[Paper] A Unified Experimental Architecture for Informative Path Planning: from Simulation to Deployment with GuadalPlanner

Source: arXiv

Source: arXiv:2602.10702v1

Overview

The paper presents GuadalPlanner, a unified experimental framework that lets researchers and engineers evaluate informative path‑planning (IPP) algorithms across simulation, software‑in‑the‑loop (SITL), and real‑world deployments without rewriting code. By cleanly separating high‑level planning from vehicle‑specific control, the authors bridge a long‑standing gap that has made it hard to transfer promising IPP ideas from the lab to production robots.

Key Contributions

• Unified Architecture – A modular pipeline that decouples decision‑making, sensing, and actuation, enabling a single IPP implementation to run unchanged in simulation, SITL, or on a physical platform.

• Standardized Interfaces – Definition of ROS 2‑based topics/services, MAVLink messages, and MQTT topics that act as a lingua franca between planner, sensor suite, and vehicle controller.

• Open‑Source Research Tool (GuadalPlanner) – Fully documented, extensible codebase that supports discrete‑graph environments, plug‑and‑play planning strategies, and easy integration with existing robotics stacks.

• Cross‑Domain Validation – Demonstrated consistency of algorithmic performance from Gazebo/ROS 2 simulations through SITL to a real autonomous surface vehicle (ASV) conducting water‑quality monitoring.

• Real‑Time Sensor Feedback Loop – Integration of on‑board environmental sensors (e.g., turbidity, temperature) that influence the planner’s utility function on the fly, showcasing true “informative” behavior.

Methodology

1. Abstraction Layers – The authors split the system into three layers:

   • Planner Layer – High‑level IPP logic that operates on a graph of waypoints.
   • Sensing Layer – Collects and timestamps sensor data, publishing it via MQTT.
   • Execution Layer – Vehicle‑specific controller that consumes ROS 2 commands and translates them to MAVLink for the ASV.

2. Interface Design – Each layer communicates through well‑defined interfaces:

   • ROS 2 topics (e.g., plan/next_waypoint) for inter‑layer messaging.
   • MAVLink commands (e.g., SET_POSITION_TARGET_GLOBAL_INT) for low‑level actuation.
   • MQTT messages for streaming sensor data.

   This modularity makes swapping out any component trivial.

3. Implementation Stack –

   • ROS 2 Foxy – Core middleware.
   • Gazebo – High‑fidelity simulation environment.
   • PX4 autopilot – Software‑in‑the‑loop (SITL) controller.
   • Custom ROS 2 ↔ MAVLink bridge – Runs on the physical ASV to translate commands.

4. Experimental Campaign – The same IPP algorithm (a utility‑maximizing greedy planner) was evaluated in three settings:

   • Pure simulation – Entirely within Gazebo.
   • SITL – Hardware‑in‑the‑loop autopilot with the ROS 2 bridge.
   • Field trial – Deployed on a 1‑meter‑wide ASV equipped with a multi‑parameter water‑quality probe.

   Performance metrics—coverage, information gain, and mission time—were logged consistently via the unified pipeline.

Results & Findings

• Consistency Across Domains – The drop in performance from simulation to the real world was under 3 %, confirming that the architecture preserves algorithmic behavior.
• Real‑Time Adaptation – When the water‑quality sensor detected a high‑variance region, the planner dynamically re‑prioritized nearby waypoints, increasing overall information gain by ~6 % compared to a static coverage planner.
• Developer Overhead – The same binary planner node was deployed unchanged across all three environments, cutting integration time by an estimated 70 % versus traditional ad‑hoc pipelines.

Practical Implications

• Faster Prototyping – Robotics teams can develop an IPP algorithm once and test it end‑to‑end without rewriting glue code for each platform, accelerating the research‑to‑product cycle.
• Cross‑Platform Portability – Because GuadalPlanner relies on industry‑standard protocols (ROS 2, MAVLink, MQTT), it can be adopted for aerial drones, ground rovers, or marine vessels with minimal effort.
• Continuous Integration & CI/CD – The unified pipeline lends itself to automated testing pipelines: simulation runs can be part of CI, SITL can serve as a staging environment, and a “smoke test” on a real robot can be triggered automatically before release.
• Real‑World Deployments – Industries such as environmental monitoring, precision agriculture, and infrastructure inspection can embed IPP logic that reacts to live sensor data, improving data quality while keeping mission duration low.
• Open‑Source Community – GuadalPlanner’s public repository (MIT‑licensed) invites contributions of new planners, sensor models, and vehicle adapters, fostering a shared ecosystem for IPP research.

Limitations & Future Work

• Discrete Graph Assumption – The current implementation works on pre‑discretized waypoint graphs; extending it to continuous‑space planners (e.g., sampling‑based methods) will require additional abstraction layers.
• Scalability of MQTT – While sufficient for the presented experiments, high‑frequency, high‑bandwidth sensor streams (e.g., LiDAR) may stress MQTT brokers. Future work could explore ROS 2 DDS QoS tuning or alternative middleware.
• Robustness to Communication Loss – The architecture assumes reliable ROS 2/MAVLink links; handling intermittent connectivity (e.g., underwater or beyond‑visual‑line‑of‑sight scenarios) remains an open challenge.
• Broader Benchmark Suite – The paper validates the approach with a single water‑quality mission; a more extensive benchmark across diverse environments and vehicle types would strengthen generalizability.

Overall, GuadalPlanner offers a practical, production‑ready bridge between IPP research and real‑world robotics, promising to reduce the “simulation‑to‑deployment gap” that has long hampered autonomous‑systems development.

Authors

• Alejandro Casado Pérez
• Alejandro Mendoza Barrionuevo
• Daniel Gutiérrez Reina
• Dame Seck Diop
• Samuel Yanes Luis
• Sergio L. Toral Marín

Paper Information