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

Technical Summary: A Unified Experimental Architecture for Informative Path Planning (GuadalPlanner)

1. Problem Statement

Informative Path Planning (IPP) is critical for autonomous missions like environmental monitoring and search-and-rescue, where the goal is to maximize information gain rather than just reaching a destination. However, the authors identify a significant "gap" in current IPP research:

• Fragmentation: Most IPP algorithms are validated in custom, highly controlled, and non-generalizable simulation environments.
• Lack of Transferability: There is a lack of standardized tools to bridge the gap between simulation and real-world deployment. Transitioning an algorithm from a simulator to a physical robot usually requires substantial code rewriting and reconfiguration.
• Coupling: Existing frameworks often intertwine high-level decision-making logic with low-level vehicle dynamics, making it difficult to test the same algorithm across different levels of realism.

2. Methodology

The paper proposes GuadalPlanner, a modular and extensible experimental architecture designed to decouple high-level decision-making from vehicle-specific control. The methodology is built on three pillars:

A. Three-Level Abstraction Model:
The architecture allows researchers to test the same algorithmic logic across three increasing levels of fidelity:

1. Algorithm-level Simulation: Uses ideal, discrete graph-based environments and "teleporting" vehicles to validate high-level logic and strategic behavior without physical constraints.
2. Software-in-the-Loop (SITL): Integrates realistic vehicle dynamics and autopilot behavior using ArduPilot SITL, allowing for testing under realistic motion constraints.
3. Real Deployment: Executes the identical code on physical autonomous vehicles (e.g., Autonomous Surface Vehicles - ASVs) using live sensor feedback and real-time actuation.

B. Technical Stack & Middleware:
The system leverages widely adopted robotics technologies to ensure interoperability:

• ROS2: Acts as the middleware for modular communication between nodes (sensing, planning, and control).
• MAVLink & MAVROS: Facilitates communication between the ROS2 environment and the vehicle's Flight Controller Unit (FCU).
• MQTT: A lightweight messaging protocol used to bridge the decision-making module (which can run on a remote server) and the vehicle (via a companion computer), enabling remote execution and multi-agent coordination.

C. Software Design:
Implemented in Python using object-oriented principles, the architecture uses class inheritance to allow specialized environments (e.g., oil spill tracking, trash detection) to extend a base graph environment.

3. Key Contributions

• Unified Architecture: The first end-to-end experimental platform that explicitly decouples high-level IPP decision-making from low-level vehicle control.
• Interface Consistency: Provides a standardized "fleet interface" so that planning algorithms interact with simulated and real vehicles identically.
• Open-Source Tooling: An extensible research tool that supports multi-agent configurations, modular environments, and real-time sensor feedback.
• End-to-End Pipeline: A complete workflow from prototyping in discrete graphs to field deployment on physical hardware.

4. Experimental Results

The authors validated the architecture through three stages of testing:

• Simulation: Demonstrated that the system can handle multi-agent coordination in diverse scenarios (e.g., Gaussian Process estimation and deterministic oil spill tracking).
• SITL: Validated the ROS2/MQTT/MAVLink communication pipeline and provided a web-based interface for real-time mission monitoring.
• Real-World Deployment: An ASV was deployed in a lake to measure water conductivity.
  • Algorithms Tested: Greedy (uncertainty-based), Expected Improvement (Bayesian optimization), and Flooding (systematic coverage baseline).
  • Findings: All algorithms successfully reconstructed the environmental model. While the "Flooding" method provided complete coverage, the "Greedy" and "Expected Improvement" methods were more efficient at reducing local uncertainty and variance. The system proved robust against real-world noise (GPS, currents, and latency).

5. Significance

GuadalPlanner addresses a fundamental bottleneck in robotics research. By providing a "plug-and-play" framework for IPP, it enables:

• Reproducibility: Researchers can share standardized modules and environments, allowing for fair comparisons between different algorithms.
• Efficiency: It reduces the time and engineering effort required to move from a theoretical paper to a field-ready autonomous system.
• Scalability: The architecture supports the transition from single-agent to multi-agent scenarios and from local simulation to remote, cloud-based decision-making.