Edit on Github

ROS 2 Developer Guide

This page defines the practices and policies we employ when developing ROS 2.

General Principles

Some principles are common to all ROS 2 development:

  • Shared ownership: Everybody working on ROS2 should feel ownership over all parts of the system. The original author of a chunk of code does not have any special permission or obligation to control or maintain that chunk of code. Everyone is free to propose changes anywhere, to handle any type of ticket, and to review any pull request.
  • Be willing to work on anything: As a corollary to shared ownership, everybody should be willing to take on any available task and contribute to any aspect of the system.
  • Ask for help: If you run into trouble on something, ask your fellow developers for help, via tickets, comments, or email, as appropriate.

General Practices

Some practices are common to all ROS 2 development:


When filing an issue please make sure to:

  • Include enough information for another person to understand the issue. In ROS 2, the following points are needed for narrowing down the cause of an issue. Testing with as many alternatives in each category as feasible will be especially helpful.

    • The operating system and version. Reasoning: ROS 2 supports multiple platforms, and some bugs are specific to particular versions of operating systems/compilers.
    • The installation method. Reasoning: Some issues only manifest if ROS 2 has been installed from “fat archives” or from Debians. This can help us determine if the issue is with the packaging process.
    • The specific version of ROS 2. Reasoning: Some bugs may be present in a particular ROS 2 release and later fixed. It is important to know if your installation includes these fixes.
    • The DDS/RMW implementation being used; see this page for how to determine which one. Reasoning: Communication issues may be specific to the underlying ROS middleware being used.
    • The ROS 2 client library being used. Reasoning: This helps us narrow down the layer in the stack at which the issue might be.
  • Include a list of steps to reproduce the issue.
  • In case of a bug consider to provide a short, self contained, correct (compilable), example. Issues are much more likely to be resolved if others can reproduce them easily.
  • Mention troubleshooting steps that have been tried already, including:
    • Upgrading to the latest version of the code, which may include bug fixes that have not been released yet. Select “building from source” on this page and follow the instructions to get the “master” branches.
    • Trying with a different RMW implementation. See this page for how to do that.

Pull requests

  • A pull request should only focus on one change. Separate changes should go into separate pull requests. See GitHub’s guide to writing the perfect pull request
  • A patch should be minimal in size and avoid any kind of unnecessary changes.
  • Always run CI jobs for all platforms for every pull request and include links to jobs in the pull request. (If you don’t have access to the Jenkins job someone will trigger the jobs for you.)
  • Before merging a pull request all changes should be squashed into a small number of semantic commits to keep the history clear.
    • But avoid squashing commits while a pull request is under review. Your reviewers might not notice that you made the change, thereby introducing potential for confusion. Plus, you’re going to squash before merging anyway; there’s no benefit to doing it early.
  • A minimum of 1 +1 from a fellow developer is required to consider a pull request to be approved, which is required before merging.
  • Any developer is welcome to review and approve a pull request (see General Principles).
  • When you start reviewing a pull request, comment on the pull request so that other developers know that you’re reviewing it.
  • Pull-request review is not read-only, with the reviewer making comments and then waiting for the author to address them. As a reviewer, feel free to make minor improvements (typos, style issues, etc.) in-place. As the opener of a pull-request, if you are working in a fork, checking the box to allow edits from upstream contributors will assist with the aforementioned. As a reviewer, also feel free to make more substantial improvements, but consider putting them in a separate branch (either mention the new branch in a comment, or open another pull request from the new branch to the original branch).
  • Any developer (the author, the reviewer, or somebody else) can merge any approved pull request.

Development Process

  • The default branch (in most cases the master branch) must always build, pass all tests and compile without warnings. If at any time there is a regression it is the top priority to restore at least the previous state.
  • Always build with tests enabled.
  • Always run tests locally after changes and before proposing them in a pull request. Besides using automated tests, also run the modified code path manually to ensure that the patch works as intended.
  • Always run CI jobs for all platforms for every pull request and include links to the jobs in the pull request.

kanban board (waffle.io)

To help organize the work, the core ROS 2 development team is using a kanban system hosted at waffle.io: ROS 2 kanban. This board augments the capabilities of GitHub by using labels to give a custom view into issues and pull requests across multiple repositories. The data produced and edited via waffle.io are stored in the underlying GitHub objects, so there’s no requirement to use waffle.io (or for the core team to be tied to it); it just provides a useful perspective on things.

Here’s how we’re using the columns in the board:

  • Backlog: cards (issues) that nobody is yet working on. Their order in the backlog is an approximate indicator of priority, with cards higher in the column having higher priority.
  • Ready: cards on which work will be started very soon. Cards in this column should have an owner assigned. Cards should not sit in this column for more than a few days.
  • In Progress: cards on which work is currently in progress. Cards in this column must have an owner assigned. Cards should not sit in this column for more than a week. When it is determined that a card will take longer, break it up into multiple cards and put the extras in the backlog.
  • In Review: cards for which the work is done and the relevant pull request/s is/are ready for review. Cards remain in this column during review, but if review uncovers significant extra work to be done, move the card into an earlier column as appropriate.
  • Done: cards for which the work is done, meaning that the relevant pull request/s has/have been merged. This column shows recently completed cards, for informational purposes only.

Tips for working with the kanban board:

  • Requesting permission to make changes: Simply comment on specific tickets that you want to work on it. Depending on the complexity it might be useful to describe how you want to address it. We will update the status (if you don’t have the permission) and you can start working on a pull request. If you contribute regularly we will likely just grant you permission to manage the labels etc. yourself.
  • Using markup to connect issues and pull requests: see the waffle.io FAQ: https://github.com/waffleio/waffle.io/wiki/FAQs#prs-connect-keywords
  • Doing equivalent things outside waffle.io, directly via GitHub: The column a card is in is determined by the label. The first and last column do not require a specific label. For the other column a label with the same name can be assigned.

Programming conventions

  • Defensive programming: ensure that assumptions are held as early as possible. E.g. check every return code and make sure to at least throw an exception until the case is handled more gracefully.
  • All error messages must be directed to stderr.
  • Declare variables in the narrowest scope possible.
  • Keep group of items (dependencies, imports, includes, etc.) ordered alphabetically.

C++ specific

  • Avoid using direct streaming (<<) to stdout / stderr to prevent interleaving between multiple threads.
  • Avoid using references for std::shared_ptr since that subverts the reference counting. If the original instance goes out of scope and the reference is being used it accesses freed memory.

Language Versions and Code Format

In order to achieve a consistent looking product we will all follow externally (if possible) defined style guidelines for each language. For other things like package layout or documentation layout we will need to come up with our own guidelines, drawing on current, popular styles in use now.

Additionally, where ever possible, developers should use integrated tools to allow them to check that these guidelines are followed in their editors. For example, everyone should have a PEP8 checker built into their editor to cut down on review iterations related to style.

Also where possible, packages should check style as part of their unit tests to help with the automated detection of style issues (see ament_lint_auto).


We will target C99.

We will use Python’s PEP7 for our C style guide, with some modifications and additions:


Some modifications and additions:

  • We will target C99, as we do not need to support C89 (as PEP7 recommends)
    • rationale: among other things it allows us to use both // and /* */ style comments
    • rationale: C99 is pretty much ubiquitous now
  • C++ style // comments are allowed
  • Always place literals on the left hand side of comparison operators, e.g. 0 == ret instead of ret == 0
    • rationale: ret == 0 too easily turns into ret = 0 by accident

All of the following modifications only apply if we are not writing Python modules:

  • Do not use Py_ as a prefix for everything
    • Instead use a CamelCase version of the package name or other appropriate prefix
  • The stuff about documentation strings doesn’t apply

We can use the pep7 python module for style checking:


The editor integration seems slim, we may need to look into automated checking for C in more detail.


We will target C++14, using new built-in C++14 features over Boost equivalents where ever possible.

We will use the Google C++ Style Guide, with some modifications:


Line Length

  • Line Length: our maximum line length is 100 characters.

Variable Naming

  • Global variables: use lowercase with underscores prefixed with g_
    • rationale: keep variable naming case consistent across the project
    • rationale: easy to tell the scope of a variable at a glance
    • consistency across languages

Function and Method Naming

  • Name style: Google style guide says CamelCase, but the C++ std library’s style of snake_case is also allowed
    • rationale: ROS 2 core packages currently use snake_case
      • reason: either an historical oversight or a personal preference that didn’t get checked by the linter
      • reason for not changing: retroactively changing would be too disruptive
    • other considerations:
      • cpplint.py does not check this case (hard to enforce other than with review)
      • snake_case can result in more consistency across languages
    • specific guidance:
      • for existing projects, prefer the existing style
      • for new projects, either is acceptable, but a preference for matching related existing projects is advised
      • final decision is always developer discretion
        • special cases like function pointers, callable types, etc. may require bending the rules
      • Note that classes should still use CamelCase by default

Access Control

  • Access Control: drop requirement for all class members to be private and therefore require accessors
    • rationale: this is overly constraining for user API design
    • we should prefer private members, only making them public when they are needed
    • we should consider using accessors before choosing to allow direct member access
    • we should have a good reason for allowing direct member access, other than because it is convenient for us


  • Exceptions are allowed
    • rationale: this is a new code base, so the legacy argument doesn’t apply to us
    • rationale: for user facing API’s it is more idiomatic C++ to have exceptions
    • Exceptions in destructors should be explicitly avoided
  • We should consider avoiding Exceptions if we intend to wrap the resulting API in C
    • rationale: it will make it easier to wrap in C
    • rationale: most of our dependencies in code we intend to wrap in C do not use exceptions anyways

Function-like Objects

  • No restrictions on Lambda’s or std::function or std::bind


  • Boost should be avoided until absolutely required

Comments and Doc Comments

  • Use /// and /** */ comments for documentation purposes and // style comments for notes and general comments
    • Class and Function comments should use /// and /** */ style comments
    • rationale: these are recommended for Doxygen and Sphinx in C/C++
    • rationale: mixing /* */ and // is convenient for block commenting out code which contains comments
    • Descriptions of how the code works or notes within classes and functions should use // style comments

Pointer Syntax Alignment

  • Use char * c; instead of char* c; or char *c; because of this scenario char* c, *d, *e;

Class Privacy Keywords

  • Do not put 1 space before public:, private:, or protected:, it is more consistent for all indentions to be a multiple of 2
    • rationale: most editors don’t like indentions which are not a multiple of the (soft) tab size
    • Use zero spaces before public:, private:, or protected:, or 2 spaces
    • If you use 2 spaces before, indent other class statements by 2 additional spaces
    • Prefer zero spaces, i.e. public:, private:, or protected: in the same column as the class

Nested Templates

  • Never add whitespace to nested templates
    • Prefer set<list<string>> (C++11 feature) to set<list<string> > or set< list<string> >

Always Use Braces

  • Always use braces following if, else, do, while, and for, even when the body is a single line.
    • rationale: less opportunity for visual ambiguity and for complications due to use of macros in the body

Open Versus Cuddled Braces

  • Use open braces for function, class, and struct definitions, but cuddle braces on if, else, while, for, etc...
    • Exception: when an if (or while, etc.) condition is long enough to require line-wrapping, then use an open brace (i.e., don’t cuddle).
  • When a function call cannot fit on one line, wrap at the open parenthesis (not in between arguments) and start them on the next line with a 2-space indent. Continue with the 2-space indent on subsequent lines for more arguments. (Note that the Google guide is internally contradictory on this point.)
    • Same goes for if (and while, etc.) conditions that are too long to fit on one line.

This is OK:

int main(int argc, char **argv)
  if (condition) {
    return 0;
  } else {
    return 1;

if (this && that || both) {

// Long condition; open brace
if (
  this && that || both && this && that || both && this && that || both && this && that)

// Short function call
call_func(foo, bar);

// Long function call; wrap at the open parenthesis
  foo, bar, foo, bar, foo, bar, foo, bar, foo, bar, foo, bar, foo, bar, foo, bar, foo, bar,
  foo, bar, foo, bar, foo, bar, foo, bar, foo, bar, foo, bar, foo, bar, foo, bar, foo, bar);

// Very long function argument; separate it for readability
  bar, bat);


int main(int argc, char **argv) {
  return 0;

if (this &&
    that ||
    both) {
  • Use open braces rather than excessive indention, e.g. for distinguishing constructor code from constructor initializer lists


ReturnType LongClassName::ReallyReallyReallyLongFunctionName(
  Type par_name1,  // 2 space indent
  Type par_name2,
  Type par_name3)
  DoSomething();  // 2 space indent

MyClass::MyClass(int var)
: some_var_(var),
  some_other_var_(var + 1)

Bad, weird (the google way?):

ReturnType LongClassName::ReallyReallyReallyLongFunctionName(
    Type par_name1,  // 4 space indent
    Type par_name2,
    Type par_name3) {
  DoSomething();  // 2 space indent

MyClass::MyClass(int var)
    : some_var_(var),             // 4 space indent
      some_other_var_(var + 1) {  // lined up


Most of these styles and restrictions can be checked with a combination of Google’s cpplint.py and uncrustify, though we may need to modify them slightly for our above changes.

We provide command line tools with custom configurations:

We also run other tools to detect and eliminate as many warnings as possible. Here’s a non-exhaustive list of additional things we try to do on all of our packages:

  • use compiler flags like -Wall -Wextra -Wpedantic
  • run static code analysis like cppcheck, which we have integrated in ament_cppcheck


We will target Python 3 for our development.

We will use the PEP8 guidelines for code format.

We chose the following more precise rule where PEP 8 leaves some freedom:

Tools like the (ament_)pep8 Python package should be used in unit-test and/or editor integration for checking Python code style.

The pep8 configuration used in the linter is here

Integration with editors:


We will target CMake 3.5.

Since there is not an existing CMake style guide we will define our own:

  • Use lowercase keywords (functions and macros)
  • Use empty else() and end...() commands
  • No whitespace before (‘s
  • Use two spaces of indention, do not use tabs
  • Do not use aligned indentation for parameters of multi-line macro invocations. Use two spaces only.
  • Prefer functions with set(PARENT_SCOPE) to macros
  • When using macros prefix local variables with _ or a reasonable prefix


The following rules to format the markdown syntax is intended to increase readability as well as versioning.

  • Each section title should be preceded by one empty line and succeeded by one empty line.
    • Rationale: It expedites to get an overview about the structure when screening the document.
  • Each sentence must start on a new line.
    • Rationale: For longer paragraphs a single change in the beginning makes the diff unreadable since it carries forward through the whole paragraph.
  • Each sentence can optionally be wrapped to keep each line short.
  • The lines should not have any trailing white spaces.
  • A code block must be preceded and succeeded by an empty line.
    • Rationale: Whitespace is significant only directly before and directly after fenced code blocks. Following these instructions will ensure that highlighting works properly and consistently.
  • A code block should specify a syntax after the opening triple backticks.


(Speculative, not yet used)

We will target Javascript 1.5, which seems to provide the best balance of support in browsers and languages (node.js) and new features.

We will use the airbnb Javascript Style guide:


The above repository comes with a jshintrc file which allows the style to be enforced using jshint. Editor integration for jshint include vim, emacs, Sublime Text, and others:



All packages should have some level of tests. Tests can be broken down into three main categories, System tests, Integration tests, and Unit tests.

Unit tests should always be in the package which is being tested and should make use of tools like Mock to try and test narrow parts of the code base in constructed scenarios. Unit tests should not bring in test dependencies that are not testing tools, e.g. gtest, nosetest, pytest, mock, etc...

Integration tests can test interactions between parts of the code or between parts of the code and the system. They often test software interfaces in ways that we expect the user to use them. Like Unit tests, Integration tests should be in the package which is being tested and should not bring in non-tool test dependencies unless absolutely necessary, i.e. all non-tool dependencies should only be allowed under extreme scrutiny so they should be avoided if possible.

System tests are designed to test end-to-end situations between packages and should be in their own packages to avoid bloating or coupling packages and to avoid circular dependencies.

In general minimizing external or cross package test dependencies should be avoided to prevent circular dependencies and tightly coupled test packages.

All packages should have some unit tests and possibly integration tests, but the degree to which they should have them is based on the package’s category (described later).

Test Coverage

Some packages should have a mechanism setup to capture test coverage information (if applicable to the language). Coverage tools exist for some of the languages described here including C, C++, and Python, but possibly others. When possible coverage should be measured in terms of branch coverage, as opposed to statement or function coverage.


(Planned; not yet used)

We will use the Semantic Versioning guidelines for versioning:


Anything below version 1.0.0 is free to make changes at will and for most of our near-term development this will be the case. In general though for versions less than 1.0.0 we should increment the minor (as major.minor.patch) when we break existing API and increment patch for anything else.

Another part of adhering to the Semantic Versioning guidelines is that every package must declare a public API. The declaration for most C and C++ packages is simple, it is any header that it installs, but it is acceptable to define a set of symbols which are considered private. When ever possible having private symbols in public headers should be avoided. For other languages like Python, a public API must be explicitly defined, so that it is clear what symbols can be relied on with respect to the versioning guidelines. The public API can also be extended to build artifacts like configuration variables, CMake config files, etc. as well as executables and command line options and output. Any elements of the public API should be clearly stated in the package’s documentation. If something you are using is not explicitly listed as part of the public API in the package’s documentation, then you cannot depend on it not changing between minor or patch versions.

With respect to library versioning, we will version all libraries within a package together. This means that libraries inherit their version from the package. This keeps library and package versions from diverging and shares reasoning with the policy of releasing packages which share a repository together. If you need libraries to have different versions then consider splitting them into different packages.

Filesystem Layout

The filesystem layout of packages and repositories should follow the same conventions in order to provide a consistent experience for users browsing our source code.

Package layout

  • src: contains all C and C++ code
    • Also contains C/C++ headers which are not installed
  • include: contains all C and C++ headers which are installed
    • <package name>: for all C and C++ installed headers they should be folder namespaced by the package name
  • <package_name>: contains all Python code
  • test: contains all automated tests and test data
  • doc: contains all the documentation
  • package.xml: as defined by REP-0140 (may be updated for prototyping)
  • CMakeLists.txt: only ROS packages which use CMake
  • setup.py: only ROS packages which use Python code only
  • README: README which can be rendered on Github as a landing page for the project
    • This can be as short or detailed as is convenient, but it should at least link to project documentation
    • Consider putting a CI or code coverage tag in this readme
    • It can also be .rst or anything else that Github supports
  • LICENSE: A copy of the license or licenses for this package
  • CHANGELOG.rst: REP-0132 compliant changelog

Repository layout

Each package should be in a subfolder which has the same name as the package. If a repository contains only a single package it can optionally be in the root of the repository.

The root of the repository should have a CONTRIBUTING file describing the contribution guidelines. This might include license implication when using e.g. the Apache 2 License.


(API docs are not yet being automatically generated)

All packages should have these documentation elements:

  • Description and purpose
  • Definition and description of the public API
  • Examples
  • How to build and install (should reference external tools/workflows)
  • How to build and run tests
  • How to build documentation
  • How to develop (useful for describing things like python setup.py develop)

Each package should describe itself and its purpose or how it is used in the larger scope. The description should be written, as much as possible, assuming that the reader has stumbled onto it without previous knowledge of ROS or other related projects.

Each package should define and describe its public API so that there is a reasonable expectation for users what is covered by the semantic versioning policy. Even in C and C++, where the public API can be enforced by API and ABI checking, it is a good opportunity to describe the layout of the code and the function of each part of the code.

It should be easy to take any package and from that package’s documentation understand how to build, run, build and run tests, and build the documentation. Obviously we should avoid repeating ourselves for common workflows, like build a package in a workspace, but the basic workflows should be either described or referenced.

Finally, it should include any documentation for developers. This might include workflows for testing the code using something like python setup.py develop, or it might mean describing how to make use of extension points provided by you package.


Best Practices

See the design guide

See the quality guide.

Package Categories

(Planned; not yet used)

The policies will apply differently to packages depending on their categorization. The categories are meant to give some expectation as to the quality of a package and allows us to be more strict or compliant with some packages and less so with others.

(Level 1)

This category should be used for packages which are required for a reasonable ROS system in a production environment. That is to say that after you remove development tools, build tools, and introspection tools, these packages are still left over as requirements for a basic ROS system to run. However, just because you can conceive a system which does not need a particular package does not mean that it shouldn’t be called ‘Level 1’, in fact the opposite is true. If we can imagine that any reasonable production scenario where a package would be used in some essential function, then that package should be considered for this category. However, packages which we consider essential to getting a robot up and running quickly, but is a generic solution to the problem should probably not start out as ‘Level 1’.

For Example, the packages which provide in-process communication, interprocess communication, generated message runtime code, and component lifecycle should probably all be considered ‘Level 1’. However, a package which provides pose estimation (like robot_pose_ekf) is a generic solution something that most people need, but is often replaced with a domain specific solution in production, and therefore it should probably not start out as ‘Level 1’. However, it may upgrade to it at a later date, if it proves to be a solution that people want to use in their products.

Tools, like rostopic, generally do not fall into this category, but are not categorically excluded. For example, it may be the case the tool which launches and verifies a ROS graph (something like roslaunch) may need to be considered ‘Level 1’ for use in production systems.

Package Requirements

Requirements to be considered a ‘Level 1’ package:

  • Have a strictly declared public API
  • Have API documentation coverage for public symbols
  • Have 100 percent branch code coverage from unit and integration tests
  • Have system tests which cover any scenarios covered in documentation
  • Have system tests for any corner cases encountered during testing
  • Must be >= version 1.0.0

Change Control Process

The change control process requires all changes, regardless of trivialness, must go through a pull request. This is to ensure a complete memoranda of changes to the code base. In order for a pull request to get merged:

  • Changes must be reviewed by two reviewers
  • Commits must be concise and descriptive
  • All automated tests must be run in CI on all applicable platforms (Windows, versions of Linux, OS X, ARM)
  • Code coverage must stay at 100 percent
  • Any changes which require updates to documentation must be made before merging

(Level 2)

These are packages which need to be solidly developed and might be used in production environments, but are not strictly required, or are commonly replaced by custom solutions. This can also include packages which are not yet up to ‘Level 1’ but intend to be in the future.

(Level 3)

These are packages which are useful for development purposes or introspection, but are not recommended for use in embedded products or mission critical scenarios. These packages are more lax on documentation, testing, and scope of public API’s in order to make development time lower or foster addition of new features.

(Level 4)

These are demos, tutorials, or experiments. They don’t have strict requirements, but are not excluded from having good documentation or tests. For example, this might be a tutorial package which is not intended for reuse but has excellent documentation because it serves primarily as an example to others.