robotics-testing

Initialize rclpy once per module with a fixture, create the node in a per-test fixture, then invoke callbacks directly with synthetic messages and spin briefly to verify published output. Avoid time.sleep; use event-driven waits with timeouts instead.

Use launch_testing to start the nodes under a LaunchDescription with use_sim_time enabled, then publish test inputs and subscribe to expected outputs from a test node. Assert that messages arrive within a bounded spin loop rather than fixed sleeps.

Property-based testing with Hypothesis generates random inputs and checks mathematical invariants, such as FK/IK roundtrips, unit quaternion preservation, and monotonic trajectory interpolation. It catches edge cases that hand-picked test vectors miss.

Yes. Mock cameras and joint state publishers provide deterministic synthetic inputs, and simulation harnesses with fixed random seeds replace physical robots. A GitHub Actions workflow can run colcon test, launch_test, and simulation tests in ROS containers.

Flakiness usually comes from time.sleep waits, unfixed random seeds, and dependence on message timing. Replace sleeps with threading.Event waits with timeouts, seed all randomness, and use deterministic mock data instead of live sensors.

Golden-file testing saves a known-good planner output as a reference file and compares future runs against it with a tolerance. Any deviation signals a regression in planner behavior, while the first run bootstraps the reference file.