Autonomous Navigation

In this tutorial, we will show you how to perform SLAM (Simultaneous localization and mapping) and autonomous navigation on Leo Rover equipped with IMU and LiDAR sensors.

For this purpose, we prepared the leo_navigation package which makes use of many other packages available in ROS to provide autonomous capabilities. You won't have to write a single line of code, but some configuration may need to be tweaked to work best in your environment.

Prerequisites

To complete this guide, you will need to, first of all, have the IMU and LiDAR sensors integrated with the system:

Make sure you are operating on the newest image for the Raspberry Pi and you are up-to-date with the packages:

A recent version of firmware for the CORE2 board may also be required:

You will also need to have ROS installed on your computer:

Some previous experience with ROS is also recommended. If you are not familiar with it, be sure to check out our ROS Development guide:

Preparing the environment

There are a few things you will need to prepare on the rover and on your computer, in order to use the software for autonomous navigation.

On the Rover

You will need to build the leo_navigation package first. To do this, start by accessing a remote terminal session on the Rover by logging via ssh:

Create a catkin workspace:

mkdir -p ~/ros_ws/src
cd ~/ros_ws
catkin config --extend /opt/ros/melodic

Clone the leo_navigation package to the source space:

cd ~/ros_ws/src
git clone https://github.com/LeoRover/leo_navigation

Use rosdep to install dependencies:

cd ~/ros_ws
rosdep update
rosdep install --from-paths src -iy

Build the workspace:

catkin build

Source the devel space:

source ~/ros_ws/devel/setup.bash

You will have to source the devel space on every terminal session you want to use the package on. If you want to do it automatically upon logging into a session, add the above line to the ~/.bashrc file.

Checking the sensors

It is also advisable to check whether the sensors are correctly integrated into the system. Make sure the sensor-related data is published on the correct topics:

rostopic echo /imu/gyro
rostopic echo /imu/accel
rostopic echo /imu/mag
rostopic echo /scan

You will also need to have the transforms, from base_link to frame ids associated with the sensor messages, available in the tf tree.

rosrun tf tf_echo /base_link /imu
rosrun tf tf_echo /base_link /laser_frame

If you see errors, it probably means you omitted the part about extending the URDF model when integrating the sensor.

On your computer

In this guide, your computer will be used for visualizing the data processed on the rover and sending navigation goals. For this purpose, you will need to build the leo_description package, which contains the robot model and the leo_viz package which contains configuration files for RViz.

Start by installing some prerequisites:

sudo apt update
sudo apt install python-rosdep python-catkin-tools

If you have never run rosdep before on your computer, type:

sudo rosdep init

The rest of the steps should be already familiar:

mkdir -p ~/ros_ws/src && cd ~/ros_ws
catkin config --extend /opt/ros/<distribution>
cd src
git clone https://github.com/LeoRover/leo_description
git clone https://github.com/LeoRover/leo_viz.git
cd ..
rosdep update
rosdep install --from-paths src -iy
catkin build
source ~/ros_ws/devel/setup.bash

Replace <distribution> with the ROS distribution you have installed on your computer (either kinetic or melodic).

You will also need to properly configure your computer to communicate in ROS network, so be sure to correctly set these environment variables:

export ROS_IP=<your ip>
export ROS_MASTER_URI=http://10.0.0.1:11311

For more information, visit ROS Development -> Connecting other computer to ROS network

Launching the software

Now that everything is ready, we can proceed to launching some software. This whole section will describe in detail the functionalities provided in the leo_navigation package and how to use them.

Here's a brief summary of different parts of the navigation software we will run:

  • Odometry – It is necessary for any autonomous navigation system to be able to estimate its position. For this reason, we will use wheel encoders and IMU readings to track the change in position over time.

  • SLAM – Once we are able estimate the robot's position, we will use it, together with the LiDAR sensor measurements, to produce an occupancy map of the terrain. While doing so, we will also correct the estimated position based on the loop closure detection. After the whole terrain is mapped, we will save the map to a file and later use it to track robot's pose against it.

  • Navigation – With an occupancy map and an accurate enough pose estimation, we can finally try to perform some autonomous navigation. When a new navigation goal is set, a path planning algorithm will find a collision-free path to it. Another algorithm will then send velocity commands to the rover to try to follow that path while also being aware of the obstacles.

All of the functions are provided through launch files that are located under the launch/ directory. All parameters for the running nodes are loaded from YAML files located under the config/ directory.

Odometry

To estimate robot's position from wheel encoders and IMU, we will use the robot_localization package which contains two state estimation nodes: the ekf_localization_node which implements Extended Kalman Filter algorithm and the ukf_localization_node which implements Unscented Kalman Filter. We will use the first option as it is less computationally expensive but the latter is a more stable alternative.

Without further ado, let's try to run the localization software and see the result. Later, we will explain what is actually going on.

On the rover, type:

roslaunch leo_navigation odometry.launch

On your computer, type:

roslaunch leo_viz rviz.launch config:=odometry

The RViz instance should start and you should see the robot model. Now, try to steer the rover (either from the Web UI or by using a joystick). If everything works, you should see the model moving along the ground plane like in this video:

Here's a diagram that illustrates the nodes launched by the odometry.launch file and the connections between them:

The ovals represents the nodes. The rectangles with solid lines represent the topics. The blue text inside the rectangles indicate the type of messages published on the topic. The arrows between the ovals and rectangles represent the topic Subscribers and Publishers. tf rectangle represents the transforms between coordinate frames exchanged through the tf2 library.

The state estimation node requires that the input topics are stamped messages which contain covariance matrices. That's why the leo_navigation package provides the message_filter node which:

  • for each message on /wheel_odom topic, publishes a message on the /wheel_odom_with_covariance topic which contains the same data but with an added covariance matrix,

  • for each pair of /imu/accel and /imu/gyro messages, publishes a message on /imu/data_raw topic which combines the two messages and adds covariance matrices.

Apart from these topics, the ekf_localization_node listens for the base_link->imu tf transform to transform IMU readings to a common reference frame (base_link in this case).

The current state estimation is published on the /odometry/filtered topic and the pose is also broadcasted as the odom->base_link transform. That's why, having the Fixed Frame set to odomin RViz, you can see the model moving.

3D odometry

Notice that we didn't use much from the measurements returned by the IMU sensor, except for the angular velocity which was used to correct the wheel odometry.

Thanks to the imu_filter_madgwick package, we can get pretty good orientation estimate, which, combined with the wheel odometry, can be used to track the robot's position in 3D. The package uses the Madgwick's filter to fuse data from a gyroscope, an accelerometer and (optionally) a magnetometer.

To try the robot's state estimation in 3D, start the same launch file as in the previous example, but with the three_d argument set to true:

roslaunch leo_navigation odometry.launch three_d:=true

Now try to steer the rover on uneven terrain and observe the result in RViz.

The diagram for the odometry.launch file now changes a little:

We now have the /imu_filter_node which takes messages from the /imu/data_raw topic (gyroscope and accelerometer data), calculates the orientation of the sensor and publishes the data with added orientation quaternion on the /imu/data topic. The node can also optionally incorporate the /imu/mag topic (magnetometer data).

The /ekf_localization_node now uses different configuration which takes the data from /imu/data topic and uses the orientation quaternion in the state estimation.

Configuration

For the /message_filter node, there is the config/message_filter.yaml file in which you can change the values on the diagonals of the covariance matrices, which represent the variances (the uncertainty) of the measurements.

It might be better to leave the values as they are if you have no idea what they mean.

The configuration files for the /ekf_localization_node can be found in the config/ekf_localization_node/ directory. There are two files: ekf_2d.yaml and ekf_3d.yaml. The appropriate file is chosen depending on the value of 3d argument passed to the launch file.

A full description of parameters for the state estimation node can be found on the robot_localization documentation and in the ekf config template.

There is also the config/imu_filter_node.yaml file for the Madgwick's filter. The parameters are explained in the imu_filter_madgwick Wiki page. If you want to use the magnetometer data, just change use_mag parameter to true.

It is essential to have the magnetometer properly calibrated for the measured data to be useful. Magnetometers are prone to distortion, so even a slight change of the environment can make the current calibration invalid.

If you set the use_mag parameter and want use the heading information (compass) in /ekf_localization_node , just modify the imu0_config parameter so that the yaw value is fused into the state estimate:

config/ekf_localization_node/ekf_3d.yaml
imu0_config: [false, false, false,
true, true, true,
false, false, false,
true, true, true,
true, false, false]

SLAM

SLAM (Simultaneous Localization and Mapping) is a problem of constructing a map of the environment while simultaneously keeping track of the robot within it. It is hard because a map is needed for localization and a good pose estimate is needed for mapping, so this appears to be chicken-and-egg problem. There are, however, methods of solving it approximately with the use of probabilistic algorithms, such as the particle filter.

One such approach, called GMapping, can be used to perform SLAM with the laser range data (LiDAR scans) and a local odometry source. The algorithm has its ROS wrapper node in the gmapping package. To run it, just type:

roslaunch leo_navigation gmapping.launch

Mark the start position of the rover! It will be useful later.

On your computer, close any running RViz instance and open a new one with the slam configuration:

roslaunch leo_viz rviz.launch config:=slam

Now try to explore the terrain with your rover and see the map being created. You will probably experience the model doing discrete jumps on the map from time to time. This is due to gmapping correcting robot's position within the map. Here's a video demonstration:

The diagram for the gmapping.launch file is pretty straightforward:

The /slam_gmapping node takes as the input:

  • laser scan data from the LiDAR sensor (/scan topic),

  • position of the laser reference frame (base_link->laser_frame transform),

  • current position of the robot from the odometry (odom->base_link transform).

As the output, it returns:

  • an occupancy grid map of the terrain (/map and /map_metadata topics),

  • current odometry drift based on the estimated position of the robot within the map (map->odom transform).

The map, odom and base_link are coordinate frame names commonly used in mobile platforms. To learn about their semantic meaning, you can read REP105.

When the /slam_gmapping node tries to correct robot's position within the map, it does not broadcast the position as the map->base_link transform, because that would make 2 root coordinate frames (map and odom). Instead, it provides the map->odom transform which marks the difference between the odometry position and the actual position of the robot within the map (the odometry drift).

AMCL

If you already explored all the terrain you wanted to map, the /gmapping node can draw unnecessary resources for its mapping part. When map is no longer updated, it is more efficient to just use an algorithm that will only track robot's position against the map you have generated.

That's where the amcl package comes in handy. It implements the Monte Carlo localization algorithm for laser range data. To use it, first, save the current map generated by GMapping using the map_saver script:

rosrun map_server map_saver -f mymap

This will create 2 files: mymap.yaml containing map's metadata and mymap.pgm containing the actual occupancy grid in binary format.

Close the gmapping.launch instance by clicking Ctrl+C on the terminal session it is running on. Then start amcl.launch by typing:

roslaunch leo_navigation amcl.launch map_file:=<path_to_the_map_file>

Replace <path_to_the_map_file> with the absolute path to the mymap.yaml file. If the file is in your current working directory, you can instead type:

roslaunch leo_navigation amcl.launch map_file:=$(realpath mymap.yaml)

If your initial pose is far from the real one, AMCL might not be able to correctly localize the robot on the map. To fix this, you can use the 2D Pose Estimate tool in RViz to manually set the initial pose.

When your initial pose is close enough to the real one, try to steer the robot and notice how the position of the robot model in RViz is being corrected by AMCL.

Similar to the /slam_gmapping node, /amcl takes the laser range data (/scan topic) as input and provides the odometry drift information by broadcasting the map->odom transform. The difference is that instead of publishing to the /map topic, it receives the (previously generated) map from the /map_server node.

You can also notice 2 additional topics published by the /amcl node:

  • /amcl_pose − estimated pose of the robot within a map, with covariance,

  • /particlecloud − a set of pose estimates being maintained by the Monte Carlo localization algorithm.

You can visualize the data published on these topics by enabling the AMCL Pose and/or the AMCL Particle Cloud displays in RViz.

Configuration

The configuration for the /slam_gmapping node is loaded from the config/slam_gmapping.yaml file. For description of each parameter, visit the gmapping ROS wiki page.

Here are some parameters you can try to adjust:

  • map_update_interval − control how often to update the map.

  • linearUpdate, angularUpdate, temporalUpdate − control when to process new laser scans.

  • particles − increasing this number may improve the quality of the map at the cost of greater computational load.

  • minimumScore − you can experiment with this parameter if you are experiencing jumping pose estimates in large open spaces.

  • maxUrange, maxRange − try to set these parameters like this: maxUrange < maximum range of your LiDAR sensor <= maxRange.

  • delta − increasing this will lower the resolution of the map but may improve performance.

Similarly, for the /amcl node, there is the config/amcl.yaml file with the parameters that are described in the amcl ROS wiki page.

Here are some parameters worth to notice:

  • min_particles, max_particles − higher number of particles may increase the accuracy of the estimates at the cost of computational load.

  • update_min_d,update_min_a − control when to perform filter updates.

  • initial_pose_x, initial_pose_y, initial_pose_a − initial mean pose used to initialize the filter.

  • laser_max_beams − you can experiment with this parameter. Higher number may increase the accuracy at the cost of computational load.

Changing the parameters in a way that results in higher computational load may lead to violation of the real-time constraint. This in turn can make the algorithms perform even worse.

Now that you have the map of the terrain and your rover can localize itself within it (with either gmapping or amcl), you can try to make the robot drive autonomously to the designated position on the map.

To do this, we will utilize 2 packages:

  • move_base − provides a node which, given a navigation goal, will try to reach it by sending appropriate velocity commands. It incorporates many other components that have their own ROS API to achieve autonomous navigation capabilities.

  • twist_mux − provides a node which will let us multiplex several velocity commands, prioritizing one over the other. This will allow us to take control of the robot with a joystick or the Web UI if the navigation goes out of control.

To start these two nodes on you rover, just type:

roslaunch leo_navigation navigation.launch

You can start these components separately using twist_mux.launch and move_base.launch files.

On your computer, close any running RViz instance and start a new one with the navigation configuration:

roslaunch leo_viz rviz.launch config:=navigation

Now to send a navigation goal from RViz, select the 2D Nav Goal tool from the toolbar located at the top, then click somewhere on the map to set position or click and drag to set position and orientation.

If the goal is achievable, you should see the planned path in RViz and the navigation software should start sending commands to the robot. Make sure no other node is publishing velocity commands (/cmd_vel topic).

Here's a diagram for the navigation.launch file:

In the configuration we are using, the /move_base node's inputs include:

  • laser scan data (/scan topic),

  • position of the laser reference frame (base_link->laser_frame transform),

  • map generated by GMapping (/maptopic),

  • odometry from the /ekf_localization_node (/odometry/filtered topic),

  • current position estimation from the odometry and the odometry drift (odom->base_link and map->odom transforms).

The /move_base node can receive navigation goals in 2 ways:

  1. The /move_base_simple/goal topic − upon receiving a PoseStamped message on this topic, the node will cancel any running operation and try to reach the new destination. This is how the 2D Nav Goal tool in RViz sends the pose selected by the user.

  2. The /move_base action (see actionlib) − this is a preferred method to use when writing a script as it allows tracking the execution status and canceling the operation. Here you can find a simple example of sending a navigation goal through the action client.

When a navigation goal is being executed, the node publishes velocity command for the rover on the /nav_cmd topic.

The /twist_mux node chooses which topic with velocity commands to forward to the /cmd_vel topic. If no messages are sent on /ui_vel and /joy_vel topics, the /nav_cmd will be selected. To utilize the other topics when using a joystick or the Web UI, you need to make sure the velocity commands from these components are published on corresponding topics.

For the joysick, if you followed the example from the ROS Development tutorial, you can choose where to publish the commands by specifying the cmd_vel_topic argument to the launch file:

roslaunch leo_joy_example joy.launch cmd_vel_topic:=joy_vel

For the Web UI, you need to change cmd_vel to ui_vel in the /opt/leo_ui/js/leo.js file on your rover. You might also need to clear the cache data on your browser.

move_base brief explanation

Up to now, we treated the /move_base node as a black box to which we send navigation goals and get velocity commands in return. This is fine to start with, but if you want to fine-tune configuration for your environment, you might need to know about various components of move_base and how do they work with each other. This will help you to understand the meaning behind the parameters.

source: http://wiki.ros.org/move_base

move_base uses the costmap_2d package to generate layered 2D costmaps that are used by the planners to plan and execute a collision-free path to the navigation goal. Our configuration defines 3 layers:

  • Static map layer − takes the occupancy grid generated by GMapping (available through /map topic) and marks cells in the costmap as either "Unknown", "Freespace" or "Lethal" depending on the value of corresponding pixels on the map.

  • Obstacle layer − subscribes to the data from observation sources (in our case, laser scans) and uses it to mark the cells as "Lethal" if it sees an obstacle and "Freespace" if the cell lies between the robot and the observed obstacle (possibly clearing cells previously marked as "Lethal"). The clearing operation works by performing raytracing through the grid from the origin of the sensor outwards.

  • Inflation layer − propagates cost values from occupied cells (marked as "Lethal" by other layers) that decrease with distance. Up to the distance denoted by the inscribed radius of the robot, the cells are marked as "Inscribed", which means that the robot is definitely in collision if it's center lies in this cell. Further, up to the inflation radius, the costs decrease exponentially.

There are 2 costmaps being used to store information about the environment:

  • global_costmap − used for creating long-term plans over the entire environment. Combines all 3 defined layers and runs in the map coordinate frame.

  • local_costmap − used for local planning and obstacle avoidance. Utilizes only the Obstacle and Inflation layers. Uses the "Rolling Window", meaning that it will remain centered around the robot as it moves to represent only the local surroundings. Runs in the odom coordinate frame to guarantee that the pose of the robot in the map will be continuous.

move_base uses 2 planners that cooperate with each other to accomplish navigation goals:

  • Global Planner − operates on the global_costmap. Upon receiving a navigation goal, finds a safe path from the current position of the robot to the goal position. Our configuration uses the implementation from the global_planner package which utilizes Dijkstra's algorithm.

  • Local Planner − operates on the local_costmap. Given a path to follow and the costmap, it produces velocity commands to send to the robot. We use the implementation from the base_local_planner package which uses the Dynamic Window Approach.

The last bit worth mentioning are the recovery behaviors. When the robot perceives itself as stuck, move_base will try to perform some actions, in order to clear out space. The default recovery behaviors will perform the following actions:

  1. Conservative reset - clear out obstacles in the costmap outside of a user-specified region.

  2. Clearing rotation - perform an in-place rotation.

  3. Aggressive reset - clear out all obstacles outside of a rectangular region in which the robot can rotate in-place.

  4. Clearing rotation - perform an in-place rotation.

If the robot is still stuck after performing these actions, move_base will abort the navigation goal.

Configuration

As the /move_base node uses components that have their own ROS APIs, the configuration has been split into multiple files to make it more modular. There is a total of 6 files under the config/move_base/ directory. Here is a short description of each file and some parameters you can try to adjust:

  • move_base.yaml – parameters read directly by the /move_base node.

    • controller_frequency – increasing this may improve the response time, but be careful not to violate the real-time constraint.

    • clearing_rotation_allowed – you may want to disable clearing rotations when debugging the local planner.

    • conservative_reset_dist – smaller values will clean more space on Conservative reset.

  • costmaps/costmap_common.yaml – parameters common to the global and local costmaps. These include the footprint polygon and the costmap layers configuration.

    • footprint – if you have mounted additional equipment to your rover that expands the outer dimensions of the robot, you should modify this parameter.

    • obstacle_layer/obstacle_range – set it to the usable range of your LiDAR sensor.

    • obstacle_layer/raytrace_range – set it slightly larger than the usable range of your sensor but smaller than the maximum range.

    • obstacle_layer/scan/expected_update_rate – set it slightly larger than the update rate of your sensor (time between scan messages).

    • inflation_layer/inflation_radius – larger values will make the planners prefer paths more distant from the obstacles but can make it difficult to travel through narrow corridors.

    • inflation_layer/cost_scaling_factor – let's you control how the cost values decrease with distance. Larger values will make the cost decay function more steeper.

  • costmaps/global_costmap.yaml – parameters loaded by the global costmap.

    • update_frequency – increasing it improves the response time to obstacles but affects performance.

    • publish_frequency – higher values will make RViz visualization more responsive at the cost of increased network throughput.

  • costmaps/local_costmap.yaml – parameters loaded by the local costmap.

    • update_frequency – as above.

    • publish_frequency – as above.

    • width, height – let's you control how much of the local surroundings to remember.

    • resolution – smaller values will make the map more precise at the cost of increased size.

  • planners/global_planner.yaml – parameters loaded by the Global Planner.

    • default_tolerance – you can set it if there is a possibility of an obstacle in the goal position.

    • use_dijkstra – setting this to false will make the Global Planner use A* algorithm for path searching which is more efficient than Dijktra but may not find the optimal path. Useful in very large maps.

  • planners/local_planner.yaml – parameters loaded by the Local Planner.

    • xy_goal_tolerance – a too small value can make the robot oscillate in goal position.

    • yaw_goal_tolerance – you can increase it if you don't care that much about the resulting orientation of the rover.

    • pdist_scale, gdist_scale, occdist_scale – these parameters may require some trial and error to fine-tune them to your environment.

Most of these parameters are available through dynamic_reconfigure so you can modify them at runtime before changing the configuration files, e.g. by running the rqt_reconfigure tool on your computer:

rosrun rqt_reconfigure rqt_reconfigure