scrubber_sim

ROS 2 Jazzy + Gazebo Harmonic (gz-sim) simulation for a large three-wheel Ackermann floor-scrubbing robot (autonomous cleaning vehicle). Full perception–navigation stack: SLAM mapping + Nav2 path planning + a custom behavior tree tuned for Ackermann kinematics.

大型三轮阿卡曼洗地车的自动驾驶仿真。ROS 2 Jazzy + Gazebo Harmonic，SLAM 建图 + Nav2 导航 + 阿卡曼定制行为树。

Robot

• Body: 1.4 × 1.0 × 1.3 m, mass ~200 kg, three-wheel Ackermann (single front wheel: drive + steer)
• Wheel radius 0.15 m, wheelbase 1.0 m, track 0.9 m, min turning radius 1.5 m
• Sensors: 360° 2D LiDAR (gpu_lidar), IMU
• Drive: front-wheel steer + rear-wheel drive (no reversing, no in-place rotation)

Package layout

scrubber_sim/
├── urdf/scrubber.urdf.xacro     # robot model + gz-sim plugins (AckermannSteering / lidar / imu)
├── worlds/simple_world.sdf      # walled room + box obstacles
├── config/
│   ├── nav2.yaml                # Nav2 params (SmacPlannerHybrid + RegulatedPurePursuit)
│   ├── slam_toolbox.yaml        # online async mapping
│   └── ackermann_bt.xml         # custom BT: Spin/BackUp removed (Ackermann can't)
├── launch/
│   ├── bringup.launch.py        # sim + robot + ros_gz bridge
│   ├── slam.launch.py           # bringup + slam_toolbox
│   └── nav2.launch.py           # bringup + slam + Nav2 stack
└── scrubber_sim/
    └── wheel_odometry.py        # custom bicycle-model odometry (see Notes)

Run

# 1. build
cd ros2_ws && colcon build --symlink-install && source install/setup.bash

# 2. bring up sim + robot
ros2 launch scrubber_sim bringup.launch.py

# 3. mapping
ros2 launch scrubber_sim slam.launch.py
#    drive around with: ros2 run teleop_twist_keyboard teleop_twist_keyboard

# 4. full navigation
ros2 launch scrubber_sim nav2.launch.py
#    send a goal in RViz, or:
ros2 action send_goal /navigate_to_pose nav2_msgs/action/NavigateToPose \
  "{pose: {header: {frame_id: 'map'}, pose: {position: {x: 3.0, y: 0.0}, orientation: {w: 1.0}}}}"

Notes / design decisions

These are hard-won lessons from getting the stack to actually run on ROS 2 Jazzy + gz-sim:

• Ackermann BT: the default Nav2 BT references the Spin recovery, which an Ackermann vehicle cannot perform. config/ackermann_bt.xml removes Spin/BackUp, keeping only Wait + DriveOnHeading. The behavior_server plugin list must match.
• Planner: SmacPlannerHybrid with motion_model_for_search: DUBIN (forward-only) + RegulatedPurePursuitController with use_rotate_to_heading: false, allow_reversing: false.
• Plugin names: Jazzy uses :: (e.g. nav2_smac_planner::SmacPlannerHybrid), not the / form from older releases.
• Lifecycle managers: slam_toolbox and the Nav2 nodes are managed by separate nav2_lifecycle_manager instances — mixing them causes a bond timeout.
• Custom odometry (wheel_odometry.py): the gz AckermannSteering system's built-in odometry is unreliable for this single-front-wheel three-wheel layout (it overshoots/drifts even with correct geometry params). We instead integrate a bicycle model from /joint_states (rear-wheel velocity → linear speed, steering-joint angle → yaw rate). This matches the real vehicle's STM32 wheel-odometry, so sim and hardware share the same model.
• gz launch must pass -r (run), otherwise the sim starts paused and no sensor data flows.

Tools

Standalone Python helpers under tools/scrubber-sim/ (no ROS 2 runtime needed, run + test locally):

• map_to_polygons/ — SLAM map.pgm → obstacle polygon list (F2C inner voids) via cv2.findContours + approxPolyDP + pixel→world transform. Feeds the complete-coverage goal so the planner routes around mapped static obstacles.
• coverage_meter/ — actual coverage measurement by footprint sweep (not path-length estimate). Stamps the cleaning footprint along the recorded trajectory onto a grid, intersects with the cleanable area (outer − voids), and reports coverage %, swept-obstacle area (collision alarm) and overspray area, with a visualization.

cd tools/scrubber-sim
python3 -m pytest tests/ -q     # 28 tests

Status

Mapping closed-loop + single-goal navigation work end to end. Complete-coverage (boustrophedon) path planning is integrated via F2C + opennav_coverage; current work is the two-layer obstacle-avoidance architecture (static voids upstream + collision-detection/recovery downstream).

License

MIT