Guidance method of underwater vehicle for rugged seafloor observation in close proximity

Obtaining detailed information about the ocean through seafloor optical imaging is crucial, yet poses significant challenges due to the high attenuation of light, which necessitates the camera to be positioned within several meters of the seafloor. This paper proposes a method for an autonomous underwater vehicle (AUV) to track the seafloor at a close range suitable for optical survey. Through the probabilistic processing of measurements from a mechanical scanning imaging sonar, the AUV can safely follow the seafloor, even in the presence of large obstacles or overhanging geometries. The method was implemented on the low‐cost AUV, HATTORI, and its effectiveness was confirmed by sea experiments around Nishinoshima island, an uninhabited volcanic island located about 1000 km south of Tokyo, Japan. Detailed images of the rugged seafloor were successfully captured under the challenging condition of strong currents.


| INTRODUCTION
The oceans are deeply connected to human life, and surveying the seafloor is essential for various purposes, such as protecting fishery resources, conducting biological surveys, exploring mineral resources, and maintaining artificial structures.Optical imaging is a widely used method for seafloor surveying as it provides detailed information on the color and shape of objects.
However, optical imaging of the seafloor is more challenging than imaging on land due to high light attenuation in the sea.
To capture photographs, the camera must be within several meters of the object.The traditional approach involves divers directly photographing the seafloor.However, this method is restricted to shallow waters, and it exposes divers to potential dangers.
Remotely operated vehicles (ROVs) are another common tool for seafloor optical imaging.They can capture images in deep-sea environments and perform intricate tasks with the help of actuators.
However, ROVs have limited operational areas as they rely on tether cables for communication and power supply, which can be affected by tidal currents and pose entanglement risks.These challenges necessitate skilled operators.
In recent years, autonomous underwater vehicles (AUVs) have been used to capture optical images of the seafloor (Bongiorno et al., 2018;Maki et al., 2011;Noguchi et al., 2022Noguchi et al., , 2020)).They have the potential to efficiently conduct seafloor exploration as they do not require tether cables or skilled pilots.Researchers are conducting studies into the simultaneous operation of multiple AUVs (Matsuda et al., 2022).It is anticipated that through collaborative interactions among AUVs, efficient marine surveys can be conducted.There have been successful instances where AUVs conducted wide-ranging underwater optical imaging without direct support from a ship (Bodenmann et al., 2023).In this research, a long-range capable AUV and satellite communication are utilized to successfully accomplish extended missions without the need for close human supervision or support.There is ongoing research to develop underwater charging station systems that enable the recharging of AUVs without the need for retrieving them to the water surface (Hasaba et al., 2023;Lin et al., 2022).By deploying these systems on the seafloor or attaching them to autonomous surface vehicles (ASVs), long-term operational capabilities for AUVs can be achieved.These research endeavors are poised to transform AUVs into indispensable platforms for conducting efficient marine surveys in the future.
In general, AUVs deployed for seafloor exploration typically survey the seafloor from above using downward-facing sensors.
However, as most of the AUVs photograph the seafloor at a fixed attitude, some rugged areas such as overhangs, are in the occluded area.One way to reduce the occluded area is to use multiple cameras, allowing AUVs to observe in various directions.However, it can be challenging to mount multiple sets of cameras and light sources in different orientations because the light needs to be fixed as far away from the camera as possible to avoid poor visibility in marine snow.This paper, therefore, proposes a method in which an AUV dynamically changes its pitch angle and tracks the seafloor surface while maintaining a constant distance.This method aims to observe steep seafloor areas without occluded areas using a single camera.
Additionally, the proposed method aims to reduce the cost of seafloor optical imaging.The existing methods using AUVs for optical imaging are often expensive due to reliance on high-cost sensors such as inertial navigation systems (INS), Doppler velocity logs (DVL), and multibeam sonar.These sensors are relatively inexpensive, with costs starting in the tens of thousands of dollars for the more affordable options.More costly systems, such as high-end INS, typically run into the hundreds of thousands of dollars.These AUVs are also large, requiring multiple operators and a sizable survey vessel with a crane.Consequently, seafloor optical imaging is costly and not easily accessible.
Our method, in contrast, relies on an inexpensive AUV equipped with commercial off-the-shelf, low-cost sensors, such as a microelectromechanical system (MEMS)-based attitude and heading reference system (AHRS), a water pressure sensor, and a mechanical scanning imaging sonar (MSIS).Among the sensors upon which the proposed method depends, the MSIS stands out as the most expensive.However, even the cost of the MSIS is around a few thousand dollars, which allows for the proposed method to be implemented using relatively inexpensive hardware.The use of small, lightweight AUVs, deployable manually from small vessels, is also anticipated.Additionally, obstacle detection with MSIS enables AUV operation regardless of seawater clarity.Therefore, this method is expected to be applicable not only for optical imaging but also for surveys requiring low-altitude observation of the seafloor using acoustics, such as high-resolution surveys of the volumetric distribution of manganese crusts (Nishida et al., 2016;Thornton et al., 2012).
The proposed method was implemented on our small AUV HATTORI, and the sea experiments confirmed its effectiveness.The experiments were conducted around Nishinoshima island, an uninhabited volcanic island in the Ogasawara Islands of Tokyo.
Detailed optical images of the seafloor, including rugged topography, were successfully captured although the experiments were conducted under strong currents.
The contributions of this study are as follows: • This research introduces a cost-effective approach for seafloor optical imaging using an inexpensive AUV equipped with low-cost, off-the-shelf sensors.A novel method is proposed, allowing the AUV to dynamically adjust its pitch angle to track the seafloor surface, including steep areas, with the aim of reducing occluded areas.
• The study presents a simplified and generalizable approach for safe path planning and following.By leveraging an occupancy grid map, basic image processing techniques, and the nonlinear guidance law (NLGL), the AUV can navigate smoothly through complex seafloor terrain, including large obstacles.
• The effectiveness of the proposed method has been confirmed through field experiments.Despite the challenging environment with rugged seafloor terrain, the AUV successfully generated detailed three-dimensional (3D) images of cliff sections using this method.The results from 27 missions have been demonstrated, underlining the practical applicability and robustness of the proposed approach in real-world scenarios.
The paper is organized as follows: Sections 2 and 3 describe the related works and the proposed method, respectively.Section 4 introduces AUV HATTORI.The experiments are detailed in Section 5, and the conclusion is presented in Section 6.

| RELATED WORK
In recent years, substantial research has been dedicated to developing affordable AUV systems for seafloor optical imaging.One strategy for reducing AUV system costs involves utilizing optical cameras to estimate the AUV's pose and interpret the surrounding environment (Arain et al., 2020;Dunbabin et al., 2019;Manderson et al., 2018).This approach includes technologies such as visual simultaneous localization and mapping (V-SLAM), where an optical camera is utilized to capture the surroundings, extract feature points in the image, and simultaneously estimate the AUV's pose and map the environment.Optical cameras are more affordable than acoustic sensors like sonars and can provide detailed information in the highly transparent sea.They are used for AUVs required to operate with high precision near the seafloor or structures, including intervention or docking to charging stations (Lin et al., 2022;Maki, Sato, et al., 2018;Nishida et al., 2019).
Concurrently, research has been conducted on reducing the cost of AUVs using acoustic sensors.While generally more expensive than NOGUCHI ET AL.
| 315 optical cameras, acoustic sensors offer the advantage of providing information over a wider area, even in turbid seawater.Some methods for AUVs to observe the seafloor at close range based on a low-cost MSIS have been proposed (McPhail et al., 2010;Schillai et al., 2019).These methods use an MSIS to detect obstacles on the vertical plane ahead of the AUV, enabling it to track the seafloor by adjusting its depth accordingly.
Previously, a method was developed that allows an AUV equipped with an MSIS and capable of maintaining an arbitrary attitude to track omnidirectional surfaces (Noguchi & Maki, 2021).This method is an extension of an artificial potential field (APF) algorithm, which calculates the attitude and surge speed references by introducing an APF around obstacles (Maki, Noguchi, et al., 2018;Warren, 1990).The APF, expressed as the sum of the attractive and repulsive potentials calculated based on a forward occupancy grid map (Thrun, 2002), guides the AUV towards the seafloor (attractive potential) and away from obstacles (repulsive potential).The AUV can evade large obstacles by altering the direction of the attractive potential, particularly when it falls into a minimum of a potential field.
However, this method has drawbacks; the AUV must reverse in the face of large obstacles, and implementation can be complex.
In this paper, we propose a novel method that enables an AUV, equipped with an MSIS and capable of maintaining an arbitrary attitude, to track a rugged seafloor.The proposed method uses an occupancy grid map and basic image processing techniques on the forward vertical plane for safe path planning and employs the NLGL (Park et al., 2007;Sujit et al., 2014) for path following.This method enables the AUV to smoothly track the seafloor, even in the presence of large obstacles without having to reverse.Moreover, the implementation process is simplified through the use of general programming libraries, due to its reliance on image-processing techniques for path generation.

| Outline
In this paper, the North-East-Down coordinate system and the forward-right-down coordinate system are used as the world coordinate system and the vehicle coordinate system, respectively.
* ˆindicates the reference value of * .
In planning missions for AUV systems for seafloor exploration, it is typical to establish linear transects on a horizontal plane in advance.Some strategies involve mapping out a route that repeats in a zigzag pattern along the straight survey line (Noguchi et al., 2022).
Another prevalent method uses a pattern referred to as "lawnmowing," characterized by some parallel survey lines (Jalving et al., 2003).The proposed method is a strategy designed for vertical planes, intended to integrate with predetermined linear mission plans following a fixed heading.
Figure 1 provides a visual representation of the AUV's movement, as guided by the proposed method.In the vertical plane (C-Z), with a reference heading angle ψ ˆ, the AUV conducts its path planning.This method aims for the AUV to follow a path that maintains a certain distance from the seafloor, ensuring that the AUV's bottom-mounted downward-looking camera is oriented perpendicularly toward them.Note, in this method, obstacles such as large rocks and structures on the seafloor are not differentiated from the seafloor itself and are considered as observation targets just like the seafloor.The position can be estimated using the dead reckoning method based on the surge velocity, attitude, and depth.

| Map maker
The Map Maker creates and updates the occupancy grid map on the vertical plane (C-Z) based on MSIS measurements (Noguchi & Maki, 2021).The space on the plane is divided into cells with a side length of a cell , as shown in Figure 3, and the probability of the seafloor presence in each cell is calculated using a binary Bayes filter (Thrun, 2002).It is important to update the map using stochastic means because the measurements of the MSIS are limited and noisy.
Note that the area to update the occupancy probability is narrowed down to the area around the AUV (n n × C Z ) according to the AUV's computing power, and the area also translates together with the AUV, as shown in Figure 3.
The presence of the seafloor in cell i is denoted by m i , and the probability of the seafloor in cell i is denoted by p m ( ) i .Here, the occupancy probability is updated in the form of log odds to improve the computation efficiency.The log odds l t i , of cell i at a certain time t is expressed as follows: where z represents MSIS observations.The probability can be calculated from the log odds as follows: The log odds is updated by a binary Bayes filter as follows: where and p m ( ) i 0, denotes the prior probability.In many cases, l i 0, can be neglected in the calculation because of the lack of the prior information (i.e., p m ( ) = 0.5 i 0, ) as follows: The algorithm needs the inverse measurement model  p m z ( ) i t , and it is calculated by MSIS observations in the proposed method as follows (Noguchi & Maki, 2021): where prevent the AUV from misrecognizing the water surface as the seafloor.

| Tracking navigator
Based on the occupancy grid map updated by the Map Maker, the Tracking Navigator calculates the reference attitude and reference surge velocity of the AUV, as shown in Figure 4.The area to update the occupancy probability is narrowed down to the area around the AUV as in the Map Maker.
First, a path is planned to track the seafloor and structures while maintaining a certain distance from them.In the path planning, it is assumed that the occupancy grid map is regarded as an image, and computation is performed using image processing libraries.A binarised image of the occupancy grid map is created by dividing it into cells with an occupancy probability above a threshold p thresh .If no cell here has an occupancy probability of p thresh or more, the AUV assumes that there is no seafloor or structure to track in the vicinity and dives vertically toward the seafloor.The reference attitude ∈ θ π π ˆ[− , ] on the vertical plane (see Figure 4) and the reference surge velocity v ˆare calculated as follows: If a cell has an occupancy probability greater than p thresh , the cell is dilated using a circular kernel whose radius is the reference observation distance d ˆmea .Then, the clockwise contour of the dilated cells is extracted as the path to be followed by the AUV.If it gets more than one contour, the contour with the closest distance from the AUV shall be selected.
The virtual target point (VTP) is then determined to track the determined path.Here, the VTP is determined using the NLGL concept (Park et al., 2007;Sujit et al., 2014) to stably track a path of arbitrary geometry.The VTP is the intersection of the path with a circle of radius r d (<2 ˆ) for mea centered on the AUV, as shown in Figure 4.However, if the distance between the AUV and the path is longer than r for and no intersection exists, the point closest to the path is the VTP.
Then, θ ˆand v ˆare calculated as follows: where (c z , auv auv ) and (c z , vtp vtp ) are the coordinates of the AUV and the VTP on the vertical plane (C Z -), respectively, and Finally, the reference 3D rotation matrix R ˆis calculated using θ ˆas follows: where the AUV may deviate from the path significantly.Therefore, the attitude difference α is calculated as follows, and if α is greater than the constant α thresh , the AUV stops moving forward and waits until the AUV's attitude approaches the reference attitude.
F I G U R E 4 Process flow of the Tracking Navigator.
where tr(*) represents the trace of the matrix * .

| AUV HATTORI
This method has been implemented in AUV HATTORI (Maki, Noguchi, et al., 2018).Figure 5 and Table 1 show the vehicle's appearance and specification table, respectively.The vehicle is smaller, lighter, and easier to deploy than conventional AUVs for seafloor optical imaging.
The vehicle, equipped with two vertical thrusters at the bow, is capable of roll and pitch motion.The horizontal thruster located at the stern of the vehicle facilitates surge motion and yaw motion.
However, it lacks the ability to perform sway motion and heave motion.The sound wave frequency of the nose-mounted MSIS is 675 kHz, and the beam width is ∘ ∘ 3 × 30 .Table 2 shows the MSIS settings.The MSIS was set to scan a range with an elevation angle of π π − ∕4 ~∕4rad relative to the vehicle.Each round-trip scan took approximately 5 s.The MSIS is set to yield 400 bins of reflection intensity values.To reduce computational load, we condense these bins to 40.This reduction involves substituting the reflection intensities of 10 consecutive bins with the maximum value among them, aiming to prevent any potential oversight of obstacles.The attitude of the vehicle is estimated using the complementary filter (Valenti et al., 2015) based on the inexpensive nine-degree-offreedom (9-DOF) sensor readings (i.e., three-axis accelerometer, three-axis gyro, and three-axis magnetometer).It has a DVL to measure ground speed, but when the ground speed cannot be measured by the DVL, the surge velocity v is estimated from the reference value of thrust N ˆsurge based on a model of the fluid resistance of the vehicle as follows: The software is implemented using the robot operating system (ROS), specifically the Noetic distribution.The vehicle has an altimeter and an ultra-short baseline (USBL), but note that these are used for monitoring and evaluating the dive, not for controlling the vehicle in this paper.

| Objective
We conducted demonstration tests of the proposed method in the sea area surrounding Nishinoshima Island (Figure 6a), an uninhabited volcanic F I G U R E 5 AUV HATTORI.| 319 island located in the Ogasawara Islands, Tokyo, Japan (Noguchi et al., 2021).Spanning 4 days, from July 9 to 12, 2021, the survey aimed to capture optical images of the seafloor near the island.Although it was known that the terrestrial part of the island was engulfed in lava and volcanic ash due to a large-scale eruption in 2020, the underwater section remained unexplored.Therefore, it was unclear how far the lava and ash had spread underwater and what the condition of marine life, including fish and benthic organisms, was.

| Procedure
HATTORI was deployed manually from a research vessel and initiated the proposed method while on the sea surface.After a designated period, it ascended and returned to the sea surface.In the study area, the tidal currents were fast (approximately 0.5 m/s) and changed direction frequently.For safety reasons, HATTORI was set to always travel in the same direction during the same dive (in principle, the direction of the current on the surface at the commencement of the observation).In addition, as HATTORI did not have a ballast releaser, HATTORI's buoyancy was slightly positive so that HATTORI could float to the surface if it stopped working during a dive due to malfunction, and so on.
Table 3 shows the parameters of the proposed method.d ˆmea was set between 1 and 2.5 m for each dive.When HATTORI's ground speed could not be measured by the DVL, a Gaussian filter with a standard deviation of 0.4 m was applied to the occupancy grid map in the horizontal direction once every 1 s.Due to the depth limitation of the onboard sensors, the occupancy probability of cells in a region deeper than a depth of 60 m was overwritten to 1.

| Results
Over the course of the 4-day survey, HATTORI executed 27 dives, 13 of which were carried out untethered. Figure 6b shows HATTORI during seafloor tracking (Maki Lab., 2018).The cumulative dive time reached 203 min, covering a total dive distance of 4.7 km (with a horizontal distance of 3.3 km).Table 4 provides details for each dive, including date, the presence or absence and type of tether, reference

| Dive 15
The duration, maximum depth, dive distance, and horizontal dive distance were 960 s, 48.1, 376, and 285 m, respectively.The depth and attitude of HATTORI with respect to dive time are shown in    (Zuiderveld, 2003).
buoyancy or adding a vertical thruster to control the speed in the heave direction (i.e., the z-direction in the vehicle coordinate system) to 0. It may also be effective in reducing the effect of external forces by increasing k v to increase the thrust in the forward direction.

| Dive 21
Dive 21, conducted on the north side of the island, is described.The duration, maximum depth, dive distance, horizontal dive distance, d ˆmea and median d mea were 1888 s, 58.7, 831, 718, 2.0, and 3.1 m, respectively.
HATTORI identified the school of fish as the seafloor, attempting to navigate over it for observation.As a result, the d mea was higher than the d ˆmea .During the experiments, HATTORI was often followed by schools of fish and sharks.For example, Figure 12  The proposed method was effective to track the seafloor without DVL measurements if the error of the velocity control was insignificant.HATTORI succeeded in tracking the seafloor without a collision in Dive 21, although the DVL malfunctioned during the dive.Figure 14 shows the occupancy grid map, HATTORI's trajectory, and VTP path before and after the ground speed became unmeasurable.The purple line indicates the trajectory when the ground speed cannot be measured.Before the ground speed became unmeasurable, HATTORI lowered its nose and moved forward at a higher altitude than the VTP path as in Dive 15 (see Figure 9c).
Although it avoided hitting the seafloor, HATTORI temporarily overshot the VTP path.When HATTORI could no longer measure ground speed at 508 s, it temporarily and rapidly reduced its altitude to maintain a horizontal attitude with respect to the seafloor, leading to a temporary overshooting of the VTP path.This was because of the parameter setting by the on-the-spot judgment.Specifically, we had applied the rule that HATTORI forced the surge thruster output to 8 N or more when ground speed could not be measured and forward speed control was in open-loop control.The setting was implemented because HATTORI was positively buoyant and there was a concern that it could not descend from the water surface.
During the vertical descend immediately after the start of a mission,

| CONCLUSION
This paper proposed a method for AUVs to track the seafloor at close range.This method was developed to allow low-cost AUVs to safely observe optical images of the seafloor with large obstructions and overhanging terrain.The method was implemented on the low-cost AUV HATTORI and its effectiveness was confirmed by sea experiments around Nishinoshima island, an uninhabited island in the Ogasawara Islands.Although the survey was conducted under strong currents, detailed optical images of the seafloor, including rugged topography, were successfully captured.
In this study, successful seafloor optical imaging missions were conducted under the assumption that the heading angle was consistently maintained throughout each mission.Going forward, we aim to successfully conduct missions composed of various survey lines, such as zigzag patterns or lawn-mowing patterns, using the method proposed in this paper.This constitutes a crucial step towards areal observations of expansive seafloor areas.In addition, we plan to explore the use of reinforcement learning and similar methods to automatically optimize the several parameters inherent in our method and verify their effectiveness.As a part of this plan, we also intend to dynamically fine-tune the reference surge speed, which in this method, was limited to either maximum speed or a complete stop.
The proposed method consists of two parts: Map Maker and Tracking Navigator.The structure of the method is shown in Figure2.The Map Maker creates an occupancy grid map on the forward vertical plane based on the MSIS measurements and AUV's pose (i.e., position and attitude) on the vertical plane.The Tracking Navigator calculates the reference attitude and surge velocity based on the occupancy grid map and AUV's pose on the vertical plane.The AUV is assumed to have the following three features.First, the AUV has actuators that allow forward motion, rolling, pitching, and yawing.Second, the AUV has a forward-looking MSIS to observe the vertical plane, as shown in Figure 1.Third, the AUV has the ability to measure its surge velocity, attitude, and position on the vertical plane in real-time.The AUV can use low-cost sensors to estimate the surge velocity, attitude, and position.The surge velocity can be estimated based on the thrust output and a dynamic model of the AUV.The attitude can be estimated based on an MEMS-based AHRS.F I G U R E 1 Visual representation of the AUV's movement guided by the proposed method.F I G U R E 2 Structure of the proposed method.
r denotes the distance from the MSIS and m r denotes the cells observed by the MSIS at distance r. S r ( ) denotes the reflection intensity at distance r.As MSIS observations generally contain more noise at near and far distances, only data of distances between r min and r max are used to update the occupancy probability.The function f r ( ) 1 indicates the hypothesis that a strong reflection should correspond to the presence of the seafloor.The constant S thresh is the threshold value where  p m z ( )=0.5 i t .In addition, a higher intensity value makes the occupancy probabilities of the farther cells higher.It is to estimate the occupancy probability in the shadowed area higher.When the AUV tracks the seafloor at a low altitude, obstacles can create large shadows.Since the MSIS mechanically rotates the direction of sound emission, it takes time to update the occupancy probability in the shadowed area.Therefore, the occupancy probability of the area is estimated to be higher, reducing the risk of colliding with obstacles in the area.The function f r ( ) 2 has the constants k r min and k rmax (∈[0, 1]) and indicates the hypothesis that the measurements at farther distances are less reliable (i.e., method, a Gaussian filter in the horizontal direction is applied to the occupancy grid map to deal with errors in the AUV's horizontal position estimation.If necessary, the occupancy probabilities on the map can be forced to be overwritten to enable more efficient diving.For example, depth limitation can be applied by overwriting the occupancy probability of cells in a region deeper than a certain depth to 1.In addition, by overwriting the occupancy probability of cells near the water surface with 0, it is possible to F I G U R E 3 Computational domain moving with the AUV.The area indicated by the light blue grid shows the area where the update calculation is performed by the proposed method. the reference attitude R ˆand the current attitude R (a 3D rotation matrix) are very different and the AUV moves forward, of seawater, the frontal projected area, and the drag coefficient, respectively.

Fs
I G U R E 6 (a) Nishinoshima island.(b) HATTORI during the experiment.(c) HATTORI's path.The bathymetric chart was made for reference from the data obtained by KAIYO MARU No. 3 during the comprehensive academic survey of Nishinoshima island by the Ministry of the Environment, Government of Japan.The dive numbers are written as D*.T A B L E 3 Parameters of the proposed method.thresh was 80 during Dives 1-6.observation distance d ˆmea , duration, maximum depth, and median observation distance (median d mea ).The longest duration per dive was 1888 s in Dive 21.The maximum depth of HATTORI was 58.7 m in Dive 21.The DVL malfunctioned during Dive 21, and HATTORI collided with the seafloor once during Dive 22.

Figure
Figure6cshows HATTORI's trajectory measured by the USBL of the small ASV BUTTORI(Horimoto et al., 2021).The zigzag patterns seen in Dives 6, 14, 15, 21, and so on are due to errors in acoustic positioning.Dives 1-5, 9, 17, and 23-27 are estimated by dead reckoning based on the integration of HATTORI's velocity measurements.Dive 19 has been excluded because it was carried out to verify the emergency surfacing command.

Figure 7
Figure 7 shows images of swimming creatures and the seafloor taken by HATTORI.The camera successfully captured various underwater scenes, including sharks and fumes.The performance of the proposed method is discussed by detailing typical dives, Dives 15 and 21.

| 321 F
I G U R E 7 Photos of nektons and the seafloor photographed by HATTORI.The dive numbers are written as D*.F I G U R E 8 Depth and attitude of HATTORI in Dive 15.

Figure 8 .
Figure 8. HATTORI's trajectory during seafloor tracking and the occupancy grid map are also shown in Figure 9a.Although the seafloor was rugged, with height differences of several meters, HATTORI successfully tracked the seafloor without a collision.
shows a photo taken by the forward-looking camera at the dive time of 822 s in Dive 21.The school of fish continued to rotate around HATTORI for a while.The occupancy grid map and HATTORI's trajectory are shown in Figure 13.As the proposed method determines the occupancy probability based on the MSIS reflection intensity, HATTORI recognized the school of fish as the seafloor and kept ascending to avoid it.
the DVL often failed to measure ground speed, and open-loop control was used.The setting was added to increase the surge thrust F I G U R E 11 HATTORI's trajectory during cliff tracking in Dive 15.F I G U R E 12 Photo taken by the forward-looking camera at the dive time of 822 s in Dive 21.F I G U R E 13 HATTORI's trajectory and occupancy grid map before and after the Dive 21 fish school was observed.The orange line indicates HATTORI's trajectory.Orange arrows indicate the position and attitude of HATTORI every 10 s. so that HATTORI could descend smoothly even with positive buoyancy as a temporary measure.This resulted in rapid acceleration and rapid loss of altitude when ground speed could not be measured.In Dive 22, conducted immediately after Dive 21, the setting resulted in the same overshoot as in Dive 21, and HATTORI collided with the seafloor.For dives after Dive 22, the setting to increase surge thrust that caused the collision was removed.In a subsequent dive (Dive 26), the DVL continued to malfunction; however, HATTORI did not collide again and observed the seafloor safely.It is assumed that open-loop control with an accurate model could have avoided the collision in Dive 22.

F
I G U R E 14 HATTORI's trajectory and occupancy grid map before and after ground speed could no longer be measured at Dive 21.The orange line indicates HATTORI's trajectory.The purple line shows the trajectory when ground speed cannot be measured.Orange arrows indicate the position and attitude of HATTORI every 10 s.The blue line shows the VTP path of the proposed method.Green dots indicate the seafloor surface estimated based on altimeter observations.
T A B L E 1 Specification of AUV HATTORI.
T A B L E 4 Dive table of the experiment at Nishinoshima island.It shows the date of each dive, the presence or absence and type of tether, reference observation distance d ˆmea , duration, maximum depth of HATTORI, and median observation distance (median d mea ).The median altitude in Dive 22 is not included due to HATTORI making prolonged contact with the seafloor.