While I don’t profess to be an absolute expert at control theory, a lot of the information about the matter that I’ve found online appears to be either inaccessible to the average Joe or simply flat-out unhelpful. Control Theory is often explained with mechanical controllers that don’t really have any application to the software world, and it has been difficult for me personally to comprehend at all what is going on. In light of this, I’ve decided to do a write up about what control theory is from the perspective of an implementor, rather than that of a theorist.

That being said, control theory is still theory. That is, to understand the concept itself, you must understand the theory. I do plan on going over the practical use of control theory in code later on in this post, but I will first begin with theory.

Feel free to drop me a message on Telegram if I’ve made any errors in my explanation, because I again do not profess to be an expert at all in the subject.

The Premise of Control Theory

Control Theory is a branch of Computer Science that deals with the manipulation of a physical system. When people talk about Control Theory, they tend to use temperature control as an example. Another common use of Control Theory is in robotics, where the physical systems are the moving parts of the robot such as arms and legs. The need for Control Theory arises because the state of the environment changes during the operation of a physical mechanism. In order to control the distance that a robot moves for example, one would need to know how powerful the motor is, the frictional forces that will be experienced, such as from the surface, from the ball bearings, potentially wind resistance, etc. By taking all of these factors into account, one can can calculate parameters such as current and the time that a motor will need to run to move some distance. This model, where all the information is collected beforehand to compute a command that controls a system is called an open loop system.

Open Loop

However, the problem is that it would be impractical to have to measure or require a human to input the potentially large number of different variables that affect the system. This problem is solved by using a sensor to measure the actual output of the system (in this case, the actual distance the wheel of a robot rolls) and then sending that information back to the software, which then computes a new set of parameters to compensate for overshoot or undershoot of some target. This is called a closed-loop system, because information travels both in the direction of the system being controlled, and in the direction of the software controlling that system as well.

Closed Loop

And while you will still need to understand certain operating parameters of the physical system, you will not need to account for every possible factor that will affect the the physical result of a command.

(As a quick side note, one would use a sensor when dealing with autonomous systems. When humans are operating a system, it is the human who takes note of whatever state the system is currently in, and therefore, a system that would otherwise be an open-loop becomes a closed-loop system with human intervention)

While I will be referring to the software control of physical systems, control loops can also be built with mechanical controls in the place of the software component as well.

Abstract View of Closed-Loop Systems

Physical systems do not react instantaneously. As such, a closed-loop system will usually sample and send updates over time at fixed intervals.

At the beginning of a procedure, a goal (or target or setpoint) is defined for a system. This could be rolling a particular distance for a wheel, or a room heating up to a particular temperature, or a particular speed at which to spin an axle, etc, so long as whatever sensor is equipped on the device is capable of measuring it (e.g. an encoder or a thermocouple). Then, the error is computed by subtracting the current state of the system from the setpoint, where error is simply a metric of how far away the current state of the mechanism is from where the setpoint is. Finally, a command will be sent to compensate for the error. This will be repeated at a set time interval until the error becomes, or gets close enough to 0, meaning that the mechanism has reached its setpoint.

Abstract Loop

The process of computing the output based on the error value is known as PID (sometimes PIDF), where the letters stand for proportional, integral, derivative, and feed-forward. Because PID is the main “controller” of a mechanism, one often uses the term “PID control” or “PID loop” to refer to a closed-loop control system. On the above diagram, PID would be the “Compute output” portion of the control loop.

When you look up “PID Controller” on Wikipedia, you get this image here:

PID equation

(You will notice that there’s only 3 terms, and that I’ve dropped the F from PID several times already. I’ll get into why that is later on)

This is a good starting point at which we can start to implement a PID control loop. All the K values are simply “gain” constants, which are used to control the power (for lack of a better word) of the output. These are adjustible in order to change the order of magnitude of the output, so for example, a motor that takes an input between -1 m/s and 1 m/s would have smaller gain constants than a motor that takes input between -10 m/s to 10 m/s. A user provides these constants to the PID controller in advance. It takes some degree of testing to figure out the best combination of gain values, a topic which is out of scope of this discussion of the PID equation for the time being.

Each gain constant precedes a value that is proportional to the instantaneous error, then the accumulated error, then the rate at which the error is changing.

PID Under The Magnifying Glass

The PID formula from the previous section takes all 3 terms and adds them together to produce the final output value.

The error values that go into a PID controller are almost always some form of distance or displacement. For example, this could be the distance a robot is from a target, or the number of degrees off a desired temperature. However, error can also be a rate in, such as the current rate of a pump compared to the desired rate of a pump. The concept for both of them are exactly the same, however, and the PID computation shouldn’t need to change so long as they are using distinct PID controllers. Either way, the bottom line is that PID is simply a way to derive an output value based on the input values. The behavior of the mechanism ends up being the same one way or another anyways, so don’t fret if you’re worried about the rate being a derivative of distance - it doesn’t matter. The end result is that you want to obtain an output that gets you to where you want to go, or to a rate at which you want to go at.

The output of a PID controller is always a rate. It doesn’t really make any sense to output a distance or displacement value, because physical systems can control rate in order to affect the distance from a setpoint. For example, a PID controller can output a greater value to increase the speed at which a wheel turns, or decrease the temperature of a heating element to reduce the speed at which heat is distrubuted into a room.

The P term

The P term looks like the following:

K_p*e(t)

Meaning that it is derived from the proportional gain constant multiplied by the most recently recorded error.

The P term is the most powerful because it responds to all changes to the value of the error, and the gain constant for the proportional term is usually tuned first. Having a P term alone usually leads to erratic output values in close proximity to the error beacuse the P gain amplifies the error to the negatives if the mechanism overshoots, and vice versa.

A motion curve is typically used in robotics to show the distanced travelled by a robot over time. I’ve used a motion curve below to show the response to a PID loop controlling the state of a system, where state can be a temperature or a volume or a distance.

A system with an overly large P value looks like the following:

P unstable loop

It is often possible to acheive reasonably good results using just the P term. One can adjuts the P gain until a decent motion curve is produced, leaving all other gains to 0. This is simply called a “P controller” or “P loop.”

The I term

The I term looks like the following:

K_i*integrate(0, t, e(T), T)

Meaning that it is derived from multiplying the integral gain by the accumulated error from time 0 to the current time.

Because the I term is related to the accumulated error, the longer the robot is off target, the more powerful the I term becomes. The problem with the I term on its own is that it tends to overcompensate because the accumulated error must be reduced by an opposing physical action, which then overshoots and starts another oscillation cycle.

This can be seen in a I only term loop here:

I only loop

The most common type of control loop that I’ve personally seen is a PI loop, in which only the P and I terms are used. It’s rare to see a full PID(F) loop because it is often more work to redetermine the gain constants than it is worth the marginal gains produced by adding an extra term.

For example, consider the following P only loop, which demonstrates the ability of a single P term to produce relatively decent results:

P stable

An I term can be introduced to make a PI controller. This will smooth out the motion curve when it reaches the setpoint, and prevent a quick decceleration. You can see, however, that there is some degree of overshoot when an I term is introduced:

PI

According to PID Theory Explained:

Some amount of overshoot is always necessary for a fast system so that it could respond to changes immediately

The D term

The D term looks like the following:

K_d*derive(e(t), t)

Meaning that it is derived from the product of the derivative gain and the instantaneous rate of change for the error.

The D term is related to the rate at which error changes. The faster the mechanism moves away from its setpoint, the more powerful the D term becomes. Unlike the other two terms, the D term does not respond directly to rate, but rather the change in rate. Therefore, it is not possible to have a D only system. The effect of the D term is to reduce the rate at which the error grows, therefore flattening the error line. This is useful for more stable mechanisms that require the output to be adjusted to reduce the power of the P term.

It’s probably easier to see it on a graph. Here’s a PD loop with the D gain set to 0:

PD - D = 0

Now, compare that to a PD loop with the D gain raised slightly:

PD - D = 0.6

The D term is problematic sometimes because setting the D gain to to high of a value will cause the output to oscillate extremely rapidly. This can also happen when there are a lot of environmental disturbances that cause the D term to overreact:

PD - D = massive

In my experience, I tend to see controllers using mostly P and I terms, but usually no D term. The Wikipedia page has this to say:

Derivative action is seldom used in practice though – by one estimate in only 25% of deployed controllers

Although this claim requires citation, it is easy to see that the D term isn’t too significantly useful for the vast majority of applications where PI term controllers are used instead. When the D gain is raised, the D term tends to “fight” against the other two terms because it wants to flatten the motion curve, when the other two terms actually want the curve to get closer to the setpoint:

PD - D = 0.7

When the D term is used, it tends to be used in a full PID loop to control the I term overshoot. PD loops are also found in the wild, but again, anything with a D term tends to be quite rare compared to P and PI loops.

The F term

The F term is special because it is often left out of the “PID” initialism. In part, this is due to the fact that feed-forward is the application of a constant output, regardless of the state of the mechanism. The actual equation for a full PIDF loop looks like the following:

u(t) = K_p*e(t) + K_i*integrate(0, t, e(T), T) + K_d*derive(e(t), t) + K_f*SP

Where the F term is derived from the product of the feed-forward gain and the current setpoint.

The purpose of a feed-forward in a PIDF computation is to provide “stability.” The F term acts as kind of a base rate that pads the output from changes in the environment. Because the F term doesn’t respond to feedback from a sensor (see the derivation of the F term), it is doesn’t truly belong with the other P, I, and D terms that do depend on sensor feedback.

The other part of the reason why I think that the F term is dropped is because many systems tend to change quite often. As a matter of fact, FRC Programming Done Right has this to say about using the F term:

Feedforward control is necessary on all but the absolute simplest of systems. It’s incredibly difficult to get a good response without a feedforward calculation.

In my opinion, this is actually backwards. For simpler, more predictable systems, it is easier to use an F term because it is easier to predict and measure the various different factors that impact the performance of a specific device. This all goes back to the open-loop and closed-loop system. The simpler the system is, the easier it will be to determine the appropriate F gain. This also means that it is difficult to use for more varied, dynamic environments you’d typically find PID control loops running in, which is why PID tends to be referred to without reference to the F term.

The F term also is only really useful for constant-rate mechanisms (i.e. those that use difference in rate as the error value rather than difference in displacement as is customoary). Becase the F term cannot respond to environmental changes, having a pre-defined F term in order to drive a mechanism and using the other terms to control error to that predefined rate is the most useful application of the F term that I can personally see.

Implementation

The derivation of the PID(F) formula is actually more daunting than the implementation in code.

We can imagine the PID loop as looking something like this:

PidfController controller = // ...

// Begin the loop by setting a target
controller.setSetpoint(setpoint);

while (true) {
    // Set the last state recorded by a sensor
    controller.setLastState(state);

    // If the error is 0, the PID procedure is done
    if (controller.computeError() == 0) {
        break;
    }

    // Compute the output
    double output = controller.computePidf();

    // Send output to the mechanism
    mechanism.setOutput(output);

    // Wait a constant time interval before repeating 
    Thread.sleep(INTERVAL);
}

(Once again, it should be noted that often a PID(F) controller stops once the error is close enough to 0, in general it’s a poor idea to directly compare floating point numbers anyways, so you probably get the idea)

To write the PidfController, the easiest step is to firstly define our gain and interval constants:

static final double INTERVAL;

double pGain;
double iGain;
double dGain;
double fGain;

The next portion would be to implement a way to calculate error. We will need a way to define the setpoint to which the PID(F) controller is to travel to, and then a way to define the last state recorded by a sensor:

double setpoint;
double lastState;

// Called by a user to set the setpoint
public void setSetpoint(double setpoint) {
    this.setpoint = setpoint;
}

// Called by a sensor to set the state of the system
public void setLastState(double lastState) {
    this.lastState = lastState;
}

public double computeError() {
    return this.setpoint - this.lastState;
}

Now that we have our gain constants a way to compute the current error, we can now head to computing each term.

Proportional:

double computeProportional() {
    double error = this.computeError();
    return this.pGain * error;
}

For integral, we need to hold the amount of error that has accummulated since the PID loop began. This can be acheived by using a variable to hold the error multiplied by the interval time each time the loop runs:

double accumulatedError;

double computeIntegral() {
    double error = this.computeError();
    this.accumulatedError += error * INTERVAL;

    return this.iGain * this.accumulatedError;
}

For derivative, since the rate of change of the error is instantaneous, we can simply use the slope formula for the last recorded error and the current error taken over the interval time:

double lastError;

double computeDerivative() {
    double error = this.computeError();
    double derivative = (error - this.lastError) / INTERVAL;

    this.lastError = error;

    return this.dGain * derivative;
}

For feed-forward, it is the gain multiplied by setpoint:

double computeFeedforward() {
    return this.fGain * this.setpoint;
}

Finally, tying everything together:

double computePidf() {
    return this.computeProportional()
            + this.computeDerivative()
            + this.computeIntegral()
            + this.computeFeedforward();
}

And then cleanup our code and write it all into a coherent class:

public class PidfController {
    private static final double INTERVAL = 0.5;

    private final double pGain;
    private final double iGain;
    private final double dGain;
    private final double fGain;

    private double setpoint;
    private double lastState;

    private double accumulatedError;
    private double lastError;

    public PidfController(double pGain, double iGain, double dGain, double fGain) {
        this.pGain = pGain;
        this.iGain = iGain;
        this.dGain = dGain;
        this.fGain = fGain;
    }

    public void setSetpoint(double setpoint) {
        this.setpoint = setpoint;
    }

    public void setLastState(double lastState) {
        this.lastState = lastState;
    }

    public double computeError() {
        return this.setpoint - this.lastState;
    }

    private double computeProportional() {
        double error = this.computeError();
        return this.pGain * error;
    }

    private double computeIntegral() {
        double error = this.computeError();
        this.accumulatedError += error * INTERVAL;

        return this.iGain * this.accumulatedError;
    }

    private double computeDerivative() {
        double error = this.computeError();
        double derivative = (error - this.lastError) / INTERVAL;

        this.lastError = error;

        return this.dGain * derivative;
    }

    private double computeFeedforward() {
        return this.fGain * this.setpoint;
    }

    public double computePidf() {
        return this.computeProportional()
                + this.computeDerivative()
                + this.computeIntegral()
                + this.computeFeedforward();
    }
}

Users should change the INTERVAL time as they see appropriate.

It should be noted that all the gain values can even be changed during the PID loop, and so it is not necessary for them to be final. The INTERVAL time doesn’t even need to be constant, so long as one takes into account the time that elapses between each iteration of the PID loop. That being said, this is simply just the bare-bones minimum code, and it is up to the implementor to determine if those features are needed.

As far as units go, it is not necessarily required to have a consistent unit conversion. A PIDF loop simply outputs the rate as it relates to the gain constants and the error value. It is up to users to determine if any consistency is really required here.

Conclusion

Control Theory and control engineering are broad topics that I’ve only briefly touched over in this blog post. With a better understanding of the theory and implementation of PID control, one can build more precise physical mechanisms such as robots, temperature controls, pumps, etc.

I’ve done my best to speak on general terms so that the concepts can be applied to the widest selection of scenarios possible. Sometimes, that has made it more difficult to undestand what I’m talking about, but I hope that giving several real world examples to demonstrate a single concept has made it easier to grasp.

Again, I will reiterate that I’m by no means an expert in the field of Control Theory. Feel free to contact me for corrections, but I’m definitely not the right person to ask if you need clarification, as I’ve done my best already to try and clarify everything up front.

For FRC Robots, you definitely want to use PIDController over the hand-rolled implementation I have here, by the way.

I probably won’t be doing any more posts about robotics unless some extenuating circumstance forces me to.