# PD Algorithm for a Quadrotor [Simulation]

I have a big problem trying to stabilize a quadrotor with a PD controller. The model and the program has been written in C++ and the model dynamic has been taken from this source in internet:

Well, in my code I wrote the model like in the eq. system ( see eq. 3.30 on page 21):

/* Calculate the acceleration about all 6 axis */
body_pos_current_.x_dot_2 =  ( thrust_.total / masse_ ) * ( sin( body_ang_current_.theta ) * cos( body_ang_current_.phi ) * cos( body_ang_current_.psi ) + sin( body_ang_current_.psi ) * cos( body_ang_current_.phi ) );
body_pos_current_.y_dot_2 =  ( thrust_.total / masse_ ) * ( sin( body_ang_current_.theta ) * sin( body_ang_current_.psi ) * cos( body_ang_current_.phi ) - cos( body_ang_current_.psi ) * sin( body_ang_current_.phi ) * cos( body_ang_current_.psi ) );
body_pos_current_.z_dot_2 =  ( thrust_.total / masse_ ) * ( cos( body_ang_current_.theta ) * cos( body_ang_current_.phi ) ) - 9.81;

body_ang_current_.phi_dot_2   = ( torque_.phi   / Jxx_ );
body_ang_current_.theta_dot_2 = ( torque_.theta / Jyy_ );
body_ang_current_.psi_dot_2   = ( torque_.psi   / Jzz_ );


where body_ang_current.<angle> and body_pos_current_.<position> are structures defined in a class to store position, velocities and accelerations of the model given the 4 motor velocities about all 3 axis.

$$\large \cases{ \ddot X = ( \sin{\psi} \sin{\phi} + \cos{\psi} \sin{\theta} \cos{\phi}) \frac{U_1}{m} \cr \ddot Y = (-\cos{\psi} \sin{\phi} + \sin{\psi} \sin{\theta} \cos{\phi}) \frac{U_1}{m} \cr \ddot Z = (-g + (\cos{\theta} \cos{\phi}) \frac{U_1}{m} \cr \dot p = \frac{I_{YY} - I_{ZZ}}{I_{XX}}qr - \frac{J_{TP}}{I_{XX}} q \Omega + \frac{U_2}{I_{XX}} \cr \dot q = \frac{I_{ZZ} - I_{XX}}{I_{YY}}pr - \frac{J_{TP}}{I_{YY}} p \Omega + \frac{U_3}{I_{YY}} \cr \dot r = \frac{I_{XX} - I_{YY}}{I_{ZZ}}pq - \frac{U_4}{I_{ZZ}} }$$

Once I get the accelerations above I m going to integrate them to get velocities and positions as well:

/* Get position and velocities from accelerations */
body_pos_current_.x_dot = body_pos_current_.x_dot_2 * real_duration + body_pos_previous_.x_dot;
body_pos_current_.y_dot = body_pos_current_.y_dot_2 * real_duration + body_pos_previous_.y_dot;
body_pos_current_.z_dot = body_pos_current_.z_dot_2 * real_duration + body_pos_previous_.z_dot;

body_ang_current_.phi_dot   = body_ang_current_.phi_dot_2   * real_duration + body_ang_previous_.phi_dot;
body_ang_current_.theta_dot = body_ang_current_.theta_dot_2 * real_duration + body_ang_previous_.theta_dot;
body_ang_current_.psi_dot   = body_ang_current_.psi_dot_2   * real_duration + body_ang_previous_.psi_dot;

body_pos_current_.x = 0.5 * body_pos_current_.x_dot_2 * pow( real_duration, 2 ) + ( body_pos_previous_.x_dot * real_duration ) + body_pos_previous_.x;
body_pos_current_.y = 0.5 * body_pos_current_.y_dot_2 * pow( real_duration, 2 ) + ( body_pos_previous_.y_dot * real_duration ) + body_pos_previous_.y;
body_pos_current_.z = 0.5 * body_pos_current_.z_dot_2 * pow( real_duration, 2 ) + ( body_pos_previous_.z_dot * real_duration ) + body_pos_previous_.z;

body_ang_current_.phi   = 0.5 * body_ang_current_.phi_dot_2   * pow( real_duration, 2 ) + ( body_ang_previous_.phi_dot   * real_duration ) + body_ang_previous_.phi;
body_ang_current_.theta = 0.5 * body_ang_current_.theta_dot_2 * pow( real_duration, 2 ) + ( body_ang_previous_.theta_dot * real_duration ) + body_ang_previous_.theta;
body_ang_current_.psi   = 0.5 * body_ang_current_.psi_dot_2   * pow( real_duration, 2 ) + ( body_ang_previous_.psi_dot   * real_duration ) + body_ang_previous_.psi;

/* Copy the new value into the previous one (for the next loop) */
body_pos_previous_.x = body_pos_current_.x;
body_pos_previous_.y = body_pos_current_.y;
body_pos_previous_.z = body_pos_current_.z;

body_pos_previous_.x_dot = body_pos_current_.x_dot;
body_pos_previous_.y_dot = body_pos_current_.y_dot;
body_pos_previous_.z_dot = body_pos_current_.z_dot;

body_ang_previous_.phi   = body_ang_current_.phi;
body_ang_previous_.theta = body_ang_current_.theta;
body_ang_previous_.psi   = body_ang_current_.psi;

body_ang_previous_.phi_dot   = body_ang_current_.phi_dot;
body_ang_previous_.theta_dot = body_ang_current_.theta_dot;
body_ang_previous_.psi_dot   = body_ang_current_.psi_dot;


The model seems to work well but, as like reported in many papers, is very unstable and needs some controls.

The first approach for me was to create a controller (PD) to keep the height constant without moving the quadcopter, but just putting a value (for example 3 meter) and see how it reacts.

Here the small code I tried:

/* PD Controller */
double e = ( 3.0 - body_pos_current_.z );  // 3.0 is just a try value!!!
thrust_.esum = thrust_.esum + e;
thrust_.total = 1.3 * e + 0.2 * real_duration * thrust_.esum;


The problem, as you can see here in this video, is that the copter starts falling down into the ground and not reaching the desired altitude (3.0 meters). Then it comes back again again like a spring, which is not damped. I tried already many different value for the PD controller but it seems that it doesn't affect the dynamic of the model.

Another strange thing is that it goes always to a negative point under the ground, even if I change the desired height (negative or positive).

What s wrong in my code? Could you me please point to some documents or code which is understandable and well documented to start?

Thanks

EDIT: Many thanks to your suggestion. Hi was really surprise to know, that my code had lots of potential problems and was not very efficient. So I elaborate the code as your explanation and I implementers a RK4 for the integration. After I ve read those articles: here and here I got an idea about RK and its vantage to use it in simulations and graphics PC. As an example I rewrote again the whole code:

/* Calculate the acceleration about all 6 axis */
pos_.dVel.x =  ( ( thrust_.total / masse_ ) * ( -sin( body_position_.angle.theta ) * cos( body_position_.angle.phi ) * cos( body_position_.angle.psi ) - sin( body_position_.angle.phi ) * sin( body_position_.angle.psi ) ) );
pos_.dVel.y =  ( ( thrust_.total / masse_ ) * ( sin( body_position_.angle.phi ) * cos( body_position_.angle.psi ) - cos( body_position_.angle.phi ) * sin( body_position_.angle.theta ) * sin( body_position_.angle.psi ) ) );
pos_.dVel.z =  ( ( thrust_.total / masse_ ) * ( -cos( body_position_.angle.phi ) * cos( body_position_.angle.theta ) ) - 9.81 );

pos_.dOmega.phi   = ( torque_.phi   / Jxx_ );
pos_.dOmega.theta = ( torque_.theta / Jyy_ );
pos_.dOmega.psi   = ( torque_.psi   / Jzz_ );

/* Get position and velocities from accelerations */
body_position_ = RKIntegrate( body_position_, real_duration );


which is much more clear and easy to debug. Here some useful functions I implemented:

QuadrotorController::State QuadrotorController::evaluate( const State &initial, const Derivative &d, double dt ) {

State output;
output.position.x = initial.position.x + d.dPos.x * dt;
output.position.y = initial.position.y + d.dPos.y * dt;
output.position.z = initial.position.z + d.dPos.z * dt;

output.velocity.x = initial.velocity.x + d.dVel.x * dt;
output.velocity.y = initial.velocity.y + d.dVel.y * dt;
output.velocity.z = initial.velocity.z + d.dVel.z * dt;

output.angle.phi   = initial.angle.phi   + d.dAngle.phi   * dt;
output.angle.theta = initial.angle.theta + d.dAngle.theta * dt;
output.angle.psi   = initial.angle.psi   + d.dAngle.psi   * dt;

output.omega.phi   = initial.omega.phi   + d.dOmega.phi   * dt;
output.omega.theta = initial.omega.theta + d.dOmega.theta * dt;
output.omega.psi   = initial.omega.psi   + d.dOmega.psi   * dt;

return output;

};

Derivative output;

output.dPos = sampleState.velocity;
output.dVel.x = pos_.dVel.x;
output.dVel.y = pos_.dVel.y;
output.dVel.z = pos_.dVel.z;

output.dAngle = sampleState.omega;
output.dOmega.phi   = pos_.dOmega.phi;
output.dOmega.theta = pos_.dOmega.theta;
output.dOmega.psi   = pos_.dOmega.psi;

return output;

};

const double C1 = 0.0f;
const double C2 = 0.5f, A21 = 0.5f;
const double C3 = 0.5f, A31 = 0.0f, A32 = 0.5f;
const double C4 = 1.0f, A41 = 0.0f, A42 = 0.0f, A43 = 1.0f;

const double B1 = 1.0f/6.0f, B2 = 1.0f/3.0f, B3 = 1.0f/3.0f, B4 = 1.0f/6.0f;

Derivative k1 = sampleDerivative( 0.0f, state );
Derivative k2 = sampleDerivative( C2 * dt, evaluate( state, k1 * A21, dt ) );
Derivative k3 = sampleDerivative( C3 * dt, evaluate( state, k1 * A31 + k2 * A32, dt ) );
Derivative k4 = sampleDerivative( C4 * dt, evaluate( state, k1 * A41 + k2 * A42 + k3 * A43, dt ) );

const Derivative derivativeSum = k1 * B1 + k2 * B2 + k3 * B3 + k4 * B4;

return evaluate( state, derivativeSum, dt );
}


Now I m really lost because...because the simulated qudrotor has the same behavior as before. Nevertheless I ve implemented the same PD algorithm as discussed in the paper, it stabilize on Z (height) but it get really crazy due to unstable behavior. So... I dunno what is wrong in my code and my implementation. And above all I cannot find any source in internet with a good self explaned dynamic model for a quadrotor.

Regards

At the end I discovered Simulink and the possibility to write a non linear system with it. Simply take the differential equations in the paper and put them as they are in simulink. Really strightforward. After I ve written a PID controller for my system it seems to work really good.

The ide a linearized system to implement was good but for some reason didn t have the expected reaction during simulation.

Anyway many thanks to all explanations and the suggestions that other people gave me.

Regards

I don't have any recommended reading for you. However I see a lot of potential problems here.

First, and probably most obvious, there could be a bug. I don't see anything immediately wrong in the code you have provided but then there is a lot there. You could improve the code by decomposing it further. For instance, make a simple function to do the integration for a scalar value and use it instead of re-writing it for every variable. On that note you would be well served to change your integration to RK4 if you want to do anything more than hover.

Another possible problem could be with your time step. If it's too large you will get odd behavior. Given what you are seeing I'm guessing this is actually the problem. But that really is a guess because there isn't really enough to go on here even though you have clearly attempted to provide a good explanation.

Still another problem could be one of numeric stability. The cos and sin functions in your equations are sensitive around certain values. For instance cos is troublesome around 0 and the output can fluctuate wildly with only small perturbations of the input. Given that you are only trying to hover I suspect this is not actually the problem. Even still seeking out a good numerical computing library is a good idea.

The problem could also be with the gains. In fact, if you're using a reasonable time step I suspect this is most likely the problem given the behavior you're seeing. Unfortunately PD control on a quadrotor isn't ideal because the system is non-linear and the guarantees for PID controllers only apply to linear time invariant systems. As such the best way to find gains for the system is to hand tune them. You would be well served to linearize the system and apply LQR control instead.

• Hi Daemon, your Idea to linearize the system and then to implement a LQR controller seems reasonable and could be much better than a PD controller. Do you have any example about linearization of a 2nd ODE system? Regards
– Dave
Sep 20, 2014 at 10:18

Well I saw in the first line an error.

body_pos_current_.x_dot_2 =  ( thrust_.total / masse_ ) * ( sin( body_ang_current_.theta ) * cos( body_ang_current_.phi ) * cos( body_ang_current_.psi ) + sin( body_ang_current_.psi ) * cos( body_ang_current_.phi ) );


I think it should be:

body_pos_current_.x_dot_2 =  ( thrust_.total / masse_ ) * ( sin( body_ang_current_.theta ) * cos( body_ang_current_.phi ) * cos( body_ang_current_.psi ) + sin( body_ang_current_.psi ) * sin( body_ang_current_.phi ) );


Last cos() changed to sin(), then it should represent the same as the formula

I'm not sure if this solve your problem, there may be other code-errors i didn't noticed on the first glance