Organizing Command-Based Robot Projects

As robot code becomes more complicated, navigating, understanding, and maintaining the code takes up more and more time and energy. Making changes to the code often becomes more difficult, sometimes for reasons that have very little to do with the actual complexity of the underlying logic. For a simplified example: putting the logic for many unrelated robot functions into a single 1000-line file makes it difficult to find a specific piece of code within that file, particularly under stress at a competition. But spreading out closely related logic across dozens of tiny files is often just as difficult to navigate.

This is not a problem unique to FRC, and in fact, good organization only becomes more and more critical as software projects become bigger and bigger. The « best » organization system is a perennial topic of debate, much like the « best » programming language, but in the end, the choice (in both cases) comes down to the specific task at hand and the programmer (or programmers) implementing said task. Even in the relatively small space of FRC robot programming, there is no right answer. The best choice for a given team will depend on the nature of the specific robot code, team structure, and pure personal preference.

This article discusses various facets of command-based robot program design that advanced FRC programmers may want to be aware of when writing code. It is not a prescriptive tutorial, though it presents some recommended best practices. If this level of choice seems daunting, however, many teams have been highly successful while sticking closely to WPILib’s example code and guidelines. However, this discussion may be of interest to intermediate and advanced programmers who want to make their code not only effective, but flexible, easily changeable, and sometimes even beautiful.

Why Care About Organization?

Good code organization will rarely make or break a team’s competitive ability—but it does mean easier debugging, faster modifications, nicer-looking code, and happier programmers. While it’s impossible to define « good » organization by way of what the code looks like from the inside, it’s easier to define in terms of what the robot’s software looks like from the outside.

What Good Organization Looks Like

When code is well-designed and well-organized, the code’s internal structure is intuitive and easily comprehensible. Cumbersome boilerplate is minimized, meaning that new robot functionality can often be added with just a few lines of code. When a constant value (such as the speed of the robot’s intake) needs to be changed, it only needs to change in one place. If multiple programmers are working together, they can easily understand each others” work. Bugs are rare, since it is difficult to accidentally introduce unintended behavior (such as creating a command that does not require necessary subsystems). Implementing more advanced functions like unit tests is easier, since the code is abstracted away from the physical hardware. Programmers are happy (most of the time).

What Bad Organization Looks Like

Poorly organized code often has internal structure that makes little to no sense, even to whoever wrote it. When functionality has to be added or changed, it often breaks unrelated parts of the robot: adding automatic shooter control might introduce a bug in the climbing sequence for unclear reasons. Alternatively, the organizational framework might be so strict that it’s impossible to implement necessary behavior, requiring nasty hacks or workarounds. Many lines of boilerplate code are needed for simple robot logic. Constants are scattered across the codebase, and changing basic behavior often requires making the same change to many different files. Collaboration among multiple programmers is difficult or impossible.

Defining Commands

In larger robot codebases, multiple copies of the same command need to be used in many different places. For instance, a command that runs a robot’s intake might be used in teleop, bound to a certain button; as part of a complicated command group for an autonomous routine; and as part of a self-test sequence.

As an example, let’s look at some ways to define a simple command that simply runs the robot’s intake forward at full power until canceled.

Inline Commands

The easiest and most expressive way to do this is with a StartEndCommand:

Command runIntake = Commands.startEnd(() -> intake.set(1), () -> intake.set(0), intake);
frc2::CommandPtr runIntake = frc2::cmd::StartEnd([&intake] { intake.Set(1.0); }, [&intake] { intake.Set(0); }, {&intake});

This is sufficient for commands that are only used once. However, for a command like this that might get used in many different autonomous routines and button bindings, inline commands everywhere means a lot of repetitive code:

// RobotContainer.java
intakeButton.whileTrue(Commands.startEnd(() -> intake.set(1.0), () -> intake.set(0), intake));
Command intakeAndShoot = Commands.startEnd(() -> intake.set(1.0), () -> intake.set(0), intake)
    .alongWith(new RunShooter(shooter));
Command autonomousCommand = Commands.sequence(
    Commands.startEnd(() -> intake.set(1.0), () -> intake.set(0.0), intake).withTimeout(5.0),
    Commands.waitSeconds(3.0),
    Commands.startEnd(() -> intake.set(1.0), () -> intake.set(0.0), intake).withTimeout(5.0)
);
intakeButton.WhileTrue(frc2::cmd::StartEnd([&intake] { intake.Set(1.0); }, [&intake] { intake.Set(0); }, {&intake}));
frc2::CommandPtr intakeAndShoot = frc2::cmd::StartEnd([&intake] { intake.Set(1.0); }, [&intake] { intake.Set(0); }, {&intake})
    .AlongWith(RunShooter(&shooter).ToPtr());
frc2::CommandPtr autonomousCommand = frc2::cmd::Sequence(
  frc2::cmd::StartEnd([&intake] { intake.Set(1.0); }, [&intake] { intake.Set(0); }, {&intake}).WithTimeout(5.0_s),
  frc2::cmd::Wait(3.0_s),
  frc2::cmd::StartEnd([&intake] { intake.Set(1.0); }, [&intake] { intake.Set(0); }, {&intake}).WithTimeout(5.0_s)
);

Creating one StartEndCommand instance and putting it in a variable won’t work here, since once an instance of a command is added to a command group it is effectively « owned » by that command group and cannot be used in any other context.

Instance Command Factory Methods

One way to solve this quandary is using the « factory method » design pattern: a function that returns a new object every invocation, according to some specification. Using command composition, a factory method can construct a complex command object with merely a few lines of code.

For example, a command like the intake-running command is conceptually related to exactly one subsystem: the Intake. As such, it makes sense to put a runIntakeCommand method as an instance method of the Intake class:

Note

In this document we will name factory methods as lowerCamelCaseCommand, but teams may decide on other conventions. In general, it is recommended to end the method name with Command if it might otherwise be confused with an ordinary method (e.g. intake.run might be the name of a method that simply turns on the intake).

public class Intake extends SubsystemBase {
  // [code for motor controllers, configuration, etc.]
  // ...
  public Command runIntakeCommand() {
    // implicitly requires `this`
    return this.startEnd(() -> this.set(1.0), () -> this.set(0.0));
  }
}
frc2::CommandPtr Intake::RunIntakeCommand() {
  // implicitly requires `this`
  return this->StartEnd([this] { this->Set(1.0); }, [this] { this->Set(0); });
}

Notice how since we are in the Intake class, we no longer refer to intake; instead, we use the this keyword to refer to the current instance.

Since we are inside the Intake class, technically we can access private variables and methods directly from within the runIntakeCommand method, thus not needing intermediary methods. (For example, the runIntakeCommand method can directly interface with the motor controller objects instead of calling set().) On the other hand, these intermediary methods can reduce code duplication and increase encapsulation. Like many other choices outlined in this document, this tradeoff is a matter of personal preference on a case-by-case basis.

Using this new factory method in command groups and button bindings is highly expressive:

intakeButton.whileTrue(intake.runIntakeCommand());
Command intakeAndShoot = intake.runIntakeCommand().alongWith(new RunShooter(shooter));
Command autonomousCommand = Commands.sequence(
    intake.runIntakeCommand().withTimeout(5.0),
    Commands.waitSeconds(3.0),
    intake.runIntakeCommand().withTimeout(5.0)
);
intakeButton.WhileTrue(intake.RunIntakeCommand());
frc2::CommandPtr intakeAndShoot = intake.RunIntakeCommand().AlongWith(RunShooter(&shooter).ToPtr());
frc2::CommandPtr autonomousCommand = frc2::cmd::Sequence(
  intake.RunIntakeCommand().WithTimeout(5.0_s),
  frc2::cmd::Wait(3.0_s),
  intake.RunIntakeCommand().WithTimeout(5.0_s)
);

Adding a parameter to the runIntakeCommand method to provide the exact percentage to run the intake is easy and allows for even more flexibility.

public Command runIntakeCommand(double percent) {
    return new StartEndCommand(() -> this.set(percent), () -> this.set(0.0), this);
}
frc2::CommandPtr Intake::RunIntakeCommand() {
  // implicitly requires `this`
  return this->StartEnd([this, percent] { this->Set(percent); }, [this] { this->Set(0); });
}

For instance, this code creates a command group that runs the intake forwards for two seconds, waits for two seconds, and then runs the intake backwards for five seconds.

Command intakeRunSequence = intake.runIntakeCommand(1.0).withTimeout(2.0)
    .andThen(Commands.waitSeconds(2.0))
    .andThen(intake.runIntakeCommand(-1.0).withTimeout(5.0));
frc2::CommandPtr intakeRunSequence = intake.RunIntakeCommand(1.0).WithTimeout(2.0_s)
    .AndThen(frc2::cmd::Wait(2.0_s))
    .AndThen(intake.RunIntakeCommand(-1.0).WithTimeout(5.0_s));

This approach is recommended for commands that are conceptually related to only a single subsystem, and is very concise. However, it doesn’t fare well with commands related to more than one subsystem: passing in other subsystem objects is unintuitive and can cause race conditions and circular dependencies, and thus should be avoided. Therefore, this approach is best suited for single-subsystem commands, and should be used only for those cases.

Static Command Factories

Instance factory methods work great for single-subsystem commands. However, complicated robot actions (like the ones often required during the autonomous period) typically need to coordinate multiple subsystems at once. When we want to define an inline command that uses multiple subsystems, it doesn’t make sense for the command factory to live in any single one of those subsystems. Instead, it can be cleaner to define the command factory methods statically in some external class:

Note

The sequence and parallel static factories construct sequential and parallel command groups: this is equivalent to the andThen and alongWith decorators, but can be more readable. Their use is a matter of personal preference.

public class AutoRoutines {
    public static Command driveAndIntake(Drivetrain drivetrain, Intake intake) {
        return Commands.sequence(
            Commands.parallel(
                drivetrain.driveCommand(0.5, 0.5),
                intake.runIntakeCommand(1.0)
            ).withTimeout(5.0),
            Commands.parallel(
              drivetrain.stopCommand();
              intake.stopCommand();
            )
        );
    }
}
// TODO

Non-Static Command Factories

If we want to avoid the verbosity of adding required subsystems as parameters to our factory methods, we can instead construct an instance of our AutoRoutines class and inject our subsystems through the constructor:

public class AutoRoutines {
    private Drivetrain drivetrain;
    private Intake intake;
    public AutoRoutines(Drivetrain drivetrain, Intake intake) {
        this.drivetrain = drivetrain;
        this.intake = intake;
    }
    public Command driveAndIntake() {
        return Commands.sequence(
            Commands.parallel(
                drivetrain.driveCommand(0.5, 0.5),
                intake.runIntakeCommand(1.0)
            ).withTimeout(5.0),
            Commands.parallel(
                drivetrain.stopCommand();
                intake.stopCommand();
            )
        );
    }
    public Command driveThenIntake() {
        return Commands.sequence(
            drivetrain.driveCommand(0.5, 0.5).withTimeout(5.0),
            drivetrain.stopCommand(),
            intake.runIntakeCommand(1.0).withTimeout(5.0),
            intake.stopCommand()
        );
    }
}
// TODO

Then, elsewhere in our code, we can instantiate an single instance of this class and use it to produce several commands:

AutoRoutines autoRoutines = new AutoRoutines(this.drivetrain, this.intake);
Command driveAndIntake = autoRoutines.driveAndIntake();
Command driveThenIntake = autoRoutines.driveThenIntake();
Command drivingAndIntakingSequence = Commands.sequence(
  autoRoutines.driveAndIntake(),
  autoRoutines.driveThenIntake()
);
// TODO

Capturing State in Inline Commands

Inline commands are extremely concise and expressive, but do not offer explicit support for commands that have their own internal state (such as a drivetrain trajectory following command, which may encapsulate an entire controller). This is often accomplished by instead writing a Command class, which will be covered later in this article.

However, it is still possible to ergonomically write a stateful command composition using inline syntax, so long as we are working within a factory method. To do so, we declare the state as a method local and « capture » it in our inline definition. For example, consider the following instance command factory to turn a drivetrain to a specific angle with a PID controller:

Note

The Subsystem.run and Subsystem.runOnce factory methods sugar the creation of a RunCommand and an InstantCommand requiring this subsystem.

public Command turnToAngle(double targetDegrees) {
    // Create a controller for the inline command to capture
    PIDController controller = new PIDController(Constants.kTurnToAngleP, 0, 0);
    // We can do whatever configuration we want on the created state before returning from the factory
    controller.setPositionTolerance(Constants.kTurnToAngleTolerance);
    // Try to turn at a rate proportional to the heading error until we're at the setpoint, then stop
    return run(() -> arcadeDrive(0,-controller.calculate(gyro.getHeading(), targetDegrees)))
        .until(controller::atSetpoint)
        .andThen(runOnce(() -> arcadeDrive(0, 0)));
}
// TODO

This pattern works very well in Java so long as the captured state is « effectively final » - i.e., it is never reassigned. This means that we cannot directly define and capture primitive types (e.g. int, double, boolean) - to circumvent this, we need to wrap any state primitives in a mutable container type (the same way PIDController wraps its internal kP, kI, and kD values).

Writing Command Classes

Another possible way to define reusable commands is to write a class that represents the command. This is typically done by subclassing either Command or one of the CommandGroup classes.

Subclassing Command

Returning to our simple intake command from earlier, we could do this by creating a new subclass of Command that implements the necessary initialize and end methods.

public class RunIntakeCommand extends Command {
    private Intake m_intake;
    public RunIntakeCommand(Intake intake) {
        this.m_intake = intake;
        addRequirements(intake);
    }
    @Override
    public void initialize() {
        m_intake.set(1.0);
    }
    @Override
    public void end(boolean interrupted) {
        m_intake.set(0.0);
    }
    // execute() defaults to do nothing
    // isFinished() defaults to return false
}
// TODO

This, however, is just as cumbersome as the original repetitive code, if not more verbose. The only two lines that really matter in this entire file are the two calls to intake.set(), yet there are over 20 lines of boilerplate code! Not to mention, doing this for a lot of robot actions quickly clutters up a robot project with dozens of small files. Nevertheless, this might feel more « natural, » particularly for programmers who prefer to stick closely to an object-oriented model.

This approach should be used for commands with internal state (not subsystem state!), as the class can have fields to manage said state. It may also be more intuitive to write commands with complex logic as classes, especially for those less experienced with command composition. As the command is detached from any specific subsystem class and the required subsystem objects are injected through the constructor, this approach deals well with commands involving multiple subsystems.

Subclassing Command Groups

If we wish to write composite commands as their own classes, we may write a constructor-only subclass of the most exterior group type. For example, an intake-then-outtake sequence (with single-subsystem commands defined as instance factory methods) can look like this:

public class IntakeThenOuttake extends SequentialCommandGroup {
    public IntakeThenOuttake(Intake intake) {
        super(
            intake.runIntakeCommand(1.0).withTimeout(2.0),
            new WaitCommand(2.0),
            intake.runIntakeCommand(-1).withTimeout(5.0)
        );
    }
}
// TODO

This is relatively short and minimizes boilerplate. It is also comfortable to use in a purely object-oriented paradigm and may be more acceptable to novice programmers. However, it has some downsides. For one, it is not immediately clear exactly what type of command group this is from the constructor definition: it is better to define this in a more inline and expressive way, particularly when nested command groups start showing up. Additionally, it requires a new file for every single command group, even when the groups are conceptually related.

As with factory methods, state can be defined and captured within the command group subclass constructor, if necessary.

Summary

Approach

Primary Use Case

Single-subsystem Commands

Multi-subsystem Commands

Stateful Commands

Complex Logic Commands

Instance Factory Methods

Single-subsystem commands

Excels at them

No

Yes, but must obey capture rules

Yes

Subclassing Command

Stateful commands

Very verbose

Relatively verbose

Excels at them

Yes; may be more natural than other approaches

Static and Instance Command Factories

Multi-subsystem commands

Yes

Yes

Yes, but must obey capture rules

Yes

Subclassing Command Groups

Multi-subsystem command groups

Yes

Yes

Yes, but must obey capture rules

Yes