Control Presentation

PD/PID Control Quadcopters Control
PID Control for Quadcopter Stabilization

Andrew Gibiansky andrew.gibiansky@gmail.com
November 30, 2012
PD Control PID Control General Case
System
Suppose we have some system with the transfer function G (s ) = a0 s2 1 . + a1 s
We would like this system to follow some desired trajectory . However, we have a disturbance (due to modeling error or external factors) summarized by a term D (s ). In addition, this system is clearly not asymptotically stable.
PD Control
We can drive the difference between the true trajectory and the desired trajectory to zero. The open loop system would be: t (s ) = (s )G (s ). In order to drive the error to zero, we introduce a PD compensator: F (s ) = Kp + Kd s. The new input is given by U (s ) = F (s )(t (s ) (s )). Then, the output is t (s ) = (F (s )(t (s ) (s )) + D (s ))G (s ) = t (s ) = Kp (s ) D (s ) + Kd (s )s Kp + Kd s a1 s a0 s2
Final Value Theorem
Theorem: Suppose F (s ) is the unilateral Laplace transform of f (t ), and f (t ) converges to some value as t . Then,
t
lim f (t ) = lim sF (s ).
s 0
PD Control Results
Suppose we have a step trajectory (s ) = L {T } = T s
and a similar constant disturbance term D (s ) = L {D } = Then, the tracking error is E (s ) = (s ) t (s ) D s
a0 Ts2 + Ta1 s D s (Kp + Kd s a1 s a0 s2 )

Steady State Error

By the nal value theorem:
t
lim e (t ) = lim sE (s ) =
s 0
D . Kp
Steady state error! In physical systems, we cannot make the gain Kp arbitrarily large, so this becomes a real problem. A PD controller cannot accurately track a trajectory with a step in it. (The error comes from the fact that eventually the system has zero derivative error, so Kd is ineffective, and the constant disturbance balances out the proportional gain Kp .)
PID Control
Add an integral term: F (s ) = Kp + Kd s + Ki /s. This integral term ensures that if we remain in a steady state error for long enough, the build-up will overcome any constant disturbance. The error becomes: E (s ) = Ta0 s2 + Ta1 s D Ki + Kp s + Kd s2 a0 s3 a1 s2
Final value theorem:

t
lim e (t ) = lim sE (s ) = 0.
s 0
General Case: PD Control

Suppose we have a system with transfer function H (s )/s, such that H (0) = 0. Then, using a PD controller where F (s ) = Kp + Kd s we have E (s ) = (F (s )E (s ) + D (s ))H (s )/s (s ) E (s ) = F (s )H (s ) /s + D (s )H (s ) /s (s ) 1 F (s )H (s ) /s F (s )H (s ) + D (s )H (s ) s (s ) s F (s )H (s )
lim e (t ) = lim sE (s ) = lim s

s 0 s 0
= lim
sF (s )H (s ) + s (D /s )H (s ) s2 (s ) D = s 0 s F (s )H (s ) Kp
We have the steady state error shown before.

Andrew Gibiansky andrew.gibiansky@gmail.com PID Control for Quadcopter Stabilization
General Case: PID Control

With a PID control with F (s ) = Kp + Kd s + Ki /s, we obtain the same form of the result:
t
lim e (t ) = lim
sF (s )H (s ) + s (D /s )H (s ) s2 (s ) s 0 s F (s )H (s ) D = lim =0 s 0 Kp + Kd s + Ki /s
Thus, the result we found before generalize for any transfer function of the form H (s )/s where H (0) = 0.
Mechanical View Kinematics Equations of Motion
Quadcopters
Quadcopter Thrusts
Degrees of freedom provided by extra motors.
Coordinate Systems
Figure: Quadcopter Body Frame and Inertial Frame
Forces and Torques

Thrust force: 0 TB = Ti = k 0 . i =1 i 2
4
Drag force: kd x FD = kd y kd z Motor torques: Lk (1 2 3 2 ) B = Lk (2 2 4 2 ) 2 2 2 2 b 1 2 + 3 4

Equations of Motion: Linear

Inertial Frame: 0 = 0 + RTB + FD mx mg Body Frame: 0 = R 1 0 + TB + R 1 FD (mx ) mx mg
Equations of Motion: Rotational

Eulers Equations (Rigid Body Dynamics): + (I ) = I Body Frame Dynamics: Ixx I=0 0 0 Iyy 0 0 0 . Izz
Iyy Izz y z Ixx 1 I xxIxx = Iyy 1 IzzI x z yy 1 I I xx yy Izz x y

Izz
Simulation
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 % Simulation times, in seconds. start_time = 0; end_time = 10; dt = 0.005; times = start_time:dt:end_time; N = numel(times); % Initial simulation state. x = [0; 0; 10]; xdot = zeros(3, 1); theta = zeros(3, 1); % Simulate some disturbance in the angular velocity. % The magnitude of the deviation is in radians / second. deviation = 100; thetadot = deg2rad(2 * deviation * rand(3, 1) - deviation); % Step through the simulation, updating the state. for t = times i = input(t); % Get controller input , , ) omega = thetadot2omega(thetadot, theta); % Convert ( a = acceleration(i, theta, xdot, m, g, k, kd); % Compute linear acceleration omegadot = angular_acceleration(i, omega, I, L, b, k); % Compute angular acceleration omega = omega + dt * omegadot; % = + dt , , ) thetadot = omega2thetadot(omega, theta); % Convert ( , , ) theta = theta + dt * thetadot; % = + dt ( = x + dt a xdot = xdot + dt * a; % x x = x + dt * xdot; % x = x + dt x end
Visualization
YouTube Movie
PD Control PID Control Automatic Gain Tuning
PD Control
Control signal should be turned into a torque: = Iu (t ) = I We can set torques: + Kp I Kd xx + Kp = Iyy Kd + Kp Izz Kd
T 0 T 0 T 0
dt
dt dt
Our electronic gyro outputs angular velocities, so we integrate them to get the angle. This integral is multiplied by the proportional gain.
Motor Angular Velocities

We do not actually set torques, we set voltages over the motors. These correspond directly to the motor angular velocities i . We can solve for the inputs i = i 2 : + Kp I Kd Lk (1 3 ) xx + Kp = Lk (2 4 ) B = Iyy Kd b (1 2 + 3 4 ) + Kp Izz Kd
T 0 T 0 T 0
dt
dt dt
For the last constraint, choose one to keep the quadcopter aloft: mg T = cos cos
PD Control
Figure: Left: Angular velocities. Right: angular displacements (, , ). Note that there is some small steady-state error (approximately 0.3 ).
We can do better by adding an integral term to make this a PID controller. + Kp e = Kd + Kp e = Kd + Kp e = Kd

T 0 T 0 T 0
dt + Ki dt + Ki dt + Ki
T 0 T 0 T 0
T 0 T 0
dt dt dt dt dt dt
T 0
To avoid integral wind-up, only start integrating the double-integral once the angle deviation is relatively small.
Integral Windup
Figure: In some cases, integral wind-up can cause lengthy oscillations instead of settling. In other cases, wind-up may actually cause the system to become unstable, instead of taking longer to reach a steady state.
PID Control
Figure: With a properly implemented PID, we achieve an error of approximately 0.06 after 10 seconds.
Automatic Gain Tuning
Tuning the PID gains (Ki , Kp , and Kd ) can be difcult. Quality of results depends on gain values and the ratios of the different gain values. Best gains might be different for different modes of operation. Requires expert intuition and a lot of time to tune them. We would like to do this automatically.
Extremum Seeking
Dene a cost function, dening the quality of a set of parameters: J ( ) = 1 tf to
tf t0
e ( t , ) 2 dt
Use gradient descent to minimize this cost function in parameter-space: (k + 1) = (k ) J ( ) Approximate gradient numerically:
J ( ) =
J ( ), J ( ) . J ( ), Kp Ki Kd
K ) J ( u K ) J ( + u J ( ) K 2
Gradient Descent Tricks
Adjust step size as we go along, to become more precise as time goes on. Computing e (t , ) means running a simulation with some random initial disturbance. Use many simulations, take the average. As we go along, average more simulations. Repeat many times to produce many local minima, then choose the best one and call it the global minimum. Automatically choose a time to stop iterating (when were no longer improving our average cost).
Manual Gains vs. Automatically Tuned Gains
Questions?
Thank you to Professors Donatello Materassi and Rob Wood for their help.

Control Presentation

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Control Presentation

Загружено:

Авторское право:

Доступные форматы

PD/PID Control Quadcopters Control

PID Control for Quadcopter Stabilization

November 30, 2012

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

PD Control PID Control General Case

Suppose we have some system with the transfer function G (s ) = a0 s2 1 . + a1 s

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

PD Control PID Control General Case

Andrew Gibiansky andrew.gibiansky@gmail.com

PD/PID Control Quadcopters Control

PD Control PID Control General Case

Final Value Theorem

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

PD Control PID Control General Case

and a similar constant disturbance term D (s ) = L {D } = Then, the tracking error is E (s ) = (s ) t (s ) D s

a0 Ts2 + Ta1 s D s (Kp + Kd s a1 s a0 s2 )

Andrew Gibiansky andrew.gibiansky@gmail.com

PD/PID Control Quadcopters Control

PD Control PID Control General Case

Steady State Error

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

PD Control PID Control General Case

Final value theorem:

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

PD Control PID Control General Case

General Case: PD Control

lim e (t ) = lim sE (s ) = lim s

We have the steady state error shown before.

PD/PID Control Quadcopters Control

PD Control PID Control General Case

General Case: PID Control

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

Mechanical View Kinematics Equations of Motion

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

Mechanical View Kinematics Equations of Motion

Degrees of freedom provided by extra motors.

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

Mechanical View Kinematics Equations of Motion

Figure: Quadcopter Body Frame and Inertial Frame

Andrew Gibiansky andrew.gibiansky@gmail.com

PID Control for Quadcopter Stabilization

PD/PID Control Quadcopters Control

Mechanical View Kinematics Equations of Motion

Forces and Torques

Drag force: kd x FD = kd y kd z Motor torques: Lk (1 2 3 2 ) B = Lk (2 2 4 2 ) 2 2 2 2 b 1 2 + 3 4

PD/PID Control Quadcopters Control

Mechanical View Kinematics Equations of Motion