- Designing the Interface
- Component/Library Separation
Within Rock, components are implemented in C++. They are also specified in a Ruby domain-specific language that is processed by a code generation tool, oroGen. This tool ensures that the component's interface matches its specification. It also removes most of the crude boilerplate-writing code that is the declaration in C++ of the component interfaces.
From a package point of view, components are defined in an orogen package. The
orogen packages are all placed in the
/orogen/ subdirectory of one of the
package categories. Additionally, an oroGen
package and a library can share the same basename (e.g.
drivers/orogen/hokuyo). This is even a recommended behavior when an orogen
package is mainly tied to a certain library.
Follows, for instance, some packages from one of my workspaces:
base/types base/orogen/types drivers/iodrivers_base drivers/orogen/iodrivers_base control/motor_controller gui/vizkit3d_world gui/orogen/vizkit3d_world simulation/rock_gazebo simulation/orogen/rock_gazebo simulation/orogen/underwater_camera_simulation # work on top of gui/orogen/vizkit3d_world
This page will not dwell on how to create an orogen package. Check this page out for more details on the subject.
Within oroGen packages, the components are named Task Contexts. This is historically the name chosen by the developers of the Orocos RTT toolkit, the underlying C++ component implementation used by Rock and orogen.
A component is like a black box that you configure, start, feed data to (write data) to get some data out (read data). This page will give you an overview of how to create your own, along with the most important guidelines on how to write a "good" component.
To start building a component, you have to define its interface, answering four questions:
- what is its name ?
- what data will it need as input ? (input ports)
- what data will it need as output ? (output ports)
- what data will it need as configuration ? (properties)
Once the interface is defined, you can use orogen to generate a C++ scaffold, and fill the scaffold to actually implement the component.
In summary, the overall development process looks like this:
- components are implemented and unit-tested in orogen projects (design and implementation)
- you choose how the components will actually run in your system (deployment)
- the component is being executed (by Syskit)
Designing the Interface
The component interfaces are written in the package's
.orogen file. This section
will first present the role of the major parts of a component interface, to finish
with pointers to other pages for details on how to actually write the component's interface definition.
The component name
This is the one thing that is painful to change. Changing data types, adding and removing inputs, outputs or properties is painless. The easiest way to change a component's name is to copy all the component implementation from one class to another.
- suffix all component names by
- if the oroGen package you are writing should not have more than one component,
- avoid duplicating the namespace. Almost every place where you will refer to the
component name will require you to specify the namespace, so it makes names
only longer. For instance, do not name
camera_gigevision::GigEVisionCameraTask, but simply
Note about data types
Provided you followed a few rules, oroGen is meant to be using the C++ types you are using in your libraries directly. This is by far the recommended workflow, as your components will follow the progress on the library implementation naturally, without adding to the library code a dependency on orogen-generated code.
Input and Output Ports
The list of input and output ports is obviously very system-specific. While port names are relatively easy to change, port types are a rather critical choice.
It is nonetheless generally speaking a very good idea to replicate the naming scheme used by other components in your system. We only recommend a specific naming scheme for ports that generate transformations.
Reusing the types is critical. For an input to be connected to an output, they need to share the same type. This is the reason why "superset" types such as for instance Rock's base/samples/RigidBodyState exist. They are here to allow using a full state estimate with linear and angular position and velocities to be used for e.g. a simple position.
Also, for this reason, changing a port's type is close to impossible in practice, since it requires changing the type in all the system(s) the component is being used. What is done instead is creating another port with the updated type and change step by step, which is rather effort-consuming (and therefore, avoided in practice)
What has just been said obviously applies only to ports that are meant to be connected to another component in the system. As we will see, it is a very good practice to have "status" or "debug" ports in your components. These ports will only ever be connected to a logger, and for those we usually define a data structure that is specific to the component being developed.
- group together data that belong together. Do not create 5 ports to get a piece of data because, for instance, you could not find an existing data structure for that combination. Re-combining data that belong together is difficult in an asynchronous system like Rock.
- port names need to be unique, an input cannot be named the same as an output
The property interface has a lot less constraints than the ports, especially when it comes to types. It is really only a contract between Syskit and the component, that does not involve other components in the system. However, it is generally speaking good to follow the same naming scheme than other components of the same kind in your system, for simplicity of development. As for ports, it is good to group properties that really belong together in structures, to avoid creating massive unstructured configuration interfaces.
oroGen allows you to define default parameters for simple types in the specification directly, and allows you to initialize the more complex types in the C++ code. Always do so. Always go for "safe" values as defaults, abstracting yourself from the system you are currently developing as much as possible.
Actually Writing and Updating the Interface Description
Now is a good time to go read on:
Once you have created / updated a
task_context block, run
amake to do
code generation and verify that everything is fine. Code generation will create
a C++ class for you to modify in
tasks/, matching exactly the task context name.
Implementation Detail: That class depends on a
Base class that is managed
*by orogen within the
.orogen/tasks/ folder. This is how orogen manages to update
the task interface after the first generation. There are very few operations that
require a manual change in the
tasks/ folder. They are detailed in the corresponding
We strongly recommend that you develop most of your system's functionality in libraries, instead of doing within the component framework itself. For C++, this means creating C++ library packages that are then later integrated into Rock components to expose that functionality to the system. For Ruby, this means creating Ruby packages that are then used within the Ruby layers (e.g. Syskit)
Why ? Developing libraries is a matter of "general" software engineering best practices. Robotics is a small field, software engineering is not. By doing most of your work in a framework-independent manner, you ensure that you can benefit from the much bigger ecosystem. Moreover, we haven't seen the end of the robotic frameworks. By developing libraries that are framework-independent, you ensure that you can integrate them elsewhere if needs be, cutting the time and effort by a lot.
How does Rock help the library/framework separation ? Supporting this
separation during the development process is a main design driver for the
tooling. For instance, Rock's build system -
autoproj - is not assumed to be
present by the rest of the packages. Second,
orogen exposes C++ structures
directly into the type system. The widespread approach - using IDLs - usually
end up pushing the developers to integrate code-generated structures in their
libraries thus tying them to the framework itself.
Developing libraries is covered in the libraries section. This section deals with using library-integrated functionality and using them to build data-processing components.
Once the component C++ class has been generated, you need to actually implement it. Familiarize yourself with the Rock component's
And, additionally, re-read the part about initializing properties
We recommend that you unit-test your components. Rock has a built-in support using Syskit as a backend. See this page.
This documentation also provides pages on specific topics important to robotics:
In Rock, the deployment is the process to how all these components will be split in threads, processes and machines. The oroGen layer deals with the first two, Syskit allows you to handle the third.
In general, with your first components, you will want to keep with the default deployment.