Abstract
This paper presents an adaptive AI planning framework designed to coordinate heterogeneous robot fleets—including Unmanned Aerial Vehicles (UAVs), Autonomous Surface Vessels (ASVs), and Inspection-and-Repair Robots (IRRs)—for the maintenance of offshore wind farms. Traditional planning methods often fail due to inaccurate static models of robot performance in harsh maritime environments. Our approach utilizes Bayesian learning to refine task duration and energy consumption models dynamically. Experimental results in both high-fidelity simulations and real-world onshore tests demonstrate that this adaptive strategy significantly improves mission success rates and scheduling accuracy compared to traditional static planning.
Introduction
Offshore wind energy is a vital component of the global transition to renewable power. However, Operations and Maintenance (O&M) costs remain high, often reaching 23% of total investment. These tasks currently rely on human technicians operating in high-risk environments with complex logistics. While multi-robot systems offer a safer and more cost-effective alternative, their coordination is challenging. Standard AI planners require precise domain models; however, manufacturer specifications for battery life and task duration rarely hold up in the unpredictable conditions of the North Sea. Underestimating resource needs leads to mission failure, while over-conservative estimates result in operational inefficiency.
Methods
The proposed system integrates high-level temporal planning with a data-driven feedback loop to address these uncertainties.
The framework is built on three core components:
- PDDL 2.1 Formulation: The wind farm domain is modeled using the Planning Domain Definition Language (PDDL), defining 13 actions for three robot types, including complex coordinated tasks like UAV-mediated IRR deployment.
- Adaptive Monitoring Strategy (AMS): This "brain" of the system includes a Discrepancy Checker to monitor state alignment and a Model Learner. The Model Learner employs Bayesian filters to update Gaussian distributions of task performance based on real-time sensor data, allowing the planner to "learn" environmental impacts on robot efficiency.
- Integration with ROSPlan: The system utilizes the POPF temporal planner within the Robot Operating System (ROS) ecosystem to generate and dispatch action sequences.
Evaluation and Datasets
The system was validated through two primary environments: a Gazebo simulation of a four-turbine offshore farm and a real-world onshore deployment involving a UAV and an IRR prototype.
The Model Learner was initialized using data from 30 simulated and 15 real-world trips to establish baseline performance distributions before testing adaptive replanning capabilities.
Key Results
The adaptive framework demonstrated clear advantages over static planning:
- Higher Mission Success: In simulations, the AMS-enabled planner successfully executed 10 out of 10 missions, whereas the standard POPF planner failed beyond 6 missions due to resource overestimation.
- Improved Accuracy: The "makespan" (total time to complete all tasks) predicted by the AMS planner was significantly closer to actual execution times compared to baseline planners.
- Real-World Efficiency: In physical tests, the system learned that a mission initially estimated at 26 minutes actually took only 8 minutes, allowing the planner to eliminate an unnecessary refueling stop and drastically shorten the total operation time.
Conclusion
This research demonstrates that adaptive temporal planning is essential for the reliable deployment of multi-robot fleets in offshore environments. By treating robot performance as a learnable distribution rather than a static value, the system achieves higher robustness and efficiency. Future work will focus on scaling the Bayesian learning to larger wind farms and improving communication resilience in maritime settings.
Summary updated