Tutorial: Getting Started

In this tutorial, we will show users how to configure fuse to combine sensor data from multiple sensors and produce a state estimate.

To start, you’ll need a launch file, a yaml configuration file, and the supporting bag file and rviz display for this tutorial:

mkdir ~/fuse_tutorials
cd ~/fuse_tutorials
wget https://locus-ros-public.s3.amazonaws.com/fuse_tutorial_data.tar.xz
md5sum fuse_tutorial_data.tar.xz # Should return ef82dceb57ed83dd597cf4973f31a175
tar -xf fuse_tutorial_data.tar.xz && rm fuse_tutorial_data.tar.xz
touch fuse_simple_tutorial.launch
touch fuse_simple_tutorial.yaml

Launch File

The core state estimation node in fuse is known as the fixed_lag_smoother_node. We’ll begin by adding it to our fuse_simple_tutorial.launch file, along with the rosbag play and rviz nodes:

<launch>
  <param name="/use_sim_time" value="true"/>

  <node name="state_estimator" pkg="fuse_optimizers" type="fixed_lag_smoother_node">
    <rosparam command="load" file="$(env HOME)/fuse_tutorials/fuse_simple_tutorial.yaml"/>
  </node>

  <node name="bag_play" pkg="rosbag" type="play" args="$(env HOME)/fuse_tutorials/turtlebot3.bag --clock -d 3"/>

  <node name="rviz" pkg="rviz" type="rviz" args="-d $(env HOME)/fuse_tutorials/fuse_tutorials.rviz"/>
</launch>

Basic Configuration

Our initial configuration will simply fuse a single wheel odometry sensor. Open up fuse_simple_tutorial.yaml and paste this into it:

# Fixed-lag smoother configuration
optimization_frequency: 20
transaction_timeout: 0.01
lag_duration: 0.5

motion_models:
  unicycle_motion_model:
    type: fuse_models::Unicycle2D

unicycle_motion_model:
  #                         x      y      yaw    vx     vy     vyaw   ax   ay
  process_noise_diagonal: [0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.1, 0.1]

sensor_models:
  initial_localization_sensor:
    type: fuse_models::Unicycle2DIgnition
    motion_models: [unicycle_motion_model]
    ignition: true
  odometry_sensor:
    type: fuse_models::Odometry2D
    motion_models: [unicycle_motion_model]

initial_localization_sensor:
  publish_on_startup: true
  #                x      y      yaw    vx     vy     vyaw    ax     ay
  initial_state: [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]
  initial_sigma: [0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.100]

odometry_sensor:
  topic: "odom"
  twist_target_frame: "base_footprint"
  linear_velocity_dimensions: ['x', 'y']
  angular_velocity_dimensions: ['yaw']

publishers:
  filtered_publisher:
    type: fuse_models::Odometry2DPublisher

filtered_publisher:
  topic: "odom_filtered"
  base_link_frame_id: "base_footprint"
  odom_frame_id: "odom"
  map_frame_id: "map"
  world_frame_id: "odom"
  publish_tf: true
  publish_frequency: 10

There’s a lot to unpack here, so we’ll look at one section at a time.

# Fixed-lag smoother configuration
optimization_frequency: 20
transaction_timeout: 0.01
lag_duration: 0.5

In this section, we specify the optimization_frequency, which is the how often we run our solver and produce a state estimate (technically, it is the frequency with which all variables in the graph are updated).

We also specify the transaction_timeout, which specifies how long we wait for motion models to be generated when adding constraints to the graph. If this time is exceeded, the constraints are not added to the graph.

The lag_duration parameter specifies the length of the smoothing window. Variables added to the fixed-lag smoother will stay in the graph for at least lag_duration seconds. After that time, old variables are removed/marginalized out.

motion_models:
  unicycle_motion_model:
    type: fuse_models::Unicycle2D

unicycle_motion_model:
  #                         x      y      yaw    vx     vy     vyaw   ax   ay
  process_noise_diagonal: [0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.1, 0.1]

This section specifies the motion (kinematic) model that we will use in this problem. As our robot is a differential-drive bot, we use a 2D unicycle model. Note that fuse supports multiple motion models to be used, but most applications will only require one.

The motion model will be used to add constraints to the graph between sensor measurements. The model we have specified is of type fuse_models::Unicycle2D, which is a plugin with its own parameters. Those parameters are specified in the next block.

The process_noise_diagonal specifies the error growth for each of our state variables when we apply the kinematic model. This is equivalent to the process noise covariance you might see in an EKF application. Here, we just specify the diagonals for that matrix.

sensor_models:
  initial_localization_sensor:
    type: fuse_models::Unicycle2DIgnition
    motion_models: [unicycle_motion_model]
    ignition: true
  odometry_sensor:
    type: fuse_models::Odometry2D
    motion_models: [unicycle_motion_model]

initial_localization_sensor:
  publish_on_startup: true
  #                x      y      yaw    vx     vy     vyaw    ax     ay
  initial_state: [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]
  initial_sigma: [0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.100]

odometry_sensor:
  topic: "odom"
  twist_target_frame: "base_footprint"
  linear_velocity_dimensions: ['x', 'y']
  angular_velocity_dimensions: ['yaw']

In this section, we specify two sensor models.

The first is an “ignition” model of type fuse_models::Unicycle2DIgnition. It is responsible for adding a constraint to our graph for the robot’s initial pose.
  • The publish_on_startup parameter will cause it to add a constraint to the graph as soon as it initializes
  • The initial_state and initial_sigma provide the starting state and covariance diagonal values
The second sensor model is of type fuse_models::Odometry2D. This particular sensor model takes in ROS nav_msgs/Odometry messages and creates graph constraints from them.
  • The topic parameter is the ROS topic on which to listen for ROS nav_msgs/Odometry messages.
  • The twist_target_frame is the frame into which we want to transform the twist (velocity) data in the incoming message. In this case, we want to transform it into the base_footprint frame.
  • The fuse_models::Odometry2D model allows users to specify which dimensions should be fused into the state estimate. In this case, we are fusing x velocity, y velocity, and yaw velocity.
publishers:
  filtered_publisher:
    type: fuse_models::Odometry2DPublisher

filtered_publisher:
  topic: "odom_filtered"
  base_link_frame_id: "base_footprint"
  odom_frame_id: "odom"
  map_frame_id: "map"
  world_frame_id: "odom"
  publish_tf: true
  publish_frequency: 10

Here, we configure the plugin that will publish our state estimate. The fuse_publishers::Odometry2DPublisher publishes a ROS nav_msgs/Odometry message, as well as a transform from the frame specified in the world_frame parameter to the frame specified in the base_link_frame_id parameter.

  • The topic is the ROS topic on which the output will be published.
  • The *_frame_id parameters specify the various coordinate frame IDs that will be used when publishing the nav_msgs/Odometry message.
  • The publish_tf parameter can be used to enable or disable publishing the transform for use by tf2.

Try running the launch file:

cd ~/fuse_tutorials
roslaunch fuse_simple_tutorial.launch

You should see the state estimate output. The covariance display for the output odom_filtered topic is not enabled by default.

Adding a Second Sensor

The example so far fuses only a single odometry source, which isn’t especially useful. In order to benefit from actual sensor fusion, we should add another sensor. In this case, we will add an IMU. We will augment our existing configuration.

# Fixed-lag smoother configuration
optimization_frequency: 20
transaction_timeout: 0.01
lag_duration: 0.5

motion_models:
  unicycle_motion_model:
    type: fuse_models::Unicycle2D

unicycle_motion_model:
  #                         x      y      yaw    vx     vy     vyaw   ax   ay
  process_noise_diagonal: [0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.1, 0.1]

sensor_models:
  initial_localization_sensor:
    type: fuse_models::Unicycle2DIgnition
    motion_models: [unicycle_motion_model]
    ignition: true
  odometry_sensor:
    type: fuse_models::Odometry2D
    motion_models: [unicycle_motion_model]
  imu_sensor:
    type: fuse_models::Imu2D
    motion_models: [unicycle_motion_model]

initial_localization_sensor:
  publish_on_startup: true
  #                x      y      yaw    vx     vy     vyaw    ax     ay
  initial_state: [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]
  initial_sigma: [0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.100, 0.100]

odometry_sensor:
  topic: "odom"
  twist_target_frame: "base_footprint"
  linear_velocity_dimensions: ['x', 'y']
  angular_velocity_dimensions: ['yaw']

imu_sensor:
  topic: "imu"
  angular_velocity_dimensions: ['yaw']
  linear_acceleration_dimensions: ['x', 'y']
  twist_target_frame: "base_footprint"

publishers:
  filtered_publisher:
    type: fuse_models::Odometry2DPublisher

filtered_publisher:
  topic: "odom_filtered"
  base_link_frame_id: "base_footprint"
  odom_frame_id: "odom"
  map_frame_id: "map"
  world_frame_id: "odom"
  publish_tf: true
  publish_frequency: 10

Note that we have added an imu_sensor section to sensor_models, and then specified the parameters for that new model.

  • The topic specifies the topic on which to listen for the sensor_msgs/IMU IMU data.
  • As with the odometry model, we can specify which state dimensions we want to fuse from this sensor. In this case, we want to fuse yaw velocity.
  • Also in keeping with the odometry model, we specify a twist_target_frame into which the incoming data must be transformed before being fused.

Now running the launch file again:

cd ~/fuse_tutorials
roslaunch fuse_simple_tutorial.launch