Fusion of Wheel Odometry and IMU
Wheel odometry estimates the robot's position and orientation changes by monitoring wheel rotation, but it is susceptible to special conditions such as slipping and uneven terrain. An IMU measures acceleration and angular velocity, providing high-frequency motion information, but it is prone to integration drift and noise. Fusing their data compensates for each other's shortcomings, delivering more reliable and accurate localization results by leveraging their respective strengths and reducing limitations. In ROS2, the robot_localization package is available to conveniently help developers achieve this fusion.
1. Introduction to robot_localization
robot_localization is a ROS2 package used for multi-sensor data fusion in robot localization and navigation. It provides a solution that fuses data from different sensors (such as wheel odometry, IMU, GPS, LiDAR, etc.) to achieve more accurate and robust localization results. In this package, the fusion algorithm is already encapsulated within nodes, and the user only needs to inject parameters when calling the node.
Please use the following command to install robot_localization:
sudo apt install ros-<distro>-robot-localization
Please replace <distro> with the codename of the ROS2 distribution you are using.
2. robot_localization Node Description
The robot_localization package provides multiple nodes, including ekf_node, ukf_node, and navsat_transform_node, to implement robot state estimation and localization. The different nodes are described as follows:
ekf_nodeis a node based on the Extended Kalman Filter (EKF) that fuses data from multiple sensors for state estimation. It receives data from sensors such as odometry, IMU, and LiDAR, and fuses them together to provide a more accurate estimate of the robot's state. The EKF node uses the robot's motion model and sensor measurement model to estimate state variables such as the robot's position, orientation, and velocity.ukf_nodeis a node based on the Unscented Kalman Filter (UKF), used for robot state estimation and localization. The UKF node employs the Unscented Transform to handle nonlinear systems, providing more accurate state estimation by estimating the robot's position and orientation in space. Compared to the EKF, the UKF node can better handle estimation problems in nonlinear systems.navsat_transform_nodeis used to fuse Global Navigation Satellite System (GNSS) navigation satellite data with Inertial Measurement Unit (IMU) data to provide precise position estimation. This node corrects errors in the GNSS signal and converts it into position information with high accuracy. This helps improve robot localization accuracy in indoor environments or areas with unstable GNSS signals.
These nodes are core components of the robot_localization package. Each node can fuse any number of sensors (IMU, odometry, indoor positioning systems, GPS receivers, etc.). They support nav_msgs/msg/Odometry (position, orientation, linear and angular velocity), geometry_msgs/msg/PoseWithCovarianceStamped (position and orientation), geometry_msgs/msg/TwistWithCovarianceStamped (linear and angular velocity), and sensor_msgs/msg/Imu (orientation, angular velocity, and linear acceleration) messages to track the robot's 15-dimensional state (x, y, z, roll, pitch, yaw, ẋ, ẏ, ż, roll̇, pitcḣ, yaẇ, ẍ, ÿ, z̈).
Based on these measurement data, the state estimation node publishes the filtered position, orientation, and linear and angular velocities (nav_msgs/Odometry) to the /odometry/filtered topic, and when enabled, publishes the filtered acceleration to the /accel/filtered topic.
Additionally, they can (enabled by default) publish the corresponding transforms as tf2 transforms, whether it's the odom → base_link transform or the map → odom transform.
3.robot_localization Usage Example
Assuming you want to fuse wheel odometry with an IMU, first ensure that the robot chassis can publish odometry messages and IMU messages. In the MyCar navigation robot, the Stm32 chassis meets this requirement, but the Arduino chassis does not include an IMU sensor and needs to be installed separately. For fusion, you can use the ekf_node or ukf_node node. In this case, a stm32 4-wheel chassis combined with ekf_node is used. The general steps are as follows:
- Write the launch file.
- Write the configuration file.
- Compile;
- Execute and view the results.
Since robot_localization provides a well-packaged implementation for sensor fusion, the overall implementation process is relatively straightforward.
(1) Write the launch file
Create a new file named driver_ekf.launch.py in the launch directory of the ros2_stm32_bridge package, and enter the following content:
from launch import LaunchDescription
from ament_index_python.packages import get_package_share_directory
import launch_ros.actions
import os
def generate_launch_description():
MYCAR_MODEL = os.environ['MYCAR_MODEL']return LaunchDescription([
launch_ros.actions.Node(
package='robot_localization',
executable='ekf_node',
name='ekf_filter_node',
output='screen',
parameters=[os.path.join(get_package_share_directory("ros2_stm32_bridge"), 'params', 'ekf.yaml')],
),
launch_ros.actions.Node(
package='tf2_ros',
executable='static_transform_publisher',
arguments=['--frame-id', 'base_footprint', '--child-frame-id', 'imu_link', '--x', '-0.15']
),
launch_ros.actions.Node(
package="ros2_stm32_bridge",
executable="base_controller",
parameters=[
os.path.join(get_package_share_directory("ros2_stm32_bridge"), "params", MYCAR_MODEL + "_ekf.yaml"),],
)
])
In this launch file, the robot chassis driver is started, the ekf_node node from the robot_localization package is launched, and the coordinate transform between base_footprint and imu_link is also published.
(2) Write the configuration file
- In the
paramsdirectory of theros2_stm32_bridgepackage, create a new file namedstm32_4w_ekf.yamland enter the following content:
/mini_driver:
# /**: ros__parameters: base_frame: base_footprint
baud_rate: 115200 control_rate: 10 encoder_resolution: 44.0 kd: 130 ki: 0 kp: 100 maximum_encoding: 100.0 model_param_acw: 0.45 model_param_cw: 0.45 odom_frame: odom
port_name: /dev/mycar
publish_tf: false qos_overrides:
/parameter_events:
publisher: depth: 1000 durability: volatile
history: keep_last
reliability: reliable
/tf:
publisher: depth: 100 durability: volatile
history: keep_last
reliability: reliable
reduction_ratio: 90.0 use_sim_time: false wheel_diameter: 0.080
In this file, an important operation is to set the publish_tf parameter to false. When set to false, the chassis driver no longer publishes the coordinate transform between base_footprint and odom; instead, this is handled by ekf_node.
- In the
paramsdirectory of theros2_stm32_bridgepackage, create a new file namedekf.yamland enter the following content:
### ekf config file ###
ekf_filter_node:
ros__parameters:
# The frequency, in Hz, at which the filter will output a position estimate. Note that the filter will not begin
# computation until it receives at least one message from one of the inputs. It will then run continuously at the
# frequency specified here, regardless of whether it receives more measurements. Defaults to 30 if unspecified.
frequency: 30.0
# The period, in seconds, after which we consider a sensor to have timed out. In this event, we carry out a predict
# cycle on the EKF without correcting it. This parameter can be thought of as the minimum frequency with which the
# filter will generate new output. Defaults to 1 / frequency if not specified.
sensor_timeout: 0.1
# ekf_localization_node and ukf_localization_node both use a 3D omnidirectional motion model. If this parameter is
# set to true, no 3D information will be used in your state estimate. Use this if you are operating in a planar
# environment and want to ignore the effect of small variations in the ground plane that might otherwise be detected
# by, for example, an IMU. Defaults to false if unspecified.
two_d_mode: false
# Use this parameter to provide an offset to the transform generated by ekf_localization_node. This can be used for
# future dating the transform, which is required for interaction with some other packages. Defaults to 0.0 if
# unspecified.
transform_time_offset: 0.0
# Use this parameter to provide specify how long the tf listener should wait for a transform to become available.
# Defaults to 0.0 if unspecified.
transform_timeout: 0.0
# If you're having trouble, try setting this to true, and then echo the /diagnostics_agg topic to see if the node is
# unhappy with any settings or data.
print_diagnostics: true
# Debug settings. Not for the faint of heart. Outputs a ludicrous amount of information to the file specified by
# debug_out_file. I hope you like matrices! Please note that setting this to true will have strongly deleterious
# effects on the performance of the node. Defaults to false if unspecified.
debug: false
# Defaults to "robot_localization_debug.txt" if unspecified. Please specify the full path.
debug_out_file: /path/to/debug/file.txt
# Whether we'll allow old measurements to cause a re-publication of the updated state
permit_corrected_publication: false
# Whether to publish the acceleration state. Defaults to false if unspecified.
publish_acceleration: false
# Whether to broadcast the transformation over the /tf topic. Defaults to true if unspecified.
publish_tf: true
# REP-105 (http://www.ros.org/reps/rep-0105.html) specifies four principal coordinate frames: base_link, odom, map, and
# earth. base_link is the coordinate frame that is affixed to the robot. Both odom and map are world-fixed frames.
# The robot's position in the odom frame will drift over time, but is accurate in the short term and should be
# continuous. The odom frame is therefore the best frame for executing local motion plans. The map frame, like the odom
# frame, is a world-fixed coordinate frame, and while it contains the most globally accurate position estimate for your
# robot, it is subject to discrete jumps, e.g., due to the fusion of GPS data or a correction from a map-based
# localization node. The earth frame is used to relate multiple map frames by giving them a common reference frame.
# ekf_localization_node and ukf_localization_node are not concerned with the earth frame.
# Here is how to use the following settings:
# 1. Set the map_frame, odom_frame, and base_link frames to the appropriate frame names for your system.
# 1a. If your system does not have a map_frame, just remove it, and make sure "world_frame" is set to the value of
# odom_frame.
# 2. If you are fusing continuous position data such as wheel encoder odometry, visual odometry, or IMU data, set
# "world_frame" to your odom_frame value. This is the default behavior for robot_localization's state estimation nodes.
# 3. If you are fusing global absolute position data that is subject to discrete jumps (e.g., GPS or position updates
# from landmark observations) then:
# 3a. Set your "world_frame" to your map_frame value
# 3b. MAKE SURE something else is generating the odom->base_link transform. Note that this can even be another state
# estimation node from robot_localization! However, that instance should *not* fuse the global data.
map_frame: map # Defaults to "map" if unspecified
odom_frame: odom # Defaults to "odom" if unspecified
base_link_frame: base_footprint # Defaults to "base_link" if unspecified
world_frame: odom # Defaults to the value of odom_frame if unspecified
# The filter accepts an arbitrary number of inputs from each input message type (nav_msgs/Odometry,
# geometry_msgs/PoseWithCovarianceStamped, geometry_msgs/TwistWithCovarianceStamped,
# sensor_msgs/Imu). To add an input, simply append the next number in the sequence to its "base" name, e.g., odom0,
# odom1, twist0, twist1, imu0, imu1, imu2, etc. The value should be the topic name. These parameters obviously have no
# default values, and must be specified.
odom0: odom
# Each sensor reading updates some or all of the filter's state. These options give you greater control over which
# values from each measurement are fed to the filter. For example, if you have an odometry message as input, but only
# want to use its Z position value, then set the entire vector to false, except for the third entry. The order of the
# values is x, y, z, roll, pitch, yaw, vx, vy, vz, vroll, vpitch, vyaw, ax, ay, az. Note that not some message types
# do not provide some of the state variables estimated by the filter. For example, a TwistWithCovarianceStamped message
# has no pose information, so the first six values would be meaningless in that case. Each vector defaults to all false
# if unspecified, effectively making this parameter required for each sensor.
odom0_config: [false, false, false,
false, false, false,
true, false, false,
false, false, false,
false, false, false]
# If you have high-frequency data or are running with a low frequency parameter value, then you may want to increase
# the size of the subscription queue so that more measurements are fused.
odom0_queue_size: 2
# [ADVANCED] Large messages in ROS can exhibit strange behavior when they arrive at a high frequency. This is a result
# of Nagle's algorithm. This option tells the ROS subscriber to use the tcpNoDelay option, which disables Nagle's
# algorithm.
odom0_nodelay: false
# [ADVANCED] When measuring one pose variable with two sensors, a situation can arise in which both sensors under-
# report their covariances. This can lead to the filter rapidly jumping back and forth between each measurement as they
# arrive. In these cases, it often makes sense to (a) correct the measurement covariances, or (b) if velocity is also
# measured by one of the sensors, let one sensor measure pose, and the other velocity. However, doing (a) or (b) isn't
# always feasible, and so we expose the differential parameter. When differential mode is enabled, all absolute pose
# data is converted to velocity data by differentiating the absolute pose measurements. These velocities are then
# integrated as usual. NOTE: this only applies to sensors that provide pose measurements; setting differential to true
# for twist measurements has no effect.
odom0_differential: false
# [ADVANCED] When the node starts, if this parameter is true, then the first measurement is treated as a "zero point"
# for all future measurements. While you can achieve the same effect with the differential paremeter, the key
# difference is that the relative parameter doesn't cause the measurement to be converted to a velocity before
# integrating it. If you simply want your measurements to start at 0 for a given sensor, set this to true.
odom0_relative: false
# [ADVANCED] If your data is subject to outliers, use these threshold settings, expressed as Mahalanobis distances, to
# control how far away from the current vehicle state a sensor measurement is permitted to be. Each defaults to
# numeric_limits<double>::max() if unspecified. It is strongly recommended that these parameters be removed if not
# required. Data is specified at the level of pose and twist variables, rather than for each variable in isolation.
# For messages that have both pose and twist data, the parameter specifies to which part of the message we are applying
# the thresholds.
odom0_pose_rejection_threshold: 5.0
odom0_twist_rejection_threshold: 1.0
imu0: imu
imu0_config: [false, false, false,
false, false, true,
false, false, false,
false, false, true,
false, false, false]
imu0_nodelay: false
imu0_differential: false
imu0_relative: true
imu0_queue_size: 5
imu0_pose_rejection_threshold: 0.8 # Note the difference in parameter names
imu0_twist_rejection_threshold: 0.8 #
imu0_linear_acceleration_rejection_threshold: 0.8 #
# [ADVANCED] Some IMUs automatically remove acceleration due to gravity, and others don't. If yours doesn't, please set
# this to true, and *make sure* your data conforms to REP-103, specifically, that the data is in ENU frame.
imu0_remove_gravitational_acceleration: true
# [ADVANCED] The EKF and UKF models follow a standard predict/correct cycle. During prediction, if there is no
# acceleration reference, the velocity at time t+1 is simply predicted to be the same as the velocity at time t. During
# correction, this predicted value is fused with the measured value to produce the new velocity estimate. This can be
# problematic, as the final velocity will effectively be a weighted average of the old velocity and the new one. When
# this velocity is the integrated into a new pose, the result can be sluggish covergence. This effect is especially
# noticeable with LIDAR data during rotations. To get around it, users can try inflating the process_noise_covariance
# for the velocity variable in question, or decrease the variance of the variable in question in the measurement
# itself. In addition, users can also take advantage of the control command being issued to the robot at the time we
# make the prediction. If control is used, it will get converted into an acceleration term, which will be used during
# predicition. Note that if an acceleration measurement for the variable in question is available from one of the
# inputs, the control term will be ignored.
# Whether or not we use the control input during predicition. Defaults to false.
use_control: true
# Whether the input (assumed to be cmd_vel) is a geometry_msgs/Twist or geometry_msgs/TwistStamped message. Defaults to
# false.
stamped_control: false
# The last issued control command will be used in prediction for this period. Defaults to 0.2.
control_timeout: 0.2
# Which velocities are being controlled. Order is vx, vy, vz, vroll, vpitch, vyaw.
control_config: [true, false, false, false, false, true]
# Places limits on how large the acceleration term will be. Should match your robot's kinematics.
acceleration_limits: [1.3, 0.0, 0.0, 0.0, 0.0, 3.4]
# Acceleration and deceleration limits are not always the same for robots.
deceleration_limits: [1.3, 0.0, 0.0, 0.0, 0.0, 4.5]
# If your robot cannot instantaneously reach its acceleration limit, the permitted change can be controlled with these
# gains
acceleration_gains: [0.8, 0.0, 0.0, 0.0, 0.0, 0.9]
# If your robot cannot instantaneously reach its deceleration limit, the permitted change can be controlled with these
# gains
deceleration_gains: [1.0, 0.0, 0.0, 0.0, 0.0, 1.0]
# [ADVANCED] The process noise covariance matrix can be difficult to tune, and can vary for each application, so it is
# exposed as a configuration parameter. This matrix represents the noise we add to the total error after each
# prediction step. The better the omnidirectional motion model matches your system, the smaller these values can be.
# However, if users find that a given variable is slow to converge, one approach is to increase the
# process_noise_covariance diagonal value for the variable in question, which will cause the filter's predicted error
# to be larger, which will cause the filter to trust the incoming measurement more during correction. The values are
# ordered as x, y, z, roll, pitch, yaw, vx, vy, vz, vroll, vpitch, vyaw, ax, ay, az. Defaults to the matrix below if
# unspecified.
process_noise_covariance: [0.05, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.05, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.06, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.03, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.03, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.06, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.025, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.025, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.04, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.01, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.01, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.02, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.01, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.01, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.015]
# [ADVANCED] This represents the initial value for the state estimate error covariance matrix. Setting a diagonal
# value (variance) to a large value will result in rapid convergence for initial measurements of the variable in
# question. Users should take care not to use large values for variables that will not be measured directly. The values
# are ordered as x, y, z, roll, pitch, yaw, vx, vy, vz, vroll, vpitch, vyaw, ax, ay, az. Defaults to the matrix below
#if unspecified.
initial_estimate_covariance: [1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1e-9]
In this file, the key configuration is as follows:
#......
odom0_config: [false, false, false,
false, false, false,
true, false, false,
false, false, false,
false, false, false]
#......
imu0_config: [false, false, false,
false, false, true,
false, false, false,
false, false, true,
false, false, false]
The above configuration primarily fuses the x-direction velocity from the odometry message with the yaw angle from the IMU message to generate new odometry data.
(3) Compilation
In the terminal, navigate to the current workspace and compile the package:
colcon build --packages-select ros2_stm32_bridge
(4) Execution
In the terminal, navigate to the current workspace and execute the launch file:
. install/setup.bash
ros2 launch ros2_stm32_bridge driver_ekf.launch.py
Restart rviz2 and the keyboard control node. In rviz2, set the reference coordinate system to odom, add the TF plugin, then add two Odometry plugins and set their topics to odom and odometry/filtered respectively. Use the keyboard to drive the robot forward. The data displayed by the two odometry plugins should be roughly consistent with the robot's actual movement. After moving forward for a while, manually rotate the robot to produce a certain yaw angle. This action should also be accurately reflected in the fused odometry and coordinate transformation, while the wheel odometry remains unchanged because the motors did not rotate (as shown in the figure below).
