A Three-Robot Inspection Testbed in ROS 2 and Gazebo Classic: Static-World A Navigation and Separate Language-Grounded Task Planning*
Year: 2026
Type: preprint
Abstract: This preprint presents a simulation-based testbed for multi-robot inspection developed using ROS 2 Humble and Gazebo Classic. The system comprises three TurtleBot3 Burger robots operating in a shared simulated environment, with robot-specific namespaces, transform frames, and sensor interfaces for RGB-D, LiDAR, IMU, and odometry. A centralized mission executor distributes ten unique inspection waypoints among the robots. Navigation uses a custom grid-based A* planner derived from static Gazebo SDF geometry, odometry-based path following, and a forward-LiDAR stop/resume guard. In a recorded nominal simulation run, all ten waypoints were completed in 197.4 seconds. Final goal distances ranged from 0.048 m to 0.050 m, within the configured 0.15 m tolerance. SLAM Toolbox generated a partial occupancy map with 63.5% known-cell coverage. The work also presents a separate language-grounded semantic-planning pipeline based on a configured scene graph, a deterministic planning baseline, and an optional locally hosted Qwen 2.5 1.5B model. A deterministic validator checks proposed structured tasks, while scripted robot- and target-unavailability events demonstrate software-level plan updates. A recorded regression-test run collected 61 tests: 60 passed, one was skipped, and none failed. The navigation and semantic-planning pipelines were evaluated separately. Semantic plans are not automatically executed by the Gazebo navigation stack, and complete mission recovery following the reported robot-failure injection was not demonstrated. All experiments described in this work are simulation-based. The preprint documents the system architecture, implementation, experimental observations, reproducibility procedure, limitations, and directions for future research. It has not undergone peer review.
Author: Sharma, Akshit
Institutions: University of Delhi, Indira Gandhi National Open University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225651169