane stabilizes.

We want to assess that our plane does not exceed speeds dangerous for its structural integrity (getting too close to Mach 1 creates important turbulences and stress on the structure of the plane). Additionally, we can also notice that the acceleration values stay well under displeasing values (reaching a max of 0.33g horizontally and 0.06g vertically). Moreover the take-off time (time before leaving ground) of 25s also matches with reality (of ~30s). Finally, the last graph allows us to estimate a take-off distance of around 1800m which is close to the one we expected (~1900m). The phenomenon can be seen in similar ways for velocity, where vertical speed starts going up at 25s and the increase rate of horizontal speed (horizontal acceleration) lowers. Acceleration is only horizontal at first, and then both horizontal and vertical when Vr (rotation velocity) has been reached (~25s) and the plane rotates. On these 3 graphs, we can see the evolution over time of acceleration, velocity, and position during the take-off phase. To compute the dynamic of the plane at a given time step in the frame of reference of the ground we will use the following relations:Īcceleration, velocity and position vs time at take-off Throughout this study, we will be using the ground frame of reference. The aim of the plane model is to compute the acceleration, speed, and position (which I will be calling dynamics in the rest of the article) of the plane for any given timestep based on conditions of the previous timestep.

Fuel consumption (specific fuel consumption estimation and plane mass variation due to consumed fuel.

Drag (due to air density variation with altitude and due to the angle of attack variations).

Lift and drag coefficient (variations due to angle of attack, including stall and variations due to Mach number impact on lift and drag).

Thrust (Variations due to altitude and air rarefaction).

I have chosen to simulate an Airbus A320 because of its wide use across the world as well as the number of data available for this aircraft (The illustrations will still be made using my wooden toy plane but the figures will be based on the actual plane, such figures can be found in ).

Let’s create our airplane simulation, computing the airplane’s next position, velocity, and acceleration based on its previous position, velocity, and acceleration of the plane. I am not going to get too technical about flight mechanics and the origin of lift in this article, but if you wish to know more about this subject (Bernoulli’s Principle, Newton’s Laws and Navier-Stokes equations) I highly recommend you watch this video from David Louapre (I have added English captions for the occasion, they are currently awaiting approval) : I. In the second article ( here ), we’ll talk about the rest of the environment (states, actions, and rewards) and the reinforcement learning agent (model selection, hyperparameters optimization, and agent training and testing). We will basically be creating a basic airplane simulation. In this first article, we’ll define the transition function allowing the environment to compute the next state based on the previous state and the agent actions: in our case computing the airplane’s next position, velocity, and acceleration based on its current position, velocity and acceleration and the pilot’s action (thrust value and pitch angle). The agent will interact with the environment, selecting actions, and observing their results (new states and rewards) while trying to optimize its return (the sum of its rewards).

#Learn to fly 3 codes for learn to fly 2 how to

Reinforcement Learning is the branch of machine Learning (making algorithms learn how to do things rather than telling them how to do it) that deals with the training of an artificial intelligence through an action-and-reward process. Nowadays AI is used in all sorts of applications from biped robots to autonomous cars, but so far no AI Airplane has been seen, let’s change that!

#Learn to fly 3 codes for learn to fly 2 series

In this series of articles, I am going to write about my journey into using artificial intelligence to make a plane fly. From Icarus burning his wings to the Wright brothers soaring through the sky, it took mankind thousands of years to learn how to fly, but how long will it take an AI to do the same? Introduction