A multirobot system for autonomous deployment and recovery of a blade crawler for operations and maintenance of offshore wind turbine blades

Offshore wind farms will play a vital role in the global ambition of net zero energy generation. Future offshore wind farms will be larger and further from the coast, meaning that traditional human‐based operations and maintenance approaches will become infeasible due to safety, cost, and skills shortages. The use of remotely operated or autonomous robotic assistants to undertake these activities provides an attractive alternative solution. This paper presents an autonomous multirobot system which is able to transport, deploy and retrieve a wind turbine blade inspection robot using an unmanned aerial vehicle (UAV). The proposed solution is a fully autonomous system including a robot deployment interface for deployment, a mechatronic link‐hook module (LHM) for retrieval, both installed on the underside of a UAV, a mechatronic on‐load attaching module installed on the robotic payload and an intelligent global mission planner. The LHM is integrated with a 2‐DOF hinge that can operate either passively or actively to reduce the swing motion of a slung load by approximately 30%. The mechatronic modules can be coupled and decoupled by special maneuvers of the UAV, and the intelligent global mission planner coordinates the operations of the UAV and the mechatronic modules for synchronous and seamless actions. For navigation in the vicinity of wind turbine blades, a visual‐based localization merged with the location knowledge from Global Navigation Satellite System has been developed. A proof‐of‐concept system was field tested on a full‐size decommissioned wind‐turbine blade. The results show that the experimental system is able to deploy and retrieve a robotic payload onto and from a wind turbine blade safely and robustly without the need for human intervention. The vicinity localization and navigation system have shown an accuracy of 0.65 and 0.44 m in the horizontal and vertical directions, respectively. Furthermore, this study shows the feasibility of systems toward autonomous inspection and maintenance of offshore windfarms.


| INTRODUCTION
As a consequence of increasing climate change awareness and oil shortage, research and policies for renewable energy have been gaining greater attention in recent years. The EU Green Deal, as one of the policies, aims to make the EU become the first climate-neutral continent by 2050 (Tsiropoulos et al., 2020). With this increasing demand, wind energy has become one of the most promising ways to produce renewable energy. The use of turbines has reached a global cumulative wind power capacity of 743 GW and has rapidly been increasing with 50 and 93 GW of new installations in 2019 and 2020, respectively (IRENA, 2019; Lee & Zhao, 2021). Offshore wind farms are particularly attractive as the farms and turbines could exploit stronger uninterrupted winds, be made larger, and minimize conflict of interests with other aspects of society (e.g., visual disturbance) (Bergström et al., 2014). According to IRENA (2019), offshore wind farm installations are predicted to reach 228 GW in 2030, compared to the 23 GW installed in 2018. However, due to their remote location, wind farms are exposed to unpredictable and harsh weather, which results in highly variable operational conditions and, in turn, to intense and costly maintenance operations (Zion Market Research, 2019). The current practice utilizes vessels to transport technicians and equipment to the turbines to perform close-up visual and nondestructive testing (NDT) inspections. Following this, rope access technicians are deployed to carry out maintenance wherever required.
To maintain offshore wind farms performing at their optimal level using this conventional method, it costs approximately 25% of the wind turbine installation cost over 25 years of service (Nilsson & Bertling, 2007). As a comparison, the Levelized costs of electricity (LCOE) for an onshore windfarm constitutes only 5% of its total investment cost (Ren et al., 2021). This costly process can be attributed to two factors: (i) the length of turbine downtime when rope-access technicians perform close-up visual or NDT inspections, and (ii) the daily use of crew transfer vessels, including the rope access technicians for round trips to and from the farms (Kabbabe Poleo et al., 2021). Additionally, technicians are subject to health-and-safety risks due to harsh environments and weather conditions. Recent advancement in the development of robotic platforms has opened up new opportunities in the automation of the operation and maintenance (O&M) of offshore wind farms. Employing autonomous systems to maintain offshore assets provides the benefits of reducing the cost significantly while also removing the health-andsafety risks associated with the technicians (Stetco et al., 2019). One prominent research direction in wind farm O&M is the use of blade inspection, maintenance, and repair (IMR) robots such as BladeBug (see Figure 2) where these robots are remotely operated and controlled to perform inspection and repair on wind turbine blades.
However, to ensure robustness and safety of operations on and around wind turbines, many current practical implementations rely on manual processes where the deployment and retrieval processes of the IMR robots on the turbine blades are performed by rope access technicians. As such, crew transfer vessels are still required and health-and-safety risks are still present to technicians.
Within the multiplatform inspection, maintenance and repair in extreme environments (MIMREE) (Bernardini et al., 2020) project, a multirobot system is proposed to automate the process of deploying and retrieving the blade IMR robot on O&M of offshore wind farms. It is believed that multiple autonomous robots are more robust to failures and can achieve better overall performance in their execution due to their collaborative nature (Piacentini et al., 2019;Welburn et al., 2019).
However, utilizing multiple heterogeneous robots, as suggested in Bernardini et al. (2020), requires robust coordination and implementation as each robot may have its set of own functionalities and roles. As shown in Figure 1, the autonomous process consists of six steps: (A) an unmanned aerial vehicle (UAV) and an IMR robot are transported by an ASV from the onshore control center to the offshore wind farms; (B) the UAV takes off with the IMR robot; (C) the IMR robot is safely deployed on the wind turbine blade by the UAV; (D) the robot is recovered by the UAV after completing its mission; (E) the UAV carries the robot and returns to the ASV; and (F) the ASV transports the entire system back to the onshore control center. The scope of this paper focuses on steps C and D, introducing a novel deployment and retrieval system to overcome the challenges in autonomously and safely deploying and recovering an IMR robot from a wind turbine blade, to fully eliminate the need for human involvement in close-up inspection and repair of offshore wind turbine blades using rope access techniques.
The main contribution of this paper is (1) novel designs of mechatronic systems for deployment and retrieval of a blade IMR robot such as a BladeBug MK II robot (see Figure 2); (2) an UAV equipped with a LiDAR-based navigation and control system; (3) a centralized AI planning software for autonomous coordination and control between the UAV and the mechatronics. These prototype systems have been developed and tested in field trials demonstrating the robustness of our proposed approach. Results show that the proposed approach is effective in retrieving an IMR robot from a wind turbine blade safely and robustly without the need for human intervention.

| RELATED WORK
This section lays out important developments in recent years related to our work. They are grouped into three different sections according to our most critical contributions.

| Air-based ground object retrieval systems
In recent times, helicopter hoist systems are frequently used in disaster rescue operations to perform vertical lift of a ground payload (Nonnenmacher & Jones, 2016). As the helicopter hovers above the payload, a rescue basket is lowered to the ground level to accommodate the payload, which is then retrieved. Such a process is often manual. The rescue basket is usually a passive device that cannot control its touchdown location or possess any equipment to grasp the payload automatically. Young (2007) described an advanced rescue basket that has two Tilt Nacelle Ducted-Fan thrusters installed on the sides, effectively transforming it into an aerial vehicle. The rescue basket can be remotely controlled to fly in the desired trajectory to reach locations inaccessible with the traditional helicopter hoist system. The technology provides local control of the slung load position which can be potentially used to drive a winch hook to engage pickup features on an IMR robot.
However, miniaturization of such thrusters on a UAV of much less weight and power than a helicopter may be a challenging task.
Adding autonomy to the vectored hoist may also significantly increase control complexity in a gusty environment.
Before helicopter hoist systems were developed, the Fulton surface-to-air recovery system (Heinrich et al., 2018) was used by the US military to recover ground targets. The system requires deployment of a helium balloon from the ground that sends a lift line into the sky with one end of the lift line attached to the payload. The lift line can then be captured by a V-shaped fork installed on the nose of an aircraft (Leary, 2008). However, this system is entirely manual, and there are many challenges to adopting this concept in a much smaller system using autonomous agents. First, a balloon releasing mechanism is needed on the IMR robot, which could add significant weight and increase the physical size of the robot. Second, gust wind and turbulence could cause the balloon to move violently and unpredictably, requiring an agile maneuver of the aircraft to catch the lift line, which is difficult in such an environment. Third, the lift line may be entangled in the aircraft risking failing the entire system. Ivanovic et al. (2018) demonstrated a novel technology for retrieving a surfaced autonomous underwater vehicle (AUV) using a UAV. During a mission, a compliant net is winched down from the UAV until afloat. Once the AUV navigates itself into the net, the net is lifted, automatically contracting around the AUV. This device is specifically designed for robots with streamlined shapes that allow them to safely navigate into the net without getting entangled.
However, in our case, the shape of an IMR robot is much more complex. Additionally, maintaining the net on the curved surface of a wind turbine blade is challenging, especially in a gusty environment. Roderick et al. (2021) studied the anatomy of bird feet and developed a biomimetic UAV that uses bird-like robotic legs called stereotyped nature-inspired aerial grasper (SNAG) for its landing gear. SNAG features claws that can grasp and release objects of F I G U R E 2 BladeBug MKII IMR robot F I G U R E 1 Illustration of the MIMREE multirobot system for offshore wind farm operations and maintenance JIANG ET AL. | 75 various shapes up to 105 g in the demonstration. Thomas et al. (2013) developed an agile micro-UAV system with a gripper that can grasp an object at a high approaching velocity. The demonstrated system weighing 658 g could grasp a 27 g cylindrical target at speeds up to 3 m/s.  developed and tested a quadcopter with a custom Fin Ray gripper that can grasp and transport small objects of different geometries and surfaces reliably at a speed of 1 m/s. They used a motion tracking system to localize the test objects and the UAV. Inspired by eagle's talon, Tscholl et al. (2022) developed a UAVmounted gripper system that uses hydraulically amplified self-healing electrostatic (HASEL) actuation. The above authors assumed a static lab environment and real-world environmental factors such as wind and associated gusts were not considered. In addition, the dynamics of the UAVs and the grippers were demonstrated with small and light-weight payloads that weighed, for example, less than 4% of the entire system in Thomas and Polin's system. In contrast, an IMR robot may weigh over 35% of the deployment and retrieval system. Fiaz et al. (2018) described a passive magnetic grasping and impulsive release system for aerial transport of ferrous objects. The design used four neodymium magnets that allowed a maximum payload of 2.6 kg in a static environment. Due to the sliding effect of Coriolis force (Fiaz et al., 2017), the maximum feasible payload can become much less depending on the flying environment, and the weight, shape, and center of mass of the payload. In their field experiments, the system successfully picked up ferrous disks of 350-500 g under a wind speed of 27 kph. The pickup required the UAV to descend very close to the ferrous disks.
All above airborne ground object retrieval systems present practical methodologies in their specific use cases. However, some of them (e.g., Heinrich et al., 2018;Nonnenmacher & Jones, 2016;Young, 2007) involve complex manual processes which are difficult to automate in a compact robotic platform, and others are not suitable for retrieving a blade IMR robot for the reasons already discussed above. The magnetic grasping system (Fiaz et al., 2018) and a few seemingly straightforward approaches, for example, (a) landing the UAV on the top of BladeBug or (b) landing the UAV near BladeBug and commanding BladeBug to walk into the UAV's undercarriage, are challenging to achieve and have associated risks. These risks are mainly due to the reduced maneuver precision of a UAV in a gusty environment, which may cause the UAV to crash land, and the lack of precise leg posture control in BladeBug that may cause difficulties for the robot to navigate into the UAV's undercarriage.

| Autonomous navigation around objects
Autonomous UAV flights in the vicinity of objects are very sensitive operations for which high precision is required. Around wind turbines, rough weather conditions and uncertainty in the knowledge of the exact location of a turbine blade are the main factors to consider when designing a navigation system for this environment.
Venugopalan, Taher, and Barbastathis in Venugopalan et al. (2012) developed an algorithm for the autonomous landing of a commercially available quadrotor (AR Drone) on an autonomous kayak considering the harsh marine environment. The algorithm consists of two sections: (1) landing pad detection (image processing) and (2) control for navigation. In the first section, a landing pad was designed based on distinctive patterns and colors, and detection was carried out offline in MATLAB. A traditional PID controller was used for the control part, with a Ziegler-Nicholas method and trial-and-error approach utilized to tune the gains. Navigation was achieved through a cascade architecture where the UAV's hover mode (inner-loop) was wrapped around by the developed controller (outer-loop). A success rate of 75-80% with an accuracy of better than ±20cm was achieved. As deduced from the literature, rather than relying solely on GPS/GNSS for accurate navigation, researchers have adapted other methods among which onboard LiDAR and vision systems are popular. In this paper, an approach merging GNSS localization with LiDAR data has been adopted.

| Cooperation and coordination amongst autonomous agents
Inspired by Bernardini et al. (2020), in this study, inspecting components of a wind turbine that might be faulty and repairing them is performed by heterogeneous robots and, in particular, UAVs and IMR robots, as it is believed that multirobot systems tend to be more robust to failures and achieve better performance in their execution. However, the approach of utilizing heterogeneous robots requires solid coordination, high-level cooperation, and the proper execution tools to achieve autonomy. Furthermore, it becomes more  PDDL Fox and Long (2003) to model missions, which is defined as planning problems, and how each system should interact and coordinate with one another in achieving the missions, which is defined as the planning domain. Both the problems and the planning domain were fed and translated by a planner to get a plan, that is, a series of actions, for each individual robot to execute. However, these works consider only homogeneous robots with similar capabilities, which limits their applicability in contexts with heterogeneous robots.
Unlike the works above, the involvement of different types of robots in this study means that each type has a different set of actions. This differentiation of actions among different robots creates coordinated actions that require more complex low-level control that not only involves the main robot but also requires the partner's robots to be actively engaged in. For example, the implementation of retrieving an IMR robot requires the robot controller to perform a series of actions such as preparing its attaching features, orientating the robot in an agreed direction, and performing an embrace position for the retrieval while, on the other side, the UAV flies toward the IMR robot and prepares its attaching module.

| MECHATRONIC SYSTEMS FOR DEPLOYMENT AND RETRIEVAL
Three mechatronic systems, including the robot deployment interface (RDI), the link-hook module (LHM), and the on-load attaching module (OLAM), have been developed for deployment and retrieval of an IMR robot, that is, BladeBug MK II. As briefly introduced in Section 1, BladeBug MK II is a relatively compact hexapod designed as an experimental wind turbine blade IMR platform. It is approximately 450 mm in length, 360 mm in width, and 160 mm in height with a mass of 4.5 kg. Figure 3 shows the integration of the mechatronic systems with BladeBug and a UAV platform named Goliath, forming F I G U R E 3 System diagram for the deployment and retrieval mechatronic systems JIANG ET AL. | 77 three major systems, that is, IMR robot system, retrieval UAV system (UAS), and deployment UAS. The global mission planner (GMP) is the brain that collects feedback from different robotic systems, calculates a mission plan, and broadcasts system-specific instructions for performing a coordinated operation.  Table 1.
The mechatronic systems, in specific, refer to the robot deployment interface (RDI), the link-hook module (LHM), and the OLAM. RDI and LHM are part of the systems for deployment and retrieval, respectively, that are mounted in the undercarriage of Goliath. The OLAM is attached to BladeBug providing coupling interfaces for both the RDI and the LHM.
Detailed designs of the mechatronic systems are discussed further in this section, and GMP is discussed in Section 5.

| Deployment and retrieval procedures
The proposed aerial-based deployment and retrieval system emphasizes the safety of the assets, such as wind turbines, and operational reliability when flying in a gusty environment with a heavy payload (at least 4.5 kg).
Figures 5 and 6 illustrate the deployment and retrieval procedures used by the proposed system, respectively. As can be depicted, the deployment procedures involve four stages and assume the system is prepared onshore where the IMR robot system is manually coupled with the UAS and locked in the undercarriage. As the UAV arrives at the waypoint near the target wind turbine blade, the blade surface must be scanned to determine a safe location for deployment. Details of this technology are introduced in Section 4. The IMR robot system is only released when the UAS lands securely to ensure the safety of the wind turbine and the robotic systems. After the IMR robot system is safely deployed on the blade, the UAS returns to a designated area on the ASV.
As the IMR robot system completes its mission on the wind turbine blade, it communicates with GMP to request recovery. The retrieval UAS first navigates near the wind turbine blade based on GPS coordinates, and meanwhile, the OLAM unfolds itself and prepares for engagement. After the GMP confirms the OLAM is ready for engagement, the UAS aligns with the OLAM using RTK GPS and performs an engagement maneuver. Coupling between the two F I G U R E 4 (a) The UAS for deployment (Goliath + RDI) and the IMR robot system (OLAM + BladeBug); (b) The UAS for retrieval (Goliath + LHM) and the IMR robot system

| Robot deployment interface
The RDI, as part of the deployment system, is mounted on Goliath, as can be seen in Figure 4a. It can be coupled with the interface on OLAM, which is attached to BladeBug, for locking the IMR robot system in the undercarriage of Goliath and releasing the system when requested.
As shown in Figure 8, the RDI has a relatively simple construction, where two servo locking devices (Tarot TL2961-02) are used as the locking and releasing mechanism and four guide blocks are designed for accurate and repeatable alignment with the interface on the OLAM. The RDI is designed to carry a payload of maximum 10 kg and operates with PWM input from the UAV directly with a low power consumption of maximum 2.5 W.

| Link-hook module
The LHM consists of three major parts, a controller compartment, a 2-DOF hinge and a hook as shown in Figure 9. (Stage D). Swing motion reduction is necessary because the nonrigid setup may result in significant swing motion during the flight, which could cause the hook to miss the pickup line. Swing motion reduction can be achieved in both passive and active manners. Due to magnetic induction, rotation of the servos caused by swing motion produces an electromotive force that opposes the direction of rotation, providing passive dampening to the swing motion. In an active swing reduction mode, both servos operate in their built-in current control algorithm.
By setting the current target to zero, motors are driven in the opposite direction of the swing motion, which reduces the current induced by the back EMF generated by the swing motion.

| UAV NAVIGATION AND CONTROL NEAR WIND TURBINE BLADE
To navigate around wind turbine blades, specifically for precisely deploying the IMR robot, a method combining GNSS and LiDAR localization is utilized. Initially, GNSS is used to take the UAV to a predetermined location in front of the target turbine blade. At this stage, a scanning LiDAR is employed to detect the blade, and a control law, a combination of position and velocity commands, is utilized to guide the UAV to the highest point on the blade's crosssection. A prototype multirotor in an H configuration was developed to validate the developed algorithm, Figure 10.
The components used in the prototype UAV have been summarized in Table 2 below. To scan the vertical plane, a Lightware (LW20) LiDAR has been attached to a servo which has been mounted to the front of the UAV. This system has been calibrated in a lab environment to find the relationship between the input PWM to the servo and the LiDAR angle to the horizon. Figure 11 illustrates how a LiDAR scanning is conducted, the blue region being a typical scan range, along with the reference frames of the blade cross-section and the LiDAR. The scanning angle resolution is controlled through the input PWM to the servo.
Wing detection is achieved by defining a threshold and detecting the jumps in the measured range vector, as can be seen in Figure 12.
F I G U R E 8 CAD design and specifications of the robot deployment interface (RDI), dimension units in mm (Mass = 0.34 kg, payload capacity = 10 kg, and rated power consumption = 2.5 W).
F I G U R E 9 CAD design of the link-hook module, dimension units in mm JIANG ET AL.

| 81
The landing target is then defined as the highest point on the detected blade. Figure 13 shows the navigation waypoints in landing the UAV on a turbine blade. The first two waypoints are reached by sending position commands in the local coordinate frame of the UAV. An outer loop proportional velocity controller is then developed and utilized to get to waypoints 3 and finally 4 (landing target). A block diagram of the navigation algorithm can be seen in Figure 14.

| GLOBAL MISSION PLANNING SYSTEM
To automate missions to inspect, repair, or maintain wind turbines involving heterogeneous robots, a central control manager is necessary to (1)  blade IMR robots, and autonomous surface vessels (ASV) (Jovan & Bernardini, 2021). For this study, we stripped down the control of the ASVs, as shown in Figure 15, due to the difficulty of using a real ASV for our experiments.   Table 3 for an example). The action dispatcher module disassembles the plan into individual actions according to (1)  Each robot's control interface (i.e., UAV control interface module, and IMR robot control interface) decides how the high-level actions are executed. For example, the low-level implementation of the action uav_deploy_imr (?uav,?imr,?uav_wp,?imr_wp) requires the use of our novel technique of UAV navigation around a wind blade, which is described in Section 4.

| Monitoring and replanning strategy
Automating the maintenance process using a multirobot system based on temporal planning has the potential of significantly reducing the costs spent annually because automation with parallel actions typically makes the process efficient. However, in practice, the efficacy of a planningbased system depends on how accurately the domain experts design the planning domain and how precisely the problem description represents the world and its operating conditions. Inaccuracies in any of those two may lead to inefficient plans and mission failures.
In our approach, the GMP tries to maintain states of the world that are as accurate as possible by reducing the discrepancy between the expected states, that is, the conditions of the world resulting from simulating action of a generated plan, and the actual states, that is, the actual conditions obtained when a robot executes an action of the generated plan. These discrepancies usually come from sensor/ We use a procedural approach to action refinement and execution monitoring. A function that refines an action operator, for example, uav_inspect_wt(?uav,?wp), into low-level commands resides on each vehicle control interface (shown in Figure 15). Each of these functions communicates its completion of the action to the problem generator. A failed output from a function triggers the problem generator to remove the corresponding predicates and request a replanning to ROSPlan, while a successful output triggers the problem generator to add the effect predicates of the action to the KB. Figure 16 shows a plan of the IMR robot retrieval mission with its crucial state transition during the execution of an action. Each action block implementation resides in the UAV/IMR robot control interface module, while the state transition process (i.e., changing the effect of an action) is implemented in the problem generator module. One should note that the precondition states and effect states of each action are not limited to what is pictured in Figure 16. Other states such as robot's battery/fuel level, the availability of the sensor/actuator, and whether the involving robots are busy are also considered as preconditions of each action to ensure robustness in the generated plans. 4 Replanning is also implicitly tailored to the action that has failed. For example, in the case of insufficient fuel during a mission, the current mission will be aborted, and a replan must be requested that prioritizes all robots returning to shore/ground.

| EXPERIMENTS
To validate the developed prototype systems, four experiments have been conducted, including field tests of UAV navigation, LHM verification experiments, a field tests of the RDI and field tests of T A B L E 3 Generated plan for a mission involving a UAV for a camera inspection and an IMR robot for an NDT inspection  Figure 17. The wind turbine blade was horizontally supported on several pedestals that raised the blade approximately 1-2 m from the ground.

| Field test of UAV navigation
A novel navigation system, as outlined in Section 4, has been developed for the UAV to accurately place the IMR robot at the highest point on the turbine blade at a given distance from the wind turbine's hub. To verify the performance, a retired turbine blade placed horizontally near ground level (see Figure 17) was used. In these experiments the procedures detailed in Figure 13 in Section 4 have been followed. To investigate repeatability as well as accuracy, the test was repeated four times. Figures 18 and 19 illustrate the processed LiDAR data of two of the scans taken at WP1 (preset at F I G U R E 16 A generated plan for an IMR robot retrieval mission focusing on the internal state transition of each action. Crucial states, both precondition and effect, are presented in the action block. If the effect of an action is not met after the execution, a replan will be requested by the problem generator module. F I G U R E 17 Test site with a retired wind turbine blade F I G U R E 18 LiDAR scan from WP1 as defined in Figure 13 (preset at 6 m in front and 5 above the wing) 6 m in front and 5 m above the wing) and WP2 (preset at 3.5 m in front and 5 m above the wing). In these figures, the range vectors have been broken down into x distance in front and z distance below the UAV. The red points in these figures illustrate the target location (highest point on the wing) to deploy the IMR. The turbine blade in these figures has been identified by detecting the jumps in the LiDAR range data.  there has been a higher accuracy in the z direction compared to x direction. This could be due to the fact that UAV localization is more accurate in the z-direction (as a barometric pressure is also used). However, to gain a better understanding of the performance and accuracy, an external monitoring system is needed in the future.

| Field test of the robot deployment interface
An RDI prototype was built and tested in a field experiment using Goliath as the UAV platform. The payload for deployment is a dummy OLAM installed on a BladeBug MK II mockup, as shown in Figure 20.
The mockup has a representative size and weighs 2.3 kg. The whole payload weighs approximately 3.5 kg.
Since Goliath had not been integrated with the navigation system, it was controlled manually by a human pilot using a remote controller. Due to difficulties in manually determining the correct landing spot, Goliath was not landed on the blade during deployment. Instead, it was flying in hover closely above the blade surface. The experiment was repeated twice, and in both attempts, the payload was successfully deployed. Figure 21 shows the sequence of actions recorded during one of the attempts.

F I G U R E 19
LiDAR scan from WP2 as defined in Figure 13 (preset at 3.5 m in front and 5 above the wing)

| LHM verification experiment
As discussed in Section 3.4, the LHM is designed with a swing reduction function to reduce the pendulum motion of the hook during the engagement stage. This experiment aims to evaluate both passive and active swing reduction modes implemented on the LHM prototype (see Figure 22) when there is no payload attached to the hook to mimic the engagement scenario. In addition, the experiment was extended to evaluate the ability to stabilize a heavy payload using both passive and active swing reduction.
Four sets of tests were conducted, as detailed in The flight log produced by Goliath's flight computer was retrieved and analysed. As can be seen from Table 7, the mean velocities achieved by Goliath during the test were lower than the target velocity. This is because the UAV had been tuned for a smooth flight with stability prioritized over agility, and therefore, not able to reach the target speed within a short distance.
The longitudinal and lateral displacement of the hook was calculated using forward kinematics and the result is visualized in Figure 24. Table 8 shows the mean and standard deviation of the displacement. In passive swing reduction mode, with a 3.1 kg payload, the pendulum motion was over two times greater in magnitude than it was without a payload. The use of active mode reduces the mean displacement of the hook by approximately 30% (longitudinal) to 34% (lateral) and standard deviation by 23% (longitudinal) to 34% (lateral). Similar reductions can also be observed for the scenario without the payload, although the improvement in lateral displacement and variation is less significant. This is possibly due to the presence of strong lateral gust wind during Test #2, which could also explain that the maximum lateral displacement is slightly greater in Test #2 compared with Test #1 (as 159.9 vs. 155.9 mm).
The power consumption of the hinge was found reasonably low. In active swing reduction mode, with and without payload, the two servos drew 28.22 and 7.27 mA, respectively, at 12 V on average.
6.4 | Field test of the retrieval system An integrated retrieval system involving Goliath, the LHM prototype (see Figure 22), the OLAM prototype (see Figure 25), and the GMP has been built, and a field trial for the retrieval system was conducted. Since the OLAM prototype was manufactured with rapid prototyping methods, including 3D printing and laser cutting Perspex, the strength of its structure is lacking and the BladeBug mockup was not used. In addition, the arms of the OLAM are fixed in the raised position instead of being servo actuated.
As shown in Figure 26, the OLAM prototype was placed on the wind turbine blade and the LHM was attached to Goliath. Both Goliath and the OLAM prototype were fully automated and controlled by the GMP described in Section 5. A safety pilot was present to take over in the event of potential hazards.
The average wind speed during the experiment was 5.0 with 5.8 m/s gust.
A full plan for the whole mission was automatically generated by the GMP based on real-time data supplied by Goliath and the OLAM. These data include OLAM GPS position, LHM status, battery level, Goliath rotors' states, and so on. The trajectory of the UAV and its corresponding action during the test is illustrated in Figure 27. Figure

| CONCLUSIONS
This study looks into the benefit of combining robot retrieval and deployment with autonomous UAVs to perform IMR tasks on offshore wind turbines. These are sophisticated tasks that require particular retrieval and deployment systems, smart navigation around wind blades, and accurate autonomy that manages and guides the whole task from start to end.
A novel deployment and retrieval system is introduced to overcome the challenges in safely deploying and recovering an IMR robot from a wind turbine. A UAV navigation system has been developed to safely and autonomously land the UAV on a wind turbine blade. Mechatronic systems for unmanned deploying and retrieving IMR robots from wind turbine blades have been designed and developed. A complex PDDL model and a powerful autonomy architecture have been developed to underpin the execution of these IMR tasks. A strategy was devised to correct the domain and problem specification by monitoring the discrepancies between the expected and the actual states of the actions over time. A correction to the knowledge base is made after every action representing a current corrected state, a replanning is requested, and an appropriate plan is automatically generated.
Individual modules of the proposed system have been validated in field trials. The results show the UAV was able to safely navigate around the wind turbine blade and successfully located an appropriate landing spot for deploying the IMR robot. The manual IMR robot deployment trial suggested the effectiveness of the developed mechatronic prototype. The proposed retrieval system has been fully integrated with the UAV and the GMP, which was validated in representative scenarios with a mock-up IMR robot and a wind turbine blade. Our experiments show that the proposed approach is effective in retrieving an IMR robot from a wind turbine blade safely and robustly without the need for human interventions.
In future work, the deployment system will be fully integrated and tested to complete the proposed autonomous system. The effect of high-speed wind gusts (over 10 m/s) on the system, which is typical near wind turbine blades must also be further investigated.
The prototypes were constructed using 3D printed components, which limits the strength and hence payload capacity of the system.
Further work is needed in completing the system using more appropriate manufacturing methods and mechanical materials to test a full-size and functional IMR robot.

ACKNOWLEDGMENT
This study has received funding from UKRI through the Innovate UK Grant Agreement No. 104821 and EPSRC Grant EP/R026084/1.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request.