[Exercise Proposal] Mine/Tunnel Inspection with ground rover: New Mobile Robots exercise

[Kunal-Somani]

Hi,

Following my contributions to the repository (16 merged PRs to date), I have been working on a new exercise proposal for RoboticsAcademy and wanted to get your thoughts before opening a PR.

cc @jmplaza

What is missing from the current Mobile Robots portfolio

The current mobile robots exercises cover open environments well with warehouses, homes, roads, F1 circuits. There is no exercise covering:

• Confined-space navigation (narrow tunnels, GPS-denied)
• Multi-sensor fusion as an explicit teaching objective
• Hazard detection combined with autonomous navigation
• Environments where camera + LiDAR alone are insufficient and IMU fusion becomes necessary

This gap matters because confined-space inspection is one of the largest real-world deployment areas for autonomous ground robots - mining, utility tunnels, post-disaster USAR, underground infrastructure. The DARPA SubT Challenge ran for three years specifically because this domain is unsolved and educationally rich.

The proposed exercise

A ground rover navigates a 4-way mine tunnel junction, autonomously inspecting for:

• Structural defects (wall cracks) detected via RGB camera
• Hazardous gas pockets (CH4, CO) detected via virtual gas sensors correlated with camera zones
• Pose estimation in GPS-denied conditions, which requires fusing wheel odometry + IMU + 2D LiDAR

The student programs in Python (and C++ in a follow-up) using the standard HAL/WebGUI pattern.

What is already implemented

Across the three repos:

RoboticsInfrastructure (feature/mine-inspection-universe):

• Differential drive rover with GPU LiDAR (360 degrees), RGB camera, IMU with noise model
• Parametric 4-way tunnel world: exterior arid terrain with directional sunlight, mine portal entrance, cliff faces, 4 tunnel arms with continuous uneven floor for realistic IMU excitation, rock debris for LIDAR targets
• Mine Cart and Mine Cart Engine props (OpenRobotics, CC-BY 4.0)
• Virtual gas and crack sensor publisher (CH4, CO at 3 pocket locations, 5 crack zones) synchronized with visible world geometry
• Full ROS2-GZ bridge YAML including IMU topic
• Launcher following the labyrinth_escape pattern

RoboticsAcademy (feature/mine-inspection-app):

• HAL.py exposing: setV, setW, getPose3d, getLaserData, getImage, getIMU, getCH4, getCO, getCrackDetected
• getIMU() is new to the exercise portfolio -- returns orientation, angular velocity, and linear acceleration, enabling students to implement EKF-style sensor fusion
• WebGUI.py using MeasuringThreadingGUI (Harmonic), two-pane camera + SLAM map with live sensor overlay
• React frontend with gas/crack readings HUD
• db.sql registration: exercise id 27, PROTOTYPE status, MULTILANGUAGE tagged

The exercise has been tested locally in Gazebo Harmonic; all sensor topics publish correctly, the rover moves and responds to cmd_vel, IMU data flows, and virtual sensors respond to rover position.

Questions before opening PR

1. Is a Mine/Tunnel Inspection exercise a welcome addition to the Mobile Robots category?
2. Any preference on how the virtual gas/crack sensors should be structured; current approach uses a standalone Python publisher node that reads odometry and computes proximity to defined hazard coordinates. Is there a preferred pattern?
3. Should the initial PR include the C++ template, or is Python-only acceptable for the PROTOTYPE merge?

[jmplaza]

Hi @Kunal-Somani , nice idea. But you have to go step by step: first, create the scene in the RoboticInfrastructure repo, and the robot you want to use. They together will be the world for that exercise. Regarding the defects, they typically will be detected using a Deep learning network (so it will be interesting to create or locate a dataset to train that network model...).

We have in the roadmap two additional inspection exercises: (a) one extending the Drone-PowerTowerInspection using a drone (creating the corresponding dataset for training the model which will detect the artifacts in the Power Towers); (b) solar panel inspection with a drone. They both, as your proposed exercise, come from real world applications.

So I would suggest to go ahead with the robot and the scene to be used in your Mine/tunnel inspection exercise.

[Kunal-Somani]

Hi @jmplaza, thank you for the response.

The world loads cleanly in Gazebo Harmonic and all sensor topics including LiDAR, camera, IMU and odometry are verified working via ROS2 bridge. As mentioned in my earlier message I have the full setup running in Gazebo, so since you have given it a green signal I will go ahead and open the PR for the RI repo first.

For crack and defect detection I have currently implemented proximity-based virtual sensors as placeholders so students can work on the navigation and localization aspects in the meantime. I will make the refinements you suggested to move toward a proper deep learning based detection approach as I work on finalizing the scene and robot for the RI PR. Once the RI PR is reviewed and merged I will finalize the RoboticsAcademy side and open a second PR for that.

I would really appreciate your guidance on whether there are any additions or changes to the implementation steps I have described that I should address before opening the RI PR, and if there is anything specific you would like to see in the scene or robot configuration.

Thank you.

[jmplaza]

Please share here some images and (links to YouTube) videos of the working world (=scene + robot). That will help to give you better feedback.