Simulation test: drive DiffBot and Remo in Gazebo in CI

[fjp]

The first tests that check how the robot behaves, not only that everything builds. Both CI workflows run them on every pull request: CI with industrial_ci, and Dev container.

What the test does

A rostest per robot, diffbot_gazebo/test/sim_drive_diffbot.test and sim_drive_remo.test, starts the robot in Gazebo without a GUI, on the empty world, with the diff drive controller. It needs no display. Two tests then run:

• publishtest, a test node that comes with rostest: /diffbot/joint_states and /diffbot/mobile_base_controller/odom are published.

• test_drive.py: drives 0.2 m/s for 3 s, then turns at 0.5 rad/s for 3 s, and checks each move twice:

  • odometry: what the diff drive controller computes from the wheel joints and its configuration;
  • true pose: where Gazebo says the robot is (/gazebo/model_states).

  Both must be in range (forward: 0.4–0.75 m, straight ahead; turn: 0.5–2.0 rad to the left), and they must agree: within 0.05 m forward and 0.25 rad when turning.

  The odometry and Gazebo report poses in different frames (odometry starts at 0, 0, 0, Gazebo spawns the robot at x = -2, y = -0.5). So the test compares motions, not poses. relative_motion() describes the motion between two poses as in the odometry motion model of Probabilistic Robotics (Thrun, Burgard, Fox, section 5.4): a rotation rot1 towards the end position, a translation trans, and a rotation rot2 to the end heading. These don't depend on the frame. Like ROS amcl's odometry model, rot1 is 0 for translations under 1 cm (turning on the spot). The heading comes from tf's euler_from_quaternion. Each method does one step: drive() sends the velocity command, stop() stops the robot, measure_motion() returns the motion from the odometry and from the true pose, and the tests call it with named arguments. The two tests, test_forward and test_turn, each check only their own motion, so their order doesn't matter. The script has type hints and passes flake8 with ROS 2's ament_flake8 settings (PEP 8, line length 99) and mypy.

The true pose is what makes the test useful. The odometry repeats some configuration mistakes, so on its own it would still look right.

What it catches

Each case was checked by breaking diffbot_control.yaml in a copy and running the tests:

Broken	Result
Wrong left wheel joint name	Controller doesn't load, no odometry: publishtest fails, test_drive errors
Left and right wheels swapped	Robot turns right (true pose −1.26 rad) while the odometry reports +1.36 rad: fails (2 failures over both robots)
Wheel radius 20 % too big	Odometry 0.58 m forward, true pose 0.49 m: fails (3 failures over both robots)

Changes

• diffbot_gazebo/test/: sim_drive.launch (Gazebo headless with diffbot_empty.world, controllers without rqt_robot_steering), the two .test files and test_drive.py. They're registered with add_rostest in CMakeLists.txt, with test_depends in package.xml. remo_description can't be one: it isn't a released ROS package, and the dev container's Dockerfile resolves every package.xml before setup.sh clones it. Both workflows clone it with diffbot_dev.repos.
• Install rules for diffbot_description, diffbot_control and diffbot_gazebo (urdf, config, meshes, launch, rviz, worlds, models). None of the packages installed its files. industrial_ci builds into an install space by default (catkin config --install), where xacro couldn't find diffbot.urdf.xacro. remo_description gets the same rule in Install the model files, so the package works from an install space remo_description#1. The other packages, including diffbot_base's nodes, still have no install rules: Install rules for all packages, so they work from an install space #113.
• One Gazebo port per test file (GAZEBO_MASTER_URI, 11351 and 11352). catkin runs the two test files in parallel, each with its own ROS master. On Gazebo's default port 11345, the second Gazebo server couldn't start, and Remo's test timed out. The two ports also don't collide with a simulation that's open in the dev container.
• diffbot_control.launch: a new argument rqt_robot_steering (default true, so nothing changes), so the test can leave out the GUI.
• diffbot_view.launch: the default world pointed to worlds/db_world.launch, which doesn't exist; it's worlds/db_world.world. diffbot.launch always overrode it, so this only mattered when the file was launched on its own.
• CI: industrial_ci builds remo_description from source with UPSTREAM_WORKSPACE: diffbot_dev.repos, like the dev container, because it isn't a released ROS package.

Turning on the spot reaches about 1.3 rad for DiffBot and 1.0–1.2 rad for Remo, instead of the commanded 1.5 rad. Gazebo's wheel velocity controller (diffbot_control/config/pid.yaml) has only a P gain, which leaves Remo's wheels at about 80 % of the commanded speed when turning, and the wheels slip. The test therefore only checks that the robot turns left by a clear amount, and that odometry and true pose agree. #116 tunes the gains (measurements there) and then tightens the turn check.

Testing

In the dev container image, with the commands of the Dev container workflow (catkin build --catkin-make-args run_tests, catkin_test_results build):

• Six clean runs: 10 tests, 0 errors, 0 failures, about 50 s each for both robots in parallel.

• The three broken cases above fail as described.

• From an install space, as industrial_ci builds (catkin config --install, with the install rule added to remo_description): 10 tests, 0 errors, 0 failures.

industrial_ci's Remo test needs ros-mobile-robots/remo_description#1 merged first, because industrial_ci clones remo_description's main branch.

Closes #110