ROS 2 Launch System

The launch system in ROS is responsible for helping the user describe the configuration of their system and then execute it as described. The configuration of the system includes what programs to run, where to run them, what arguments to pass them, and ROS specific conventions which make it easy to reuse components throughout the system by giving them each different configurations. Also, because the launch system is the process (or the set of processes) which executes the user’s processes, it is responsible for monitoring the state of the processes it launched, as well as reporting and/or reacting to changes in the state of those processes.

Authors: William Woodall

Date Written: 2019-09

Last Modified: 2019-09


This article describes the launch system for ROS 2, and as the successor to the launch system in ROS 1 it makes sense to summarize the features and roles of roslaunch from ROS 1 and compare them to the goals of the launch system for ROS 2.

Description of roslaunch from ROS 1

From the description of roslaunch from the wiki (

roslaunch is a tool for easily launching multiple ROS nodes locally and remotely via SSH, as well as setting parameters on the Parameter Server. It includes options to automatically respawn processes that have already died. roslaunch takes in one or more XML configuration files (with the .launch extension) that specify the parameters to set and nodes to launch, as well as the machines that they should be run on.

This description lays out the main roles of roslaunch from ROS 1 as:

  • launch nodes
  • launching nodes remotely via SSH
  • setting parameters on the parameter server
  • automatic respawning of processes that die
  • static, XML based description of the nodes to launch, parameters to set, and where to run them

Further more the wiki goes on to say (

roslaunch was designed to fit the ROS architecture of complexity via composition: write a simple system first, then combine it with other simple systems to make more complex systems. In roslaunch, this is expressed through several mechanisms:

  1. <include>s: you can easily include other .launch files and also assign them a namespace so that their names do not confict with yours.
  1. <group>s: you can group together a collection of nodes to give them the same name remappings.
  1. aliased <machine>s: you can separate machine definitions and node definitions into separate .launch files and use aliases to define which machines get used at runtime. This allows you to reuse the same node definitions for multiple robots. For example, instead of saying that a laser_assembler runs on ‘’, you can say that it runs on the ‘tilt-laser’ machine. The machine definitions then take care of which host is the ‘tilt-laser’ machine.

roslaunch also contains a variety of tools to help you write your .launch files as portably as possible. You can use the <env> tag to specify environment variables that need to be set for a particular machine or node. The $(find pkg) syntax let you specify file paths relative to a ROS package, instead of specifying their location on a particular machine. You can also use the $(env ENVIRONMENT_VARIABLE) syntax within include tags to load in .launch files based on environment variables (e.g. MACHINE_NAME).

From this, there are a few more design goals and roles for roslaunch from ROS 1:

  • composition of systems into systems of systems to manage complexity
  • use include semantic to reuse fragments rather than writing each from scratch
  • use groups to apply settings (e.g. remappings) to collections of nodes/processes/included launch files
    • also use groups with namespaces to form hierarchies
  • portability through abstraction of operating system concepts, e.g. environment variables
  • utilities to locate files on the filesystem in a relocatable and portable way, e.g. $(find <package_name>)

That covers most of the features and design goals of roslaunch from ROS 1, but in the next subsection we’ll discuss what is different for the launch system in ROS 2 due to changes in ROS 2 and how it might improve on the launch system from ROS 1.

Differences in ROS 2

One of the objectives of the launch system in ROS 2 is to emulate the features of the launch system in ROS 1, but due to architectural changes in ROS 2, some of the features, goals, and terminology need to change.

Relationship Between Nodes and Processes

In ROS 1, there could only ever be one node per process and so the goals of roslaunch from ROS 1 reflect that by using “ROS nodes” and “processes” almost interchangeably.

Even for the ROS 1 feature called ‘nodelet’ (where you could emulate having more than one node per process), the conceptual mapping from node or nodelet to process was preserved by proxy processes. For example, you would run a “NodeletManager” and then run a process for each nodelet you wanted to run in that manager. This allowed nodelet’s which exited to be detected by roslaunch from ROS 1, as well as allowing them to respond to signals that it sent to the proxy process.

Since you can have many nodes per process in ROS 2, it is no longer necessary to conflate nodes and processes. Due to this, the design and documentation for the launch system in ROS 2 will need to be clearer when talking about processes and nodes. Additionally, the way that configuration (e.g. parameters and remappings) get passed to nodes by the launch system needs to be adapted, though this part overlaps with the design documents for static remapping1 and for parameters2.

Also, since there can be multiple nodes per process, shutting down a node no longer always means sending a unix signal to a single process. Other mechanisms might need to be used to have more granular shutdown control in multi-node processes.

Launching Nodes (Processes) Remotely and Portability

The launch system in ROS 1 only really ever was supported on Linux and other Unix-like operating systems like BSD and macOS. These machines all have SSH, which is the mechanism which is specifically called out to be used when launching processes on remote machines. It also played a role in defining what you specified and how when configuring roslaunch from ROS 1 to be able to launch processes on remote machines.

In ROS 2, Windows has been added to the list of targeted platforms, and as of the writing of this document it does not support SSH natively. So unless that changes (more possible than it sounds), a different, more portable mechanism might be required to support this feature everywhere. At the very least, an alternative solution would need to be used on Windows even if SSH was still used on Unix-like operating systems.


In ROS 1, there was a global parameter server which stored all parameters and nodes would get and set all parameters through this server. The server was tightly integrated into roslaunch from ROS 1, and was also used by the other kind of parameters from ROS 1, which were called “dynamic reconfigure parameters”.

In ROS 2, there are only one kind of parameters and they work differently. In general they work more like “dynamic reconfigure parameters” from ROS 1, in that they are node specific (no truly global parameters) and they are managed by the node (the node can refuse changes and parameters can only be read and changed while the node is running). More details can be found in the parameters design document2.

There can (and probably will) still be a “global parameter server” in ROS 2, but it will simply be implemented as a node which accepts all changes and could be run along with the launch system automatically or could be invoked explicitly by the user (a la roscore from ROS 1), but it should not be required for basic functionality.

This fundamental difference in how parameters work will affect both the architecture of the launch system in ROS 2 and how users specify parameters for nodes via the launch system.

In roslaunch from ROS 1 there were only a few ways that it could react to changes in the system, and they were both related to a process “dieing” (either a clean or unclean exit):

  • respawn a process if it died
  • shutdown the whole launch system if a required process died

This is somewhere that the launch system in ROS 2 can hopefully improve on what roslaunch from ROS 1 had to offer, and it can do so by providing not only these common reactions to processes exiting, but also by providing more granular information about the process exit (and other events), and by letting the user specify arbitrary responses to these type of events.

Deterministic Startup

In the ROS 1 wiki for roslaunch, it says (

roslaunch does not guarantee any particular order to the startup of nodes – although this is a frequently requested feature, it is not one that has any particular meaning in the ROS architecture as there is no way to tell when a node is initialized.

Hopefully this is another case on which the launch system for ROS 2 can improve, at least for nodes with a lifecycle, a.k.a. Managed Nodes3. For Managed Nodes, it would not be possible to apply constraints on when something is launched, rather than how it is in roslaunch from ROS 1, where things are run in a non-deterministic order.

In order to do this, the launch system in ROS 2 will need to model the dependencies between processes and/or nodes where they exist, and the constraints on those dependencies. For example, a user might express that an image processing node has a dependency on a camera driver node with the constraint that it should not be launched (what ever the action to do that might be, e.g. run a process or something else) until the camera driver node reaches the “Active” state. These constraints can be arbitrarily defined by the user or common constraints could be modeled directly by the launch system.

Also, these constraints don’t have to be related to ROS specific events like lifecycle state changes. For example, a user might express that a plain process should be launched (in this case executed as a subprocess) after another process has been running for ten seconds. The launch system in ROS 2, could either choose to let the user define a predicate which satisfied that constraint, or it could provide a generic constraint like: “launch N seconds after another process”.

Also leveraging Managed Nodes when possible, the launch system in ROS 2 could export, aggregate and export, or react to lifecycle events of nodes. For example, it might be possible to say that a node, rather than a process, is “required” such that the launch system shutdowns if that node’s state ends up in the “Finalized” state, which would be similar to a process exiting with the “required=true” setting for roslaunch from ROS 1.

Static Description and Programmatic API

Most users of roslaunch from ROS 1 used it by defining a static XML description of what they wanted executed and which parameters they wanted to set. There is an API for roslaunch in ROS 1, but in our experience few people use this interface. We can only speculate as to why, but the API is not very well documented and is not prevalent in the tutorials and examples. Sticking strictly to the XML description has caused two different approaches to dynamic behavior/configuration to become more popular:

  • preprocessing with an XML preprocessor, like xacro or some other general purpose templating system
  • more sophisticated expressions as XML tags in the roslaunch from ROS 1 syntax, e.g. $(eval expression) (added in ROS Kinetic) or the if=$(arg ...) and unless=$(arg ...) attributes

Often when these kind of “dynamic” features are discussed the question of “why is roslaunch (from ROS 1) a static description and not a script”? The direct answer is that “it doesn’t have to be”, but the API for doing it programmatically is not very well documented or easy to use.

There are pro’s and con’s to both scripted launch files as well as static, declarative launch files, but that will be covered in its own section later in this article. But even if the preference is for a static launch file format like is common in ROS 1, it’s a goal of the launch system in ROS 2 to have a more accessible public API which is used to execute that static launch file, so a programmatic approach will always be an option.

Locating Files

It’s often the case that you need to express the location of a file when describing your system to the launch system, whether it be an executable to run, a file to be passed as an argument, or a file from which to load parameters. In the launch system for ROS 2, like the launch system for ROS 1 the concept of packages is used to group related resources and programs together to make this easier, but it will also support some other kinds of relative paths (other than just package share folders). But where ROS 1 and ROS 2 differ in this topic is how the packages will be found, which folders a package can be associated with, and therefore probably also the syntax for how to get that relative path.

Similarities with ROS 1

The previous subsection dealt with what may be different for the launch system in ROS 2, but in this subsection the similarities will be enumerated (not necessarily exhaustively). The launch system in ROS 2 will:

  • convert common ROS concepts like remapping and changing the namespace into appropriate command line arguments and configurations for nodes so the user doesn’t have to do so
  • manage complexity through composition of simpler systems (launch files)
  • allow including of other launch files
  • use groups to apply settings to collections of nodes and processes
  • provide operating system portability where possible

and possibly other things, all of which it will share in common with roslaunch from ROS 1.

Separation of Concern

The launch system can be considered in parts, separated by concern. The coarse breakdown is like so:

  • Calling Conventions for Processes and Various Styles of Nodes
  • Reporting System for Events
  • System Description and Static Analysis
  • Execution and Verification of the System Description
  • Testing

The purpose of the following sections is to enumerate what the launch system could do and the things with which it could interact, but is not the requirements list for the launch system in ROS 2. The requirements for the launch system will be enumerated in section below based on what’s possible in these sections.

Calling Conventions

In order for the launch system to execute a described system, it needs to understand how it can achieve the description. The phrase “calling conventions” is an existing phrase in Computer Science4, but this section is not talking specifically about the compiler defined calling convention, through it is appropriating the term to describe a similar relationship. In this case, the phrase “calling conventions” is meant to describe the “interface” or “contract” the launch system has with anything it is executing and monitoring. This contract covers initial execution, activity during runtime, signal handling and behavior of the launch system, and shutdown.

Operating System Processes

The most basic version of these entities, and the foundation for the other entities, are operating system processes.


For these, the launch system needs to know how to execute them, and to do that it needs:

  • name of the executable (just the name, relative path, or absolute path)
  • environment variables (PATH, LD_LIBRARY_PATH, DL_PRELOAD, etc…)
  • command line arguments
  • working directory (directory from which to execute the process)
  • launch prefix (used to inject things like gdb, valgrind, etc…)


Missing from this list is the user which should be used to execute the process. It’s possible that it would be necessary or at least useful to change the user based on the launch description. However, it can always be done in a user written script and supporting it in our Python implementation in a portable way looks to be difficult.

With this information the launch system can execute any arbitrary operating system process on the local machine.


During runtime, the launch system may monitor all operating system process’s:

  • stdout pipe
  • stderr pipe

The launch system may choose to either capture these pipes, for logging or suppressing output to the console, or it can connect the pipes to an existing pty, like the terminal’s stdout and/or stderr pipes or a null pipe (e.g. /dev/null).

When capturing the output pipes of a process, the launch system could report this data in a way that the user may process them in real-time or could pass the data through user defined filters, generating a user-handled event when the filter matches.

In addition, the launch system may interact with, or allow the user to interact with, an operating system process’s:

  • stdin pipe
  • signals (SIGINT, SIGTERM, SIGUSR1, etc…)

Regardless of how the user uses the launch system to interact with these items, they should be exposed by the launch system, which is the only entity which can interact with them directly.


If the operating system process terminates, and therefore returns a return code, the launch system will report this event and it can be handled in a user defined way. Termination covers expected termination (e.g. return from main() or use exit()) and unexpected termination (e.g. the abort trap or a segmentation fault or bus error).

Historically, ROS 1’s roslaunch allowed a few common exit handling cases:

  • require=true: if this process exits (any reason) shutdown everything else, as it’s “required”
  • respawn=true: if this process exits (any reason) restart it with the same settings as startup
    • respawn_delay=N: if restarting it, delay a number of seconds between attempts

The launch system may initiate the termination of an operating system process. This starts with the signaling of SIGINT on the child process. If this does not result in the termination of the process, then one of a few things can happen based on the configuration of the launch system:

  • after a period of time, signal SIGTERM
  • after a period of time, signal SIGKILL
  • nothing

By default, the launch system will:

  • send SIGINT
  • after a short period of time, send SIGTERM
  • after an additional short period of time, send SIGKILL

The latter two steps can be skipped, or the time until escalation can be adjusted, on a per process basis.

The launch system will initiate this process when an event (built-in or user generated) initiates shutdown, e.g. when a process with the equivalent of the require=true exit handler terminates, or when the launch system itself receives the SIGINT signal.

If the launch system itself receives the SIGTERM signal it will send the SIGKILL signal to all child processes and exit immediately.

The rationale for the previous rule is that if someone attempts to SIGTERM the launch system, they probably did so out of impatience after sending SIGINT to the launch system, and therefore the launch system should attempt to exit quickly. Exiting quickly will hopefully avoid encouraging a user to SIGKILL the launch system, which might cause the subprocesses to be improperly shutdown and perhaps even become zombie processes.

Shell Evaluation

A special case of operating system processes, shell evaluation would simply be passing shell script code as an argument to the default system shell. This is a problematic thing to support because it is hard/messy to make it portable to all operating systems.

A kind of in-between entity is an operating system process which uses shell evaluation to expand a relative executable name to an absolute path using the PATH environment variable.

Remote Operating System Processes

Any of the entities based on an operating system process can be made into a remote operating system process by simply adding the requirement information needed to gain access to the other machine and execute it. This is a feature that ROS 1’s roslaunch has, and is useful in multi machine robots.

TODO: figure out what we need to do here in terms of portability and configuration

ROS Nodes

Any operating system process can become ROS specific by having at least one ROS Node within it. Having one or more “plain” ROS nodes in a process doesn’t add new standardized ways to get information into or out of the operating system process that contains them, though ROS topics, services, parameters, etc. can be accessed during runtime. It does however, add some specific kinds of inputs during execution and it also can affect how the process reacts to signals.

This applies to “plain” ROS nodes, but there is more that the launch system can use in Managed ROS Nodes, which is described in the next section.


In addition to the “Execution” subsection of the “Operating System Processes” section, processes with ROS Nodes in them may need to consider additional elements, like:

  • “Package name + executable name” rather than “executable name + PATH” (i.e. ros2 run equivalent)
  • ROS specific environment variables (e.g. ROS_DOMAIN_ID, RMW_IMPLEMENTATION, console output formatting, etc…)
  • ROS specific command line arguments
    • Varies for single Node processes and multi Node processes
    • Change node name or namespace
    • Remap topics, services, actions, parameters, etc…
    • Initialize parameter values

The specific syntax of these extra environment variables and command line arguments are defined in other documents5 1.

In each of these cases, the ROS specific constructs can be expressed with the existing mechanisms described by the “Execution” subsection for “Operating System Processes”, i.e. the ROS specific constructs can be expanded into either command line arguments or environment variables. Therefore the launch system is able to take ROS specific declarations, e.g. “remap ‘image’ to ‘left/image’”, and convert them implicitly into terms that a normal operating system process can consume like environment variables or command line arguments, e.g. adding image:=left/image to the command line arguments. However, what a given ROS specific declaration is converted into depends on how the nodes are used within the process, but later sections will go into details about that.


During runtime a “plain” ROS node doesn’t expose anything new beyond what an operating system process does. It does have ROS topics, services, parameters, etc. but none that are standardized in a way that’s useful for the launch system at this time.

It also does not react in any special way to stdin, but processes containing ROS nodes do tend to have a signal handler for SIGINT which does a more graceful shutdown, but that is not enforced. Sending the SIGINT signal typically causes most nodes to shutdown if they are using one of the “spin” functions in rclcpp or are polling rclcpp::ok(), as is recommended.


Termination of a ROS Node (the node, not the process) is not externally observable beyond what is observed with an operating system process (the return code).

Managed ROS Nodes

For ROS nodes that have a lifecycle, a.k.a. Managed ROS Nodes3, each node will have additional runtime state, which the launch system could access and either utilize directly, pass through to the event system, or aggregate before passing it through the event system.

Building yet again on previous entities, the “Managed ROS Nodes” inherits all of the execution, runtime, and termination characteristics from normal ROS nodes and therefore operating system processes.


Managed ROS Nodes do not add any additional inputs or specific configurations at execution time on top of what “plain” ROS nodes add, at least not at this time. In the future this might change, so reference the design doc3 or future documentation on the subject.


During runtime, a Managed ROS node emits events anytime the state of the node changes. This is at least emitted on a topic, but could also be captured, aggregated, and/or communicated in other ways too. These state changes could be consumed by either the launch system itself or by the user, either of which could react to these changes.

For example, the user could express something like “when node ‘A’ enters the Active state, launch nodes ‘B’ and ‘C’” or “if node ‘A’ exits with a return code or enters the Finalized state, shutdown everything”.


Managed ROS Nodes have some additional observable effects when terminating (the node, not necessarily the process containing it). A managed node enters the Finalized state after passing through the ShuttingDown transition state on termination. Since these are state transitions, they are observable via the lifecycle event system, at least through the ROS topic lifecycle_state (subject to change, always reference the managed nodes design document3).

The mechanism for how Managed ROS Nodes transition to the Finalized state (or any other state) will not be decided in this document. Instead, the implementations or some other lifecycle specific documentation will cover that. However, you could imagine this transition could be handled by the Node itself automatically, or by launch sending a state transition request, or by launch sending a specific signal.

Process with a Single Node

In this subsection, and the following subsections of the “Calling Conventions” section, the different possible combinations of nodes and processes is explained. In each case they “inherit” any behaviors from either the “ROS nodes” or the “Managed ROS nodes” subsections above, but in these subsections the “how” of communicating ROS specific options is described in more detail.

The first is a single process with a single ROS node within it. This was likely the most commonly used type of entity launched in ROS 1, as you could only have one node per process in ROS 1. In ROS 2, this will likely be less common because you can have one to many nodes per process, but will may still be used quite a bit in the form of quickly developed scripts and drivers or GUI tools which might require control over the main thread.

Since there is only one ROS node, the command line arguments do not need to be explicit about to which node they apply. For example, changing the namespace of the single node could be expressed with the command line argument __ns:=new_namespace.

Even though there is only one node in the process, that node does not need to start when the process starts, nor does the process need to end when the node is shutdown and/or destroyed. If it is a managed node, the lifecycle of the node is best tracked using the lifecycle events. In fact, a process with a single node could start a node, run for a while, later destroy it, and then create it again.

So the biggest impact of a single node process is that the configuration, in terms of command line arguments and environment variables, can be simplified.

Process with Multiple Nodes

In a process with multiple nodes, things are much the same as with a process with a single node, but the configuration, again in terms of command line arguments and environment variables, need to be more specific in order to discriminate between the various nodes being instantiated in the process. The remapping design document1 goes into detail on how you can selectively configure multiple nodes using command line arguments, so check there for up-to-date details.

However, as an example of a process with multiple nodes, consider a program that instantiates two camera driver nodes called “camera1” and “camera2” by default. You could configure their namespaces separately by doing something like camera1:__ns:=left camera2:__ns:=right.

Dynamically loading Nodes

Dynamically loading a node means spawning it in a container process that does not know about the node until it is asked to load it. A container process is a stand alone executable that loads and executes nodes within itself.

Container Process API

While there will be standard container processes, custom container processes would allow using custom executors or client libraries. Therefore, there must be a container process API for the launch system to communicate which nodes should be loaded.

The launch system must be able tell the container process what arguments to give to a dynamically loaded node. This includes command line arguments and client library specific options (e.g. rclcpp has use_intra_process_comms). Since the launch system cannot know about all custom containers, the API must include a way to pass unknown arguments (e.g. by passing key-value pairs).

The API will not include setting environment variables per loaded node. Many languages have APIs to get environment variables, and there is no way to isolate them within a process.

The following options for an API are being considered.

API using Command Line Configuration File

One option for a container processes API is to pass a configuration file with nodes to load via the command line.


  • No waiting for an API to become available


  • Requires write access to the file system
  • Requires parsing a config file
  • Cannot tell from the outside if a container process supports this interface
  • Cannot tell if and when nodes are loaded or unloaded

This API could have very low latency to launch nodes since it does not require waiting for discovery. However, there is no way to get feedback about the success or failure of loaded nodes. There is also no way to tell a container process to unload a composable node.


Another option for a container process API is to pass configuration in via STDIN.


  • No waiting for an API to become available
  • Works with read-only file systems


  • Requires parsing a config
  • Cannot tell from the outside if a container process supports this interface
  • Cannot tell if and when nodes are loaded or unloaded
  • Cannot stop dynamically loaded nodes from reading STDIN

This API could also have very low latency to launch nodes. However, there also is no way to get feedback about the success or failure of loaded nodes. STDOUT cannot be used because a composable node logging messages to STDOUT is assumed to be very common and would conflict. Since STDIN is always available, it would be possible to unload a node via this API.

API using ROS Services or Topics

Lastly, a container process API may be defined by ROS services or topics.


  • No config file parsing
  • Works with read-only file systems
  • Can indicate if a node was successfully loaded
  • Can create API to trigger launch events
  • Can tell if a container process supports this interface


  • Must wait for the service API to become available
  • Cannot stop dynamically loaded nodes from creating the same services

This is the only option discussed which can communicate the success or failure of dynamically launched nodes. It is also the only option that allows introspection. However, this option has the highest potential delay from when the container process is spawned to when nodes may be loaded.

Proposed Container process API

This is a proposal for an API a launch system will use to interact with container processes.

Command Line Arguments

A container process must accept command line arguments including log level, remapping, and parameters. These command line arguments must not be applied to dynamically launched nodes. The launch system will pass these arguments to a container process in the same way it would pass them to a node. If a remap rule would apply to a launch service, the launch system should try to use the remapped service name instead.

ROS Services

A container process must offer all of the following services.

  • ~/_container/load_node
  • ~/_container/unload_node
  • ~/_container/list_nodes

The services are hidden to avoid colliding with user created services. load_node will be called by the launch system when a composable node is to be dynamically loaded, and unload_node destroys a composable node. list_nodes is not called by launch system, and is only provided for introspection.

  1. load_node

    If a container process is asked to load a node with a full node name matching an existing node, then it must reject the request. This is to avoid conflicts in features that assume node name uniqueness, like parameters.

    A container process must assign the node a unique id when it is loaded. The id of a loaded node instance never changes. Two nodes in the same container process must never have the same id, and there should be a significant time delay before an id is reused.

    The interface for this Service was added in ROS 2 Dashing:

  2. unload_node

    The interface for this Service was added in ROS 2 Dashing:

  3. list_nodes

    The interface for this Service was added in ROS 2 Dashing:

Exit Code

If the container process is asked to shutdown due to normal [Termination], then the exit code must be 0. If it exits due to an error then exit code must be any other number.

Parallel vs Sequential Loading of Nodes

If it is possible to load multiple nodes in parallel, then it needs to be decided how to load the nodes. The container process should load nodes as soon as it is asked. It should be up to the launch system to decide whether to load nodes in parallel or sequentially. If multiple nodes of the same type are to be launched, then the launch system should load the nodes sequentially so each is able to remap it’s name before the next is loaded. If they are of different types then the launch system may choose to try to load them in parallel, where the exact order they get loaded is determined by chance or the container process.

Registration of Composable Nodes

How Composable nodes are registered is not defined by this document. Instead, the container process is responsible for knowing how to find nodes it is asked to load. For example, a container process might use pluginlib for rclcpp nodes, or python entry points for rclpy nodes.

Event Subsystem

This section of the article covers the event subsystem within the launch system, which is responsible for generating events and reporting them to users and itself so that those events can be handled.

Categories of Events by Source

Events produced by the event subsystem of the launch system can fall broadly into two categories: events that only the launch system can directly observe and events that the launch system may relay for convenience but is directly observable by other systems too. Therefore, the events that only the launch system can observe must be exposed via the event system if we want them to be used by other applications or users.

Another way to categorize events is by their source.

Launch System Events

The most basic events are related solely to things that happen within the launch system itself. This can be as simple as a timed event, either a specific amount of time has passed, or a specific time of day has passed, or an “idle” event which might be used to implement a “call later” type of callback.

However, these events can be more specific to the launch system, like when a launch description is included, or when the launch system needs to shutdown. These events might also contain pertinent information like why a launch description was included, e.g. given as an argument to the launch system, included by another launch file, requested to be included by asynchronous request (maybe via a ROS service call), or in the case of a shutting down event, maybe why the launch system is shutting down, e.g. a required process exited, or it received the SIGINT signal.

Operating System Events

Other events will be specific to any process that is executed by the launch system, like when a process is started or when a process exits. You could also imagine events which get fired when stdout or stderr data is received from the process by the launch system, assuming it captures that information.

ROS Specific Events

ROS specific events would most likely occur in processes that launch is executing, but using ROS topics and/or services launch could observe these events and generate equivalent events within the launch event system. For example, if a process being run by launch contains a node with a life cycle, launch could observe any life cycle state transitions the node makes and create an event each time one of those transitions occur. Using this a user could, for example, wait for a node to reach the “active” state and only then start another process.

Reporting and Handling of Events

Without getting into implementation details (e.g. libraries and class hierarchies), this subsection will try to express what information should be contained within events, and how they can be accessed and what the behavior of the event system is with respect to delivery.

Information Provided by Events

Like many other event systems, the events should be capable of not only notifying that an event has occurred, but it should be able to communicate data associated with the event.

A simple example of an event without extra data might be an event for “call later”, where it doesn’t matter who initiated the “call later” or how long it has been since that occurred (though it could include that if it wished), and so this events existence is sufficient to notify waiting actions to proceed.

A simple example of an event with extra data might be a “process exited” event, which would include enough information to identify which process it was as well as the return code of the process.

Again, like many other event systems, the events should have a type (either as an attribute or as a child class) which can be used to filter events to particular handlers.

Event Handlers

Events can be handled by registering an event handler with the launch system. The only required form of event handler is one that is a function, registered locally with the launch system.

Other kinds of event handlers could be supported by building on a locally defined function. They could be something like a user-defined “lambda” defined in the description of the launch file, or even a built-in event handler function which just publishes the events as ROS messages. In the latter case, it could be either be a subscription to a topic (which needs no a priori registration with the launch system) or a service call (which was registered with the launch system a priori). So if it is a topic, the subscription object, with its callback, could be considered an event handler. In the case of a service, which would be called by the launch system and handled by a user defined service server, the service server (and it’s response) would be considered the event handler.

Handling of Events

By default, events are passed to all event handlers and there is no way for an event handler to “accept” an event to prevent it from being delivered to other events.

While event handlers have no comparison operators between one another (so no sorting), the order of delivery of events to event handlers should be deterministic and should be in the reverse order of registration, i.e. “first registered, last delivered”. Note that delivery to asynchronous event handlers (e.g. a subscription to a ROS topic for events, sent via a ROS publisher), will be sent in order, but not necessarily delivered in order.

Event Filters

Like the Qt event system, it will be possible to create event filters, which emulate the ability to accept events and prevent them from being sent “downstream”. 6

Unlike the Qt event system, an event filter is simply like any other event handler, and will not prevent other event handlers from receiving the event. Instead, each event filter will have its own list of event handlers, each of which can accept or reject an event, allowing or denying further processing of the event within the event filter, respectively.

Any event handler can be added to an event filter, but “pure event sinks” are unable to accept an event, e.g. things like a ROS topic. This is because there is no feedback mechanism, i.e. a subscription to a topic has no way of indicating if an event has been accepted or rejected as it does not have a return type. Whereas, other event handlers which are functions or lambda’s withing the launch system itself or ROS service calls can have a return type and therefore can accept or reject an event.

Sending Events

It should be possible for users of the launch system send events, in addition to the system being able to do so itself.

System Description

The system description is a declarative set of actions and reactions that describe what the user wants to launch in a format that the launch system can interpret.

The goal of the system description is to capture the intentions of the user describing the system to be launched, with as few side effects as possible. The reason for doing this is so that a launch description can be visualized and statically analyzed without actually launching the described system. Having a tool that can allow a developer to visualize and modify the launch description in a WYSIWYG (what you see is what you get) editor is an important use case for the system description.

First, this section will describe in a programming language, or text markup, agnostic way what can be expressed in the system description and how it maps to the calling conventions and event handling described in previous sections, as well as how it maps to launch system specific behaviors. After that, it will suggest how this agnostic system description can be applied to Python and XML, but also how it might be able to be extended to support other languages and markups.

Launch Descriptions

The system is described in parts which we’ll refer to here as “Launch Descriptions”. Launch descriptions are made of up of an ordered list of actions and groups of actions. It may also contain substitutions throughout the description, which are used to add some flexibility and configuration to the descriptions in a structured way.


Actions may be one of several things, and each action has a type (associated with the action’s function) and may also contain arbitrary configurations. Actions represent an intention to do something, but a launch description is parsed first, then actions are taken in order of definition later. This allows actions to be interpreted and then statically introspected without actually executing any of them unless desired.

By separating the declaration of an action from the execution of an action, tools may use the launch descriptions to do things like visualize what a launch description will do without actually doing it. The launch system will simply use the interpreted actions in the launch descriptions to actually execute the actions.

Basic actions include being able to:

  • include another launch description
  • modify the launch system configurations at the current scope
  • execute a process
  • register/unregister an event handler
  • emit an event
  • additional actions defined by extensions to the launch system

Actions may also yield more actions and groups rather than perform an actual task. For example, an action to “run a node” may end up resulting in “executing two process” or in “executing a process and registering an event handler”. These kind of actions could be thought of a launch description generators or macros, since they effectively generate the same contents as a launch description, but look like an action to the user.

This allows for more complex actions which might include, but not be limited to:

  • include a launch description from a file with a certain markup type
  • set an environment variable
  • run a single-node process
  • run a multi-node process
  • run a node container
  • run a node proxy to load into a node container
  • run a process on a remote computer
  • declare launch description arguments
    • exposed as either:
      • command line arguments for top-level launch descriptions
      • or additional arguments to the “include another launch description” action
    • stored in “launch system configuration”
  • various OS actions, e.g. touch a file, read a file, write to a file, etc…

Each of these actions would be able to generate one or more other actions. This can be used to run one or more processes with a single action statement, or to simply provide some “syntactic sugar” For example, a “run a single-node process” action might take ROS specific configurations, then expand them to generic configurations for one of the basic actions like the “execute a process” action.

Including Another Launch Description

One of the simplest actions is to include another launch description. This launch description is processed in its entirety, including parsing of any launch descriptions it includes recursively. Therefore processing of launch descriptions is in order, and depth first.

Included launch descriptions inherit all configurations of the current launch description, and any changes to the launch system configurations made in the included launch description will affect actions after the include action.

However, it should also be possible to control which configurations are inherited by an included launch description and also to “scope” an included launch description so that it cannot affect the configuration above it.

In all cases, the desired behavior may be achieved though selective use of optionally scoped group actions.

Launch System Configuration

The “modify the launch system configurations at the current scope” action mentioned above is able to mutate a local scope of configurations which can affect other actions which come after the modification. Actions may use this local state to uniformly apply certain settings to themselves.

For example, the environment variables which are set when running an operating system process would be taken from the launch system configuration, and therefore can be modified with an action. Then any “execute a process” actions which come after it will be affected by the configuration change.

Changes to the local state by included launch descriptions persist, as they should be thought of as truly included in the same file, as if you had copied the contents of the included launch description in place of the include action. To avoid this, a group without a namespace could be used to produce a push-pop effect on the launch configurations.

Execute a Process

Another basic action would be to execute a subprocess, with arguments and emitted events, as described in the calling conventions section under “operating system process”.

This action will take a few required arguments, a few optional requirements, and also take settings from the launch system configurations if they’re not explicitly given. The signature of this action should be similar to the API of Python’s function7. Basically taking things like the executable file, arguments, working directory, environment, etc. as input and reporting the return code, stdout and stderr, and any errors as emitted events.

Also, every executed process will automatically setup a few event handlers, so that the user can emit events to ask the launch system to terminate the process (following the signal escalation described in previous sections), signal the process explicitly, or write to the stdin of the process. More sophisticated calling conventions which are based on the “operating system process” may include other default event handlers.

Event Handlers

The launch description can also contain event handlers. An event handler is essentially a function which takes an event as input and returns a launch description to be included at the location of the event handler registration. The event handler will be executed asynchronously when the associated event is emitted.

There are two actions associated with event handlers, registering one and unregistering one. How event types and event handlers are represented and tracked depends on the implementation of the launch system. However, as an example, a “launch file” written in Python might represent event’s as classes which inherit from a base class. If instead the “launch file” is written in XML, event types might be expressed as a string, with launch system events using a “well-known name” and with user definable events being represented with unique strings as well. Similarly, the Python based “launch file” might use instances of objects to represent registered event handlers, therefore you might need that object to perform the unregister action. And in the XML based “launch file” the user might be required to give a unique name to all registered event handlers, so that they can unregistered with the same name later.

When an event handler finishes, it is able to return a launch description which is implicitly given to the include action at the location of the event handler’s registration. This allows an event handler to cause any action upon completion, e.g. include another launch description, unregister an event handler, emit another event, run a process, start the termination of a process by emitting an event, etc…

The lowest level of event handlers is the function which takes an event and returns a launch description. For example, a user defined event handler might look like this in Python:

# This is a made up example of an API, consider it pseudo code...

launch_description = LaunchDescription(...)
# ...

def my_process_exit_logger_callback(event: ProcessExitedEvent) -> LaunchDescription:
    print(f"process with pid '{}' exited with return code '{event.return_code}'")

    ProcessExitedEvent, my_process_exit_logger_callback, name='my_process_exit_logger')

However, to remove boilerplate code or to avoid programming in markup descriptions, common event handler patterns can be encapsulated in different event handler signatures.

For example, there might be the on_event event handler signature, which then returns a given set of actions or groups the user provides. This signature might be useful to after ten seconds start a node or include another launch file, and in XML it might look like this:

<!-- This is a made up example of a markup, consider it pseudo code... -->
<!-- Also this could be made even simpler by just having a tag which lets -->
<!-- you specify the extra actions directly, rather than emitting an event and -->
<!-- handling it, but this demonstrates the custom event handler signature -->

<emit_event after="10" type="my_custom_timer_event" />

<on_event event_type="my_custom_timer_event">
  <!-- actions to be included when this event occurs -->
  <include file="$(package-share my_package)/something.launch.xml" />
  <node pkg="my_package" executable="my_exec" />
  <group namespace="my_ns">
    <node pkg="my_package" executable="my_exec" />
Emitting Events

Another basic action that the launch system should support is the ability to emit events and if necessary declare new kinds of events before emitting them.

This feature could be used by users to filter a launch system event and then dispatch it to other user defined event handlers, or to create new or modified events based on existing events.

How events are defined is up to the implementation, but it should be possible to model the events so they can be emitted and then handled by registered event handlers.



  • can be broken into:
    • “namespace” (like roslaunch),
    • conditionals (if and unless) (see:, and
    • scope (push-pop for configurations)
    • should consider what we’re discussing to do in



  • equivalent to substitutions in ROS 1, see:
  • they’ve already been implemented in the reference implementation, they should at least be summarized as built here

Mapping to Programming Languages and Markup Languages


  • Explain in general how the features described in the previous sections would map to a programming language and/or markup language and any considerations therein.
  • How it would map to Python (likely implementation)
  • How it would map to XML (likely first markup language)

Execution and Verification of the System Description

TODO: Restructure notes on this and put them here.

Temporary summary:

Whether described via static file or programmatically, once the system is described it has to be executed, and this section will cover all of that. Most of this is already covered in the “calling conventions” section, but this section will also cover some more details about execution, and then add on to that verification (starting another discussion about what the launch system should and should not do itself). Verification is runtime assertion that mirrors the static analysis that can be done off-line.


TODO: Restructure notes on this and put them here.

Temporary summary:

In ROS 1, rostest is an important extension of roslaunch, and so far in ROS 2 we’re already using the foundation of launching (executing processes and reacting to their exit, return codes, and stdout/stderr), called ros2/launch_testing right now, to implement some tests. This section will cover how that happens and how it integrates with the static description files as well as the programmatic API, adding ROS specific concepts to what we’re already doing with ros2/launch_testing.


TODO: Reformat requirements list, possibly combine/reconcile with “separation of concerns section” (consider dropping in favor of renaming to something that implies requirements as well)

Reference Implementation Proposal

TODO: This will outline what we have and what we need to build and how it should be separated.


TODO: Anything we choose not to support in the requirements vs. the “separation of concern section”, and also any alternatives which we considered but rejected in the reference implementation proposal.