Outdoor UAV–UGV Collaborative Continuous Demining System

中文说明

A ROS Noetic / Gazebo Classic workspace for a simulated UAV–UGV collaborative mine-removal mission. A PX4 UAV surveys a mine field with a downward RGB-D camera and a YOLO11 segmentation model; confirmed locations are transformed into the UGV's egocentric frame. A Husky–UR5 UGV uses LiDAR-inertial odometry, an elevation-aware planner, and a MoveIt pipeline to execute the configured visit–grasp–return–place cycle.

Scope. This is a simulation and course/research project, not a real-world explosive-ordnance-disposal system. Run it only in the supplied Gazebo/PX4 environment.

Submission material in this repository:

• Source code and configuration: the repository itself.
• Presentation: 空地协同连续排雷系统.pptx.
• 3 min 11 s UAV mapping and terrain-visualization excerpt (1920×1080, H.264/MP4): 空地协同排雷演示视频.mp4. It is supporting footage, not a substitute for the full success checks in §10.

1. 🎓Project basic information

Group No.	3
Assignment name	空地协同户外操作系统
Project name	Outdoor UAV–UGV Collaborative Continuous Demining System / 户外空地协同连续排雷系统
Team members	武天豪、赵汝堃、杜军、吴淑林

2. Function description

Problem. Locate simulated landmines from the air and have a ground robot visit, grasp, return, and place each confirmed target without a pre-built shared global map. The UGV handoff and ground-navigation chain is egocentric and does not require GPS; PX4 SITL may still use its simulated GNSS sensors internally.

Main features:

• UAV survey with a down-facing RGB-D camera and a YOLO11s-seg landmine detector.
• Multi-frame spatial confirmation to turn noisy per-frame detections into a stable mine map.
• Egocentric handoff of confirmed targets into the UGV's local odom frame (the UGV ground mission needs neither a shared global map frame nor GPS).
• UGV autonomy: LiDAR-inertial odometry, elevation-aware path planning, ordered target visitation.
• MoveIt Task Constructor grasp pipeline for the simulated mine prop. The submitted configuration uses the analytic top-down candidate server and a Gazebo fixed joint to hold the selected landmine_* rigid body during transport; the visible jaw motion is not claimed as a friction grasp. The GPD adapter is retained as non-acceptance research code.

Method (brief — see docs/algorithm_research_references.md for the underlying algorithm survey, not reproduced here): PX4 SITL + EGO-Planner for UAV flight, Ultralytics YOLO11s-seg for the submitted CPU detection path, DLIO for UGV odometry, a CMU-derived local planner with an ANYbotics elevation_mapping cost layer for UGV navigation, and MoveIt Task Constructor with analytic grasp candidates. A TensorRT adapter is retained but is not part of the tested submission configuration.

Inputs: the supplied outdoor_city Gazebo world, the five-mine field spawned by src/uav_truth_tracker/launch/spawn_outdoor_mine_field.launch, the shipped YOLO11s-seg weights, and runtime environment variables (§8).

Outputs: a confirmed mine map (mine_detection_output/mine_map.yaml), UGV tour/grasp status on ROS topics, and optional RViz visualization.

Applicable scenario: a single-workstation Gazebo Classic simulation; not applicable to real hardware without significant additional work (see Scope above).

Mission pipeline

PX4 UAV + down-facing RGB-D camera
          │
          ▼
YOLO11s-seg landmine localization
          │  /mine_detection/raw
          ▼
multi-frame association and confirmation
          │  /mine_detection/map (MineMap)
          ▼
MineMap → WorldTarget bridge
          │  /detected_targets
          ▼
UGV target tour + elevation-aware planning
          │  /ugv/goal
          ▼
MoveIt grasp–return–place cycle

The air-ground transform is established once at startup: DLIO owns odom → base_link, and an anchor latches odom → uav0/map_local. The UGV target handoff and ground mission therefore require neither a shared global map frame nor GPS; this statement does not disable PX4's simulated GNSS inside SITL.

3. Directory layout

Path	Contents	Ownership
airground_takeoff.sh	Main integrated demo launcher and runtime defaults.	User-run; tune through environment variables rather than editing it.
one_key_takeoff.sh	Legacy UAV-only flight/mapping launcher retained for historical debugging.	Do not use it for the complete course mission; use airground_takeoff.sh.
record_uav_map.sh	UAV-only point-cloud recording helper retained for mapping/debug sessions.	Optional historical/debug tool; it does not start the complete course mission.
stop_airground.sh	Broadly stops the current user's ROS/PX4/Gazebo simulation stack; see the warning in §11.	User-run.
setup_uav.sh	Installs external dependencies and builds the workspace.	User-run; PX4_DIR may be overridden.
uav_deps.repos	Legacy EGO-Planner compatibility import manifest.	Compatibility only; the supported setup_uav.sh path fetches and checks the exact EGO commit listed in §5.
requirements-runtime.txt	Pinned Python packages for the verified CPU detector path.	Dependency file; do not edit for the default run.
src/uav_truth_tracker/	UAV ROS nodes, survey, mine localization/fusion, PX4 launch files, and custom messages.	Source under scripts/; runtime parameters under launch/ and config/.
src/mobile_manipulator/	Husky–UR5 simulation, DLIO/planning integration, elevation filters, target tour, and bridge node.	Source under scripts//src/; runtime parameters under launch/ and config/; display presets under rviz/.
src/grasp_mtc/	MoveIt Task Constructor grasp pipeline.	Source under scripts/; task launch parameters under launch/.
src/gpd_ros/	Optional GPD messages and detector wrapper.	Messages still build without libgpd; the default mission does not require GPD.
models/, worlds/	Gazebo UAV models and shared simulation worlds.	Supplied input assets.
mine_seg_v2_delivery/	Delivered YOLO segmentation weight and model manifest.	Supplied input model; keep weights/best.pt in place.
docs/	Design, research, and testing notes.	Reference material.
patches/	Project patches applied to the EGO-Planner checkout.	Applied automatically by setup_uav.sh.
build/, devel/	Catkin build products.	Generated; safe to regenerate.
mine_detection_output/	Confirmed mine-map output for the current run.	Generated and overwritten at the next run.

4. Runtime environment

Component	Recorded environment or reproducible target
OS	Ubuntu 20.04.6 LTS
ROS	ROS Noetic
Simulator	Gazebo Classic 11.15.1
Flight stack in the latest local runs	PX4 commit bda25bfcc1a817f4ba559497c8ad6962f114cfd7 (v1.17.0-alpha1-1668-gbda25bfcc1-dirty when inspected on 6 August 2026) + MAVROS + EGO-Planner commit bfda51284c8c1b476043255a8145ef925a3778a5
Clean reproduction target installed by this repository	PX4 tag v1.14.3, commit 1dacb4cdef2d7145754fc788fa8dc482eed74b40, built as px4_sitl_default gazebo-classic
Main ROS packages	MAVROS 1.20.1, Navigation 1.17.3, robot_localization 2.7.7, MoveIt 1.1.16, catkin-tools 0.9.4
Build toolchain	GCC 9.4.0, CMake 3.16.3, C++17, catkin-tools 0.9.4 (catkin build), Python 3.8.10
Verified CPU perception environment	Ultralytics 8.4.60, PyTorch 2.4.1+cpu, torchvision 0.19.1+cpu, OpenCV 4.13.0, NumPy 1.24.4, SciPy 1.10.1, PyYAML 5.3.1
Optional GPU acceleration	NVIDIA driver + CUDA + TensorRT and an exported ONNX model. This is used only when explicitly selecting UAV_DETECT_BACKEND=tensorrt; the default CPU/Ultralytics path needs none of it.
Desktop session	A desktop OpenGL/Gazebo-capable session and gnome-terminal (the launcher opens one tab per subsystem — see §7)
Actual test hardware	Lenovo ThinkBook 15 G4 IAP, Intel Core i5-1240P (16 logical CPUs), 16 GB RAM, Intel integrated graphics, no NVIDIA GPU. No physical UAV/UGV is required.
Storage requirement	At least 10 GB free for PX4 sources/build products, ROS build products, logs, and generated maps.

The local PX4 row records the exact checkout on the development workstation, including its local modifications; it is not portable. For a fresh evaluator machine, setup_uav.sh deliberately installs and checks the clean v1.14.3 reproduction target instead of pretending to reproduce a dirty checkout. There are no other OS/version combinations documented or tested; if you are not on Ubuntu 20.04 + ROS Noetic, expect to adapt package names yourself.

5. Dependency installation

Prerequisite: Ubuntu 20.04 with the ROS Noetic apt repository configured and ros-noetic-desktop-full installed (the course image used for testing already provides this). On another clean Ubuntu 20.04 host, install ROS Noetic first, then verify that /opt/ros/noetic/setup.bash exists before continuing. The project setup script intentionally stops if no ROS environment has been sourced.

Clone the repository into a catkin workspace and run the setup from the repository root:

git clone --recurse-submodules https://github.com/JunDu-cyber/Air-Ground-Collaborative-Manupulation.git learning_ws
cd learning_ws
source /opt/ros/noetic/setup.bash
bash setup_uav.sh

setup_uav.sh initializes and verifies the pinned submodules and ANYbotics helper repositories, installs the fixed Python/ROS dependencies, prepares the egocentric navigation sources, clones and patches EGO-Planner, installs or completes PX4 SITL at ${PX4_DIR:-$HOME/PX4-Autopilot}, and builds the complete workspace. It requires sudo and can take considerable time. After the PX4 revision check, it records the resolved path in ${XDG_CONFIG_HOME:-$HOME/.config}/airground/env.sh and appends one managed Gazebo setup block to ~/.bashrc; an explicitly exported PX4_DIR always takes priority.

External source	Verified revision	Installed location and handling
ZJU FAST-Lab EGO-Planner	bfda51284c8c1b476043255a8145ef925a3778a5	src/ego-planner/; fetched by commit and patched automatically. No archive extraction or manual environment variable is needed.
PX4-Autopilot	tag v1.14.3, commit 1dacb4cdef2d7145754fc788fa8dc482eed74b40	${PX4_DIR:-$HOME/PX4-Autopilot}; cloned with submodules and built as px4_sitl_default gazebo-classic. If that directory already contains a different PX4 revision, setup stops without modifying it and tells you to choose a new PX4_DIR.
direct_lidar_inertial_odometry	fc8d183f18cdcfb9bb4fc754c6d373cedc4cbd04	Git submodule at src/direct_lidar_inertial_odometry/.
autonomous_exploration_development_environment	bf0cba71365271ebff09831a05afd78578150300	Git submodule at src/autonomous_exploration_development_environment/.
robot_body_filter	b6635e9c40d0524d70e4e0059a5c6c6bb382d6f4	Git submodule at src/robot_body_filter/.
FAST_LIO / livox_ros_driver (optional baseline)	7cc4175de6f8ba2edf34bab02a42195b141027e9 / 3d240d5666129e1a3052e78ee8487a04b08fdda3	Git submodules at src/FAST_LIO/ and src/livox_ros_driver/; present for comparison but excluded from the default build.
ANYbotics message_logger, kindr, kindr_ros	commits recorded in src/elevation_mapping.repos	Imported into src/ by setup_uav.sh. The project-modified elevation_mapping source itself is already vendored at src/elevation_mapping/; see its UPSTREAM.md.
YOLO11s-seg model	manifest in mine_seg_v2_delivery/DEPLOYMENT.txt	Already included as mine_seg_v2_delivery/weights/best.pt; do not move or extract it.

The setup script installs the pinned Python perception packages from requirements-runtime.txt. To repair only that environment later, run:

python3 -m pip install --user -r requirements-runtime.txt

Updating or repairing an existing clone

After pulling a new revision, rerun the supported setup path. It rechecks source revisions, restores dependencies and patches, runs rosdep, verifies/installs PX4, and rebuilds the workspace:

source /opt/ros/noetic/setup.bash
bash setup_uav.sh
source devel/setup.bash

ROS package dependencies declared in package.xml files are resolved by rosdep inside setup_uav.sh. GPD and TensorRT are retained comparison adapters but are not reproducible submission/acceptance configurations; leave GRASP_SOURCE=analytic and UAV_DETECT_BACKEND=cpu for assessment.

6. Pre-run configuration

• Environment variables (common mission switches and advanced overrides are listed in §8). The tested defaults use CPU detection, manual UAV goals, UGV navigation/grasping, elevation costs, and analytic grasp candidates. Change one subsystem at a time when diagnosing a run.
• Model file placement. The submission model must be present at mine_seg_v2_delivery/weights/best.pt (already shipped in this repo). Its manifest is mine_seg_v2_delivery/DEPLOYMENT.txt — task, input size (960), recommended confidence (0.7337337337 per the manifest; the launch file's own default is 0.65, see §8), and validation metrics. Use best.pt, not last.pt.
• PX4_DIR. If PX4 must be installed somewhere other than ~/PX4-Autopilot, export PX4_DIR for setup_uav.sh. After its revision check, setup saves that directory in ${AIRGROUND_ENV_FILE:-${XDG_CONFIG_HOME:-$HOME/.config}/airground/env.sh}; airground_takeoff.sh reads it automatically. An explicitly exported PX4_DIR still overrides the saved value.
• The default integrated entry point does not require editing personal absolute paths; its active paths are derived from the repository root or the variables in §8. Some legacy standalone launch files retain their own defaults and are not part of this procedure.

7. 🚀Complete run procedure

Start the integrated simulation from the workspace root:

source /opt/ros/noetic/setup.bash
source devel/setup.bash

bash airground_takeoff.sh

The default is the verified CPU detector path. UAV_DETECT_BACKEND=tensorrt is retained for separately configured research machines but was not tested on the acceptance hardware.

Step 1. From the workspace root, source ROS then the workspace overlay (in that order), as shown above. Step 2. Run airground_takeoff.sh. This is the single entry point for the default demo; no second launch command is required. Step 3. The script opens gnome-terminal tabs in this fixed launch order: 1_Gazebo → 1b_Mine_Field → 2_PX4_Spawn_UAV → 3_UGV → 3b_UGV_LIO → (3c_UGV_NAV, 7a_MoveGroup, 7b_Grasp if enabled) → 4_MAVROS → 5_AirGround_ROS → 6_Takeoff → 8_UAV_Detect. Built-in delays sequence the starts; the readiness checks in Step 4, not the elapsed delay alone, determine when it is safe to operate. Step 4. Wait roughly 30–60 s for the tabs to come up. startup sequence done means that launch commands were dispatched, not that every node is ready. In a new terminal, source both setup files and verify all of the following before sending a goal (stop each continuous hz/tf2_echo command with Ctrl-C after it produces valid data):

rostopic echo -n1 /mavros/state              # connected: True
rostopic echo -n1 /state_estimation           # one UGV odometry message
rosrun tf2_ros tf2_echo odom uav0/map_local   # a stable transform
rostopic hz /terrain_map                      # non-zero rate
rostopic hz /mine_camera/rgb/image_raw        # non-zero rate
missing=0
for service in /ugv/align_to_mine /grasp/execute /grasp/place /ugv/start_tour; do
  rosservice list | grep -qx "$service" || { echo "missing: $service"; missing=1; }
done
test "$missing" -eq 0

Then proceed to the interactive steps in §10. Step 5. Inspect results per §9 (Output) and §10 (Success criteria).

The launcher starts Gazebo paused, spawns the five-mine field, brings up PX4, the UGV, DLIO, planning, the air-ground TF/elevation layer, MAVROS, the UAV takeoff bridge, and mine detection.

8. Input specification

Static and interactive inputs used by the submitted mission:

Input	Format / interface	Repository path or source	Role in the default run
outdoor_city.world	SDFormat/XML Gazebo world	src/mobile_manipulator/worlds/outdoor_city.world	Default simulation scene selected by airground_takeoff.sh.
spawn_outdoor_mine_field.launch	ROS launch/XML	src/uav_truth_tracker/launch/spawn_outdoor_mine_field.launch	Spawns the supplied five-mine test field into the default world.
best.pt	PyTorch checkpoint	mine_seg_v2_delivery/weights/best.pt	YOLO11s-seg weights used by the default CPU detector.
RViz 2D Nav Goal	Manual geometry_msgs/PoseStamped interaction	RViz tool → /move_base_simple/goal	Supplies UAV survey waypoints when UAV_SURVEY=false (the default).

All required mission inputs are repository assets, simulated sensor streams, or RViz interaction. No physical UAV, UGV, LiDAR, camera, joystick, or other external peripheral is required.

Common mission switches consumed by airground_takeoff.sh:

Variable	Default	Purpose
UAV_DETECT	true	Start the UAV mine-detection stack.
UAV_DETECT_BACKEND	cpu	cpu uses the verified Ultralytics/PyTorch path; tensorrt is retained but not part of submission acceptance.
UAV_DETECT_DEVICE	cpu	PyTorch device passed to the detector.
UAV_SURVEY	false	Enable the automatic UAV survey route.
UGV_NAV	true	Start CMU planning and the target-tour node.
UGV_GRASP	true	Start MoveIt and request a grasp on target arrival.
GRASP_SOURCE	analytic	Grasp candidate provider. Keep analytic for the submitted configuration; GPD is not part of acceptance.
GRASP_DETECTOR	color	Wrist-camera mine detector used by the grasp pipeline.
FLIGHT_H	2.0	UAV flight height in metres.
LOW_ALTITUDE	true	Use the low-altitude EGO flight/planning configuration.
START_RVIZ	true	Start the elevation-map RViz configuration.
UGV_COST_SOURCE	elevation	UGV terrain-cost source.
UGV_GLOBAL_PLANNER	far	UGV global planner selection.
UGV_UAV_PRIOR	true	Fuse the gated UAV cloud into the UGV-centred elevation map.
UGV_ELEVATION_UPDATE	false	Add live UGV LiDAR updates to that elevation map when enabled.
UGV_MAP_SIZE	120	Elevation-map side length in metres.
UGV_MAP_RES	0.35	Elevation-map resolution in metres per cell.

Path, spawn, simulation, and sequencing overrides:

Variable	Default	Purpose
AIRGROUND_ENV_FILE	${XDG_CONFIG_HOME:-$HOME/.config}/airground/env.sh	Managed file used to remember the PX4 directory verified by setup. Override only when maintaining separate installations.
PX4_DIR	$HOME/PX4-Autopilot	PX4 source/build directory.
EGO_WS, UGV_WS	repository root	Workspace overlays used by UAV and UGV launch tabs.
UGV_WORLD	src/mobile_manipulator/worlds/outdoor_city.world	Default shared Gazebo world.
GAZEBO_WORLD	value of UGV_WORLD	World passed to gazebo_ros.
LIDAR_SDF	models/iris_depth_camera_lidar_terrain/model.sdf	PX4 UAV model SDF.
MAVROS_PX4_LAUNCH	/opt/ros/noetic/share/mavros/launch/px4.launch	MAVROS launch file.
UAV_POINTS_MAP_DIR	$HOME/pointcloud_maps	Saved UAV point-cloud directory.
SPAWN_X, SPAWN_Y, SPAWN_Z	0.0, -18.0, 1.5	UAV Gazebo spawn position in metres.
SPAWN_YAW	1.5707963	UAV spawn yaw in radians.
MAP_LOCAL_Z	0.0	Vertical offset of the UAV-local origin in odom; normally leave at zero.
PX4_GAZEBO_GUI	true	Show the Gazebo client.
PX4_SIM_SPEED_FACTOR	1	PX4 simulation speed factor.
PHYSICS_STEP, PHYSICS_RATE	0.005, 200.0	Gazebo time step and maximum update rate.
AG_ENABLE_GATE	derived	Override the altitude gate; normally derived from elevation-map ownership and UGV_UAV_PRIOR.
GAZEBO_LOAD_WAIT, PX4_WAIT, UGV_SPAWN_WAIT	10, 8, 6	Startup delays in seconds before the next launch tab is dispatched.
MAVROS_WAIT, ROS_WAIT	5, 20	MAVROS and ROS-layer startup delays in seconds.

These are all environment-variable overrides read by the integrated launcher; subsystem launch files expose additional direct roslaunch arguments. The detector launch's own confidence arg is 0.65 (src/uav_truth_tracker/launch/uav_mine_detection.launch). Change it only in a standalone detector evaluation; the integrated default uses 0.65.

Runtime topic inputs the pipeline consumes:

Topic	Type	Meaning
/mine_camera/rgb/image_raw	sensor_msgs/Image	UAV down-facing RGB image.
/mine_camera/depth/image_raw	sensor_msgs/Image	Time-aligned UAV depth image.
/mine_camera/rgb/camera_info	sensor_msgs/CameraInfo	RGB optical calibration used for 3-D projection.
RViz 2D Nav Goal → /move_base_simple/goal	geometry_msgs/PoseStamped	Optional manual UAV goal when UAV_SURVEY=false (default). The UGV does not consume this topic.
/ugv/start_tour	std_srvs/Trigger (service call)	Freeze the currently collected targets and begin the UGV tour. With the five-mine field, wait for confirmed_count: 5 first; detections received after this call are intentionally ignored.

With the default manual-UAV mode, use RViz's fixed odom frame and click near the five supplied mine-field positions below. The goal bridge converts the click into the UAV-local frame and applies FLIGHT_H; the coordinates are operator waypoints for this supplied world, not detector ground truth used by the algorithm.

Suggested click order	odom XY (m)
1	(1.4, 0.9)
2	(2.3, -1.1)
3	(5.5, 1.4)
4	(8.5, -1.4)
5	(11.5, 1.3)

Wait at each area until its marker turns confirmed, then continue. Do not call /ugv/start_tour until the YAML reports all five confirmations.

9. Output specification

Generated files and their overwrite behavior:

Path	Behavior
logs/setup_build.log	Latest setup_uav.sh catkin build log; overwritten by the next setup build and ignored by Git.
mine_detection_output/mine_map.yaml	Atomic snapshot of candidates and confirmed mines. Reset at integrated detector startup and overwritten when the map revision changes.
~/pointcloud_maps/uav_points_map_latest.pcd	Latest accumulated UAV cloud in odom; overwritten every 30 s and on clean shutdown.
~/pointcloud_maps/uav_points_map_YYYYMMDD_HHMMSS.pcd	Timestamped snapshot created every 30 s; files accumulate until manually removed.
~/trajectory_logs/flight_trajectory_YYYYMMDD_HHMMSS.csv	A new UAV actual/desired trajectory log for each mapping-node start.

These are generated outputs and are not source-controlled. Copy wanted results elsewhere before another run or cleanup. The legacy GPS/UTM uav_map_origin.yaml sidecar is deliberately disabled in the default egocentric run because its cloud is recorded in odom, not the legacy UAV map frame.

Live outputs are ROS topics rather than files:

Interface	Type	Meaning
/mine_detection/raw	uav_truth_tracker/MineDetectionArray	Per-frame depth-localized detections.
/mine_detection/map	uav_truth_tracker/MineMap	Associated and confirmed mine hypotheses.
/detected_targets	mobile_manipulator/WorldTarget	Confirmed mine locations handed to the UGV.
/ugv/tour_status	std_msgs/String	Target-tour state.
/ugv/goal	geometry_msgs/PoseStamped	Current UGV navigation goal in odom.
/elevation_mapping/elevation_map_postprocessed	grid_map_msgs/GridMap	Fused elevation and post-processed terrain layers.
/terrain_map	sensor_msgs/PointCloud2	Traversability/cost cloud consumed by UGV planning.
/target_tour_markers	visualization_msgs/MarkerArray	Pending/current/processed tour targets.
/uav0/mapping/points_world	sensor_msgs/PointCloud2	UAV cloud transformed into the shared odom frame.
/mine_detection/debug_image	sensor_msgs/Image	Timestamped detector visualization.
/mine_camera/diagnostics	diagnostic_msgs/DiagnosticArray	RGB-D rate, age, synchronization, encoding, and TF health.
/uav/pose_cov	geometry_msgs/PoseWithCovarianceStamped	UAV state covariance used by elevation mapping.

Custom UAV messages are defined under src/uav_truth_tracker/msg; the UGV handoff message is WorldTarget.msg.

Visualization: with START_RVIZ=true (the default), RViz opens with the fused/raw elevation maps, /terrain_map, /target_tour_markers, UAV cloud, robot model, and /mine_detection/debug_image already configured.

10. Success criteria

The run is working if, in order:

1. All five mines are confirmed before the target set is frozen — confirm:

   rostopic hz /mine_camera/rgb/image_raw
   rostopic echo -n1 /mine_detection/map
   grep '^confirmed_count: 5$' mine_detection_output/mine_map.yaml

   The map message must contain five entries with confirmed: true, and the YAML check must print confirmed_count: 5. Use RViz 2D Nav Goal to direct the UAV if automatic survey is disabled (the default). Do not start the tour early: the tour intentionally freezes its input set and ignores later detections.
2. The UGV tour starts with all five targets — call

   rosservice call /ugv/start_tour "{}"
   rostopic echo /ugv/tour_status

   The returned message and /ugv/tour_status must show total: 5; RViz's /target_tour_markers shows the current and pending targets.
3. The elevation map is populated — /terrain_map and the elevation-map RViz display show terrain rather than an empty grid.
4. Every mine completes the simulated grasp cycle — with UGV_GRASP=true, record one successful ALIGN → GRASP → NAV_HOME → PLACE sequence and a place OK result for each of the five targets. During GRASP, the wrist-camera target drives the approach; the jaws close visually and the executor selects the nearest matching landmine/landmine_* rigid body only for a Gazebo fixed-joint transport lock. That same model remains locked until /grasp/place detaches it. Only after all five successful PLACE events may terminal DONE be accepted. DONE by itself is insufficient because the tour can advance after a navigation/alignment/grasp failure. UGV_GRASP=false is a navigation-only diagnostic mode and must be identified as such when reporting results.

No persistent errors/exceptions in any gnome-terminal tab is a baseline expectation throughout. The supplied video is an UAV mapping/terrain-visualization excerpt; use the live checks above for full mission acceptance.

11. Stopping the program

From a new terminal in the repository root, run:

bash stop_airground.sh

Do not run another ROS/PX4 session under the same user at the same time. The launcher and stop script intentionally clear stale simulation state broadly: stop_airground.sh kills all nodes on the current ROS master and may terminate other Gazebo, ROS, RViz, PX4, or MAVROS processes owned by the user. Save unrelated work first.

The script requests a ROS shutdown, then terminates matching Gazebo, PX4, MAVROS, planner, detector, and grasp processes. It does not delete maps, models, source files, or build products. Shutdown is complete when Gazebo closes and the following command prints no matching process:

pgrep -af 'gzserver|gzclient|rosmaster|px4|mavros_node'

If a prior run ended abnormally, run bash stop_airground.sh once before starting again.

12. Troubleshooting

• ego-planner missing during build: run bash setup_uav.sh, or clone it into src/ego-planner/ and apply patches/ as the setup script does.
• No TensorRT executable / CUDA failure: run with UAV_DETECT_BACKEND=cpu UAV_DETECT_DEVICE=cpu; ensure the Python dependencies above are installed.
• No detections: verify the RGB, depth, and camera-info topics, then inspect /mine_camera/diagnostics and /mine_detection/debug_image.
• UGV does not move: targets must be confirmed before /ugv/start_tour is called; inspect /detected_targets, /ugv/tour_status, /state_estimation, and /terrain_map.
• Elevation map is empty: verify the latched transform with rosrun tf2_ros tf2_echo odom uav0/map_local and check the UAV cloud topic /uav0/mapping/velodyne_points_gated.
• Grasp stage fails or behaves inconsistently: set UGV_GRASP=false for a navigation-only diagnostic run. For a grasp-enabled run, inspect the physical mine selected: landmine_* and welded ... lines, /gazebo/model_states, /grasp/execute, and /grasp/place. docs/grasp_PLAN.md is a historical investigation log, not the current acceptance status.
• CMake warns that libgpd is absent: this is expected for the reproducible default. Keep GRASP_SOURCE=analytic; the package still generates its ROS messages and the workspace continues to build. GPD is an optional comparison backend, not a default dependency.

Acknowledgements

This workspace builds on ROS, Gazebo Classic, PX4, MAVROS, EGO-Planner, ANYbotics elevation_mapping / ETH grid_map, Clearpath Husky, Universal Robots UR5, Robotiq, MoveIt, GPD, DLIO, Ultralytics YOLO, and the Gazebo model collections included or referenced by the project.