Structuring ROS Nodes

Modified 2020-10-19 by Andrea F. Daniele

All ROS nodes should be in the src directory of the respective package. If the node is called some_name, then the file that has its implementation should be called some_name_node.py. This file should always be executable. Furthermore, all the logic of the node should be implemented in a Python class called SomeNameNode.

The structure of the some_name_node.py should generally look like the following example (without the comments):

#!/usr/bin/env python

# import external libraries
import rospy

# import libraries which are part of the package (i.e. in the include dir)
import library

# import DTROS-related classes
from duckietown.dtros import \
    DTROS, \
    NodeType, \
    TopicType, \
    DTReminder,\
    DTParam, \
    ParamType

# import messages and services
from std_msgs.msg import Float32
from duckietown_msgs.msg import \
    SegmentList, \
    Segment, \
    BoolStamped

class SomeNameNode(DTROS):
    def __init__(self, node_name):
        # class implementation

if __name__ == '__main__':
    some_name_node = SomeNameNode(node_name='same_name_node')
    rospy.spin()

Observe that all nodes in Duckietown should inherit from the super class DTROS. This is a hard requirement. DTROS provides a lot of functionalities on top of the standard ROS nodes which make writing and debugging your node easier, and also sometimes comes with performance improvements.

In Python code, never ever do universal imports like from somepackage import *. This is an extremely bad practice. Instead, specify exactly what you are importing, i.e. from somepackage import somefunction. It is fine if you do it in __init__.py files but even there try to avoid it if possible.

When using a package that has a common practice alias, use it, e.g. import numpy as np, import matplotlib.pyplot as plt, etc. However, refrain from defining your own aliases.

The code in this node definition should be restricted as much as possible to ROS-related functionalities. If your node is performing some complex computation or has any logic that can be separated from the node itself, implement it as a separate library and put it in the include directory of the package.

Modified 2020-10-19 by Andrea F. Daniele

There are a lot of the details regarding the initalization of the node so let’s take a look at an example structure of the __init__ method of our sample node.

class SomeNameNode(DTROS):
    def __init__(self, node_name):
        super(SomeNameNode, self).__init__(
            node_name=node_name,
            node_type=NodeType.PERCEPTION
        )

        # Setting up parameters
        self.detection_freq = DTParam(
            '~detection_freq',
            param_type=ParamType.INT,
            min_value=-1,
            max_value=30
        )
        # ...

        # Generic attributes
        self.something_happened = None
        self.arbitrary_counter = 0
        # ...

        # Subscribers
        self.sub_img = rospy.Subscriber(
            'image_rect', 
            Image, 
            self.cb_img
        )
        self.sub_cinfo = rospy.Subscriber(
            'camera_info', 
            CameraInfo, 
            self.cb_cinfo
        )
        # ...

        # Publishers
        self.pub_img = rospy.Publisher(
            'tag_detections_image/compressed',
            CompressedImage,
            queue_size=1,
            dt_topic_type=TopicType.VISUALIZATION
        )
        self.pub_tag = rospy.Publisher(
            'tag_detections', 
            AprilTagDetectionArray, 
            queue_size=1,
            dt_topic_type=TopicType.PERCEPTION
        )
        # ...

Now, let’s take a look at it section by section.

Node Creation

Modified 2020-10-19 by Andrea F. Daniele

In classic ROS nodes, you would initialize a ROS node with the function rospy.init_node(...). DTROS does that for you, you simply need to pass the node name that you want to the super constructor as shown above.

DTROS supports node categorization, this is useful when you want to visualize the ROS network as a graph, where graph nodes represent ROS nodes and graph edges represent ROS topics. In such a graph, you mught want to group all the nodes working on the PERCEPTION problem together, say, to clear the clutter and make the graph easier to read. Use the parameter node_type in the super constructor of your node to do so. Use the values from the NodeType enumeration. Possible node types are the following,

GENERIC
DRIVER
PERCEPTION
CONTROL
PLANNING
LOCALIZATION
MAPPING
SWARM
BEHAVIOR
VISUALIZATION
INFRASTRUCTURE
COMMUNICATION
DIAGNOSTICS
DEBUG

Node Parameters

Modified 2020-10-19 by Andrea F. Daniele

All parameters should have names relative to the namespace of the node, i.e. they should start with ~. Also, all parameters should be in the scope of the instance, not the method, so they should always be declared inside the constructor and start with self..

The parameters should never have default values set in the code. All default values should be in the configuration file! This makes sure that we don’t end up in a situation where there are two different default values in two different files related to the node.

In classic ROS, you get the value of a parameter with rospy.get_param(...). One of the issues of the ROS implementation of parameters is that a node cannot request to be notified when a parameter’s value changes at runtime. Common solutions to this problem employ a polling strategy (which consists of querying the parameter server for changes in value at regular intervals). This is highly inefficient and does not scale. The dtros library provides a solution to this. Alternatively to using rospy.get_param(...) which simply returns you the current value of a paramter, you can create a DTParam object that automatically updates when a new value is set. Use self.my_param = DTParam("~my_param") to create a DTParam object and self.my_param.value to read its value.

Generic attributes

Modified 2020-10-19 by Andrea F. Daniele

Then we initialize all the non-ROS attributes that we will need for this class. Note that this is done before initializing the Publishers and Subscribers. The reason is that if a subscriber’s callback depends on one of these attributes, we need to define it before we use it. Here’s an example that might fail:

class CoolNode(DTROS):
   def __init__(...):
       self.sub_a = rospy.Subscriber(..., callback=cb_sth, ...)
       self.important_variable = 3.1415

   def cb_sth(self):
       self.important_variable *= 1.0

And something that is better:

class CoolNode(DTROS):
   def __init__(...):
       self.important_variable = 3.1415
       sub_a = rospy.Subscriber(..., callback=cb_sth, ...)

   def cb_sth(self):
       self.important_variable *= 1.0

Publishers and Subscribers

Modified 2020-10-19 by Andrea F. Daniele

Finally, we initialize all the Subscribers and Publishers as shown above. The dtros library automatically decorates the methods rospy.Publisher and rospy.Subscriber. By doing so, new parameters are added. All the parameters added by dtros have the prefix dt_ (e.g., dt_topic_type). Use the values from the TopicType enumeration. Possible types list is identical to the node types list above.

Modified 2020-07-08 by Aleksandar Petrov

All functions, methods, and variables in Duckietown code should be named using snake_case. In other words, only lowercase letters with spaces replaced by underscored. Do not use CamelCase. This is to be used only for class names.

The names of all subscribers should start with sub_ as in the example above. Similarly, names of publishers should start with pub_ and names of callback functions should start with cb_.

Initalizing publishers and subscribers should again always be in the scope of the instance, hence starting with self..

Modified 2020-07-08 by Aleksandar Petrov

Modified 2020-07-08 by Aleksandar Petrov

If you need to take care of something before when ROS tries to shutdown the node but before it actually shuts it down, you can implement the on_shutdown method. This is useful if you are running threads in the background, there are some files that you need to close, resources to release, or to put the robot into a safe state (e.g. to stop the wheels).

Modified 2020-10-19 by Andrea F. Daniele

Often we want to publish some information which helps us analyze the behavior and performance of the node but which does not contribute to the behavior itself. For example, in order to check how well the lane filter works, you might want to plot all the detected segments on a map of the road. However, this can be quite computationally expensive and is needed only on the rare occasion that someone wants to take a look at it.

A frequent (but bad design) way of handling that is to have a topic, to which one can publish a message, which when received will induce the node to start building a publishing the debug message. A much better way, and the one that should be used in Duckietown is to create and publish the debug message only if someone has subscribed to the debug topic. This is very easy to achieve with the help of dtros. Publishers created within a DTROS node exports the utility function anybody_listening(). Here’s an example:

if self.pub_debug_img.anybody_listening():
    debug_img = self.very_expensive_function()
    debug_image_msg = self.bridge.cv2_to_compressed_imgmsg(debug_img)
    self.pub_debug_img.publish(debug_image_msg)

Note also that all debug topics should be in the debug namespace of the node, i.e. ~debug/debug_topic_name.

Similarly, a Subscribers created within a DTROS node exports the utility function anybody_publishing() that checks whether there are nodes that are currently publishing messages.

Modified 2020-10-19 by Andrea F. Daniele

If you have operations that might take non-trivial amount of computational time, you can profile them in order to be able to analyze the performance of your node. DTROS has a special context for that which uses the same mechanism as the debug topics. Hence, if you do not subscribe to the topic with the timing information, there would be no overhead to your performance. Therefore, be generous with the use of timed sections.

The syntax looks like that:

with self.time_phase("Step 1"):
    run_step_1()

...

with self.time_phase("Step 2"):
    run_step_2()

Then, if you subscribe to ~debug/phase_times you will be able to see for each separate section detailed information about the frequency of executing it, the average time it takes, and also the exact lines of code and the file in which this section appears.

Modified 2020-10-19 by Andrea F. Daniele

If your node has at least one parameter, then it should have a configuration file. If there is a single configuration (as is the case with most nodes) this file should be called default.yaml. Assuming that our node is called some_node, the configuration files for the node should be in the config/some_node/ directory.

Every parameter used in the implementation of the node should have a default value in the configuration file. Furthermore, there should be no default values in the code. The only place where they should be defined is the configuration file.

Modified 2020-10-19 by Andrea F. Daniele

Assuming that our node is called some_node then in the launch directory of the package there should be an atomic launch file with the name some_node.launch which launches the node in the correct namespace and loads its configuration parameters.

The launch file content of most node will be identical to the following, with only the node name and package name being changed.

<launch>
  <arg name="veh"/>
  <arg name="pkg_name" value="some_package"/>
  <arg name="node_name" default="some_node"/>
  <arg name="param_file_name" default="default" doc="Specify a param file"/>

  <group ns="$(arg veh)">
    <node  name="$(arg node_name)" pkg="$(arg pkg_name)" type="$(arg node_name).py" output="screen">
      <rosparam command="load" file="$(find some_package)/config/$(arg node_name)/$(arg param_file_name).yaml"/>
    </node>
  </group>

</launch>

Modified 2020-07-08 by Aleksandar Petrov

If you have any questions about good practices in software development, join the Slack channel #info-developers.