Anais do XX Congresso Brasileiro de Automática Belo Horizonte, MG, 20 a 24 de Setembro de 2014

DORIS - A MOBILE ROBOT FOR INSPECTION AND MONITORING OF OFFSHORE FACILITIES Mauricio Galassi∗, Anders Røyrøy†, Guilherme P.S. de Carvalho‡, Gustavo M. Freitas‡, P˚ al J. From§, Ramon R. Costa‡, Fernando Lizarralde‡, Liu Hsu‡, Gustavo H.F. de Carvalho‡, Jose F.L. de Oliveira‡, Amaro A. de Lima¶, Thiago de M. Prego¶, Sergio L. Netto‡, Eduardo A.B. da Silva‡ ∗

Petrobras/CENPES - Research and Development Center †

TPD RD New Development Solutions, Statoil ASA ‡

§

Dept. of Electrical Eng. - COPPE/UFRJ

Dept. of Mathematical Sciences and Technology - Norwegian University of Life Sciences ¶

Dept. of Telecommunications - CEFET/RJ

Emails: [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected] Abstract— DORIS is a research project which endeavors to design and implement a mobile robot for remote supervision, diagnosis, and data acquisition on offshore facilities. The proposed system is composed of a railguided robot capable of carrying different sensors through the inspected area. This paper presents a general overview of the robot and a description of the developed mechanical designs and signal processing algorithms. Initial results validate the mechanical concepts considered so far and indicate that the signal processing algorithms are capable of detecting, in real time, multiple foreign objects and audio anomalies from a standard scenario. Keywords—

Mobile robots; Field robotics; Security and safety of HMS.

Resumo— DORIS ´ e um projeto de pesquisa que se empenha em implementar um robˆ o m´ ovel para supervis˜ ao remota, diagn´ ostico, e aquisi¸c˜ ao de dados em instala¸c˜ oes offshore. O sistema proposto ´ e composto de um robˆ o guiado por um trilho e capaz de levar diferentes sensores atrav´ es do ambiente inspecionado. Esse artigo apresenta uma vis˜ ao geral do robˆ o e uma descri¸c˜ ao do projeto mecˆ anico e dos algoritmos de processamento de sinais desenvolvidos. Resultados iniciais validam os conceitos mecˆ anicos considerados at´ e ent˜ ao e indicam que os algoritmos de processamento de sinais s˜ ao capazes de detectar, em tempo real, m´ ultiplos objetos abandonados e anomalias de ´ audio nos sinais adquiridos em um cen´ ario padr˜ ao. Palavras-chave—

1

Rob´ otica M´ ovel; Rob´ otica de Campo.

Introduction

considered the use of robots in Oil & Gas facilities in operations that require both high precision and strength, regardless of weather conditions.

Safety and efficient operation are imperative factors to offshore production sites and a main concern to all Oil & Gas companies. A promising solution to improve both safety and efficiency is to increase the level of automation on the platforms by introducing robotic systems. During the last decade, several Oil & Gas companies, research groups, and academic communities have shown an increasing interest in the use of robots for operation on offshore facilities. Recent studies project a substantial decrease in the level of human operation and an increase in automation on future offshore oil fields (Skourup and Pretlove, 2009). The studies also point out the potential increase in efficiency and productivity with robot operators, besides the improvement of Health, Safety, and Environment (HSE) conditions, as robots can replace humans in tasks performed in unhealthy, hazardous, and confined areas (From, 2010). In (Anisi et al., 2010), it is

Among the research groups interested in offshore robotics, Fraunhofer IPA is pioneer in proposing and demonstrating the applicability of mobile robots for offshore inspection and maintenance tasks in loco (Bengel et al., 2009). One example is MIMROex (Bengel and Pfeiffer, 2007), capable of navigating safely, building maps, and executing inspection tasks autonomously throughout the topside of platforms. Another robotic device applied in offshore environments is Sensabot (NREC/CMU, 2012), capable of safely inspect and monitor hazardous and remote production facilities. The robot can sustain high temperatures, is able to reach areas with difficult access, and is certified to operate in explosive and toxic environments. SINTEF-ICT is another group interested in manipulators applied to the oil and gas industry. Inspection and maintenance operations in a simu-

3174

Anais do XX Congresso Brasileiro de Automática Belo Horizonte, MG, 20 a 24 de Setembro de 2014

lated production process are performed by the cooperation of a gantry-mounted manipulator and a floor-mounted robot (Kyrkjebø et al., 2009). In this paper, we describe the DORIS project, which aims to develop a mobile robot to perform monitoring, inspection, and simple intervention tasks in an offshore platform. To this end, the system must be able to move throughout the monitored environment carrying different sensors, analyzing sensor data in loco or storing it for a posterior analysis, and interpreting the results. The sensors can identify abnormalities such as intruders in restricted areas, abandoned objects, smoke, fire, and liquid and gas leakages. Furthermore, the robot is able to make machinery diagnosis, read instruments, and perform interventions on valves and other equipment using an embedded manipulator. In the following sections, we present an overview of the DORIS project with particular focus on the mechanical designs and the signal processing algorithms. Preliminary results with the prototypes tested so far validate the considered mechanical concepts, and the capability of the signal processing algorithms to detect, in real-time, multiple abandoned objects and audio anomalies in the recorded audio signals of a noisy background. 2

(a) Robot’s operational scenario in a production plant

(b) Detailed zoom of the robot

Figure 1: Illustration of the DORIS robot operating in a production plant. lithium-ion batteries, which have a small size and a high energy capacity. Four batteries are used to power the motors and two to power the other electronics components. It is essential to monitor the batteries’ behavior so that faults can be avoided. The power management interface is implemented through System Management Bus (SMBus) connections, allowing the electronics system to receive all possible information about each battery state. The main objective of the software subsystem is to allow the implementation of high- and lowlevel control of the robot. The tools used to develop DORIS software architecture must consider two important factors: they have to be commercially available, and provide modular functionalities. These requirements led to the adoption of Qt as the graphical interface framework, Robot Operating System (ROS) as the communication middleware (Quigley et al., 2009), and Ubuntu as the operating system. The software provides autonomous control (programmed tasks) and remote control through a Graphical User Interface (GUI) in the Host Control Base (HCB) computer. The HCB is composed of a set of processes running in parallel denominated ROS nodes, which can communicate with each other. To deal with this environment, a new software architecture called Robot Package Software is proposed, dividing the software into tools (graphical windows) and components (processing and communication unities), and grouping them into a dynamic library.

General Overview and Main Challenges

The proposed system is composed of a robot with cameras, microphones, gas, vibration and temperature sensors, and a manipulator arm. The robotic device is guided by a rail and both the robot and the rail follows a modularity concept. Additional robot modules can be annexed to include other sensors, and the rail track can be modified by adding or replacing rail segments, thus enabling operation in different areas of the platform. The robot will be controlled autonomously or by teleoperation. Task managing can be either in automatic (programmed using a mission interface) or manual mode (real-time remote operation). The teleoperation and monitoring capabilities guarantee online access to the embedded sensors, providing information about the surrounding environment and the robot operating conditions with real-time processing. Figure 1 illustrates the operation in a production plant. The DORIS project can be divided into five subsystems: electronics, power supply, software, mechanics and signal processing. The electronics subsystem is responsible for providing embedded computational support for the robot control, signal processing, task managing, and local and remote communication. The device motion is controlled through drivers that can receive position, velocity, or current setpoints. The power supply system uses military-class

3175

Anais do XX Congresso Brasileiro de Automática Belo Horizonte, MG, 20 a 24 de Setembro de 2014

Considering the robot functionalities and the aggressive offshore environment, several challenges should be addressed. Temperatures in offshore facilities can vary between −30◦ C to 50◦ C, relative humidity can reach 100%, and there may be splash water, salty air, storms, and high extensive corrosion (Graf and Pfeiffer, 2007). Concerning robustness and safety required to operate in classified areas, the robot must be sealed against water and objects, resistant to a wide temperature range, protected from impact and vibration, electrically shielded to avoid explosion by ignition, and equipped with a monitoring system. Another challenge is that the embedded computers must run heavy signal processing algorithms, requiring high computational power. However, the power supply subsystem must efficiently provide power and maintain a low level of power consumption. Further complications arise because the system is designed to move in confined areas and have efficient wireless communication with operators, providing online information of sensors data. Finally, the robot must have a modular and flexible design, employing plug and play extensions. 3

lows for path modification. The track corresponds to a closed circuit, allowing the robot to perform periodic inspection and monitoring tasks. The main objectives of the mechanical project are to design the rail, the traction and passive modules, and the joints used to couple them. The design must allow the robot to move smoothly in a 3D space and to make full stops anywhere on the rail. Considering the severe corrosion and weather conditions in offshore environments, the choice of materials are imperative to the success of the project and certified solutions must be considered. The robot is composed of four modules at its default configuration, but it is conceived so that other modules can be added. The total weight of this configuration is estimated at 50 kg and we expect to have a maximum speed of 1m/s. The first adopted concept considers a tubular rail with an attached rack. The idea is inspired by R the Thyssenkrupp Flow II stairlift. Traction is provided by conical wheels supported on the tube, and auxiliary mechanisms with springs improves stability. The joint to couple two modules resembles to a spine, being composed of multiple disks guided by steel cables with springs attached to its ends, which turn the joint flexible. This design is illustrated in Fig. 2.

Mechanical Design

The DORIS robot must move in a 3D space performing horizontal, vertical, and curved movements. Thus, the robot’s mechanical system must be flexible and able to keep its orientation stable. It also has to avoid sliding and move relatively fast, in case of emergency situations. The robotics literature shows that guided robots are the most suitable motion concept for DORIS. Versatrax Vertical Crawler uses three rubber tracks to move inside a pipe (Inuktun, 2014). POBOT (Fauroux and Morillon, 2010) and Pruning Climbing Robot (Kawasaki et al., 2008) are capable of moving on vertical structures using a self-locking property to keep the position using friction between the wheels and the rail. UT-PCR (Baghani et al., 2005) is a light-weight robot that moves vertically on a rail with ordinary wheels being pressed against the rail by springs. ARTIS (Christensen et al., 2011), developed by DFKI Robotics Innovation Center, is a modular rail guided robot that moves on a rectangular cross sectioned rail and performs inspection and maintenance tasks in ballast water tanks. We propose a tubular rail for DORIS locomotion. The use of a pre-specified path reduces concerns as localization and obstacle avoidance and allows the robot to move relatively fast through its workspace. Motion is simple, as the robot has only one degree of freedom (DoF). The use of a rail limits the robot workspace, but it may be installed to pass through key areas and its modularity al-

Figure 2: Design of the first concept, considering a rail with an attached rack. The main advantage of this design is the absence of sliding due to the use of a rack and pinion mechanism. However, the rack has a complex geometry, which is difficult to machine, limits the robot speed and has low efficiency. Therefore, the following premise was adopted for further designs: the rail must be designed to be as simple as possible, leaving the complexity to the robot. This is also motivated by the fact that the rail may be long so that its cost should be kept to a minimum. The following designs incorporate the use of gimbals with wheels as guides for the module on the rail. Two gimbals, one coupled to the other with orthogonal pivot axes, are mounted on the module’s base, providing pitch and yaw rotations.

3176

Anais do XX Congresso Brasileiro de Automática Belo Horizonte, MG, 20 a 24 de Setembro de 2014

The internal gimbal comprises four equally spaced wheels that closely encompass the rail. The second concept uses two sets of gimbals and a centralized traction system, composed of two groove wheels mounted on a prismatic base that can slide horizontally and vertically on guide bars through linear bearings. This 2DoF prismatic mechanism is necessary to compensate for rail curvatures. A clamping system is designed to press the traction wheels against the rail, applying adjustable radial forces to compensate the robot’s weight. The passive modules comprises only a base and the two sets of gimbals only. Double cardan joints are designed to couple two modules, as depicted in Fig. 3.

and pitch (fundamental frequency) signatures using a single or a array of microphones. • Vibration analysis: Use of acceleration sensors to diagnose the operation mode of rotating machines, performing possible fault classification, such as misalignment and unbalancing operation. • Gas sensor: detection of gas leakages. • 3D mapping: environment 3D modeling using a laser sensor. The main idea of all these signal processing features is to make the robot perform an initial reference lap around the closed rail track, being manually validated by a system operator. In the subsequent laps, all signal processing algorithms compare the newly acquired signals with the reference data to detect any form of anomaly, as indicated above. Once an anomalous behavior is detected, an alarm is flagged to the system, which stores all associated data for immediate or future diagnosis, as represented in Fig. 4.

Video Processing Algorithms

Figure 3: Design with gimbals and two traction wheels mounted on a 2DoF prismatic base.

Measurement Processing Algorithms

Events Alarm Event Processor

Events

A prototype based on this design was built to validate the considered concepts. The tests’ results, which are presented in Section 5, show that the use of gimbals is an proper choice concerning stability, guidance, and support. Furthermore, it is possible to have a smooth vertical motion applying radial forces by the clamping mechanism. An important advantage of this design is the simplicity of the rail. However, the prismatic mechanism can lock in some situations, which is not ideal. Moreover, this model has a high weight (the traction module alone is estimated to weight 20kg) and the clamping mechanism is complex. A test was set up to analyze the behavior of polyurethane wheels and the results show that polyurethane is an appropriate material to provide grip. 4

Audio Processing Algorithms

Actions

Events

Actions Storage

Operator

Positioning Actions

Sensors

Figure 4: Diagram of signal processing capabilities incorporated to the DORIS robotic platform. 4.1

Video Signal Processing

The initial goal of the video processing techniques is to identify abandoned objects in the proposed scenarios. To do so, a reference video, without abandoned objects, must be properly compared to a target video, which possibly contains abandoned objects. For this comparison to be effective, the videos must be precisely synchronized. Below follows a more detailed description of our abandoned object detection method:

Signal Processing Algorithms

The following signal processing capabilities are devised for the DORIS robotic platform:

• Initial Video Alignment: To perform the initial video alignment, a maximum likelihood approach, based on the videos motion data and a motion model for the robot, is employed. First, the homographies between the consecutive frames of the given video sequence are calculated, and from them, the translational motion of the camera is extracted. By integrating the horizontal component of the camera motion along

• Video: use of multiple cameras (visible-light, infrared, fisheye and stereo) to detect video anomalies such as abandoned objects, smoke, fire, liquid leakage, and intruders. • Audio: detection of audio anomalies of impulsive nature, such as an explosion or the diagnosis of rotating machines based on energy

3177

Anais do XX Congresso Brasileiro de Automática Belo Horizonte, MG, 20 a 24 de Setembro de 2014

the track, the horizontal camera displacement is obtained as a function of the frame number. The obtained curve is noisy due to camera vibration, but one can obtain a noiseless motion model by performing the least-squares fitting of a piece-wise linear model composed of two straight lines of opposite angular coefficients. In this way, we obtain a template for the DORIS robot movement. By performing a matching between the movement template and the noisy curve being produced by the robot in the target video sequence, the video synchronization is obtained.

of being part of an abandoned object, in sequential frames, is larger than a given threshold, empirically set. It must be noted that, in order to correctly align the images to be compared in the temporal filtering and voting steps, homographies must be calculated between the used frames. 4.2

Audio Signal Processing

The main goal of the audio signal processing block is the detection of audio events in an acoustically adverse environment. Among the possible audio anomalies to be detected, we consider impulsive events, such as an explosion or any other abnormal background noise, and the machine monitoring through energy and pitch tracking. The main challenges for achieving such goals include high reverberation level in case of enclosed spaces, and significant background noise of possibly non-stationarity nature.

• Geometric Registration Between Frames: Considering that the target and reference video sequences have been properly aligned in time, the speeded-up robust feature (SURF) algorithm is employed to identify the points of interest (PoI) on two corresponding frames of both video sequences (Bay et al., 2008). In the following, a correspondence is determined in a point-by-point level among the two PoI sets previously identified, first by eliminating the ones that greatly deviate from the translational movement restriction, and finally by using the random sample consensus (RANSAC) algorithm (Kong et al., 2010) (Hartley and Zisserman, 2003). Based on these point correspondences, an homography (Kong et al., 2010) (Hartley and Zisserman, 2003) is computed on the reference frame to allow a proper comparison with the corresponding frame of the target video.

5

Prototypes and Preliminary Results

Two prototypes have been built to test mechanical and signal processing concepts. The first one is based on the Roomba robot and was developed to test video and audio anomalies detection employing signal processing techniques. The second protoype, DoriAna, was built to test the proposed mechanical concepts and traction system. The real scale prototype, made with low cost materials, was tested in horizontal and vertical motion on a rail composed of straight and curved modules.

• Image Comparison: As the simple subtraction between the registered frames does not work due to the excessive amount of details in the cluttered environment being surveilled, the image comparison is perform by calculating the normalized cross correlation (NCC) between the two images. This is done only in the frame regions where the absolute value of the difference between the two registered frames is larger than a threshold. A second threshold is used to binarize the result, producing areas that are candidates to have abandoned objects in the target frame. A multiscale approach, with variation of both the NCC window size and the downsample factor to be employed in the frames dimensions, is used in order to allow the detection of objects of different sizes.

5.1

Roomba

To build the first prototype, a commercial Roomba from iRobot was used. It is adapted with supports, guide wheels, a netbook to command its movements, and embedded sensors such as a camera, a microphone, and a laser range finder. The device performs a back-and-forth movement inside a cable tray with speeds up to 0.5m/s. Firstly, tests were performed in a laboratory environment and then in an emergency diesel generator plant at CENPES, the research center from Petrobras S.A. (Fig. 5). This last cluttered scenario was essential to allow initial algorithms research and development, given the real world difficulties that emerge. The produced video database was used in the study of computer vision techniques to detect abandoned objects in the surveilled scenario with a moving camera. The acquired audio database was used in the research of algorithms to detect audio anomalies of impulsive nature, eventually diagnosing machinery malfunction, also taking into consideration that the sensors were in a moving platform.

• Object Detection: In order to further reduce both false positives and false negatives, the temporal filtering described in (Kong et al., 2010) is employed on the binary NCC images. After that, to increase even further the detection robustness, it is used a voting procedure in which a detection occurs only if the number of times a pixel is a candidate

3178

Anais do XX Congresso Brasileiro de Automática Belo Horizonte, MG, 20 a 24 de Setembro de 2014

Figure 7: Recording scenario employed in audio database development emulating event of interest (played in a loudspeaker) heavily corrupted by strong background noise.

Figure 5: Roomba based prototype moving in a cable tray at an emergency diesel generator plant.

phone was able to capture a change on the operating regime of the background pump by monitoring its fundamental frequency (pitch) along time, as depicted in Fig. 8, despite no significant change on the background-sound energy.

To test the video processing algorithms, more than 14 hours of raw video were recorded, producing about 60 videos containing abandoned objects, with 6 containing 15 objects each, and the remaining containing a single object. A total of 24 different objects were used. In this database, were varied the objects’ size, types and position along the robot’s path, the amount of objects in the same scene, and the illumination. Figure 6 shows results of the detection of multiple objects of different sizes in the same scene.

(a)

Figure 8: Background noise captured by a moving microphone and associated pitch analysis, indicating a regime change around 18 s.

(b)

Background-noise filtering: In this case, an event of interest (the sound generated by an industrial rotating machine) is heavily corrupted by the noise signal generated by the background pump. A first lap performed by the prototype, however, is able to model the background noise, and observe that it is restricted to the frequency interval f ∈ [200, 600] Hz. Once this interference is reduced or practically eliminated by a simple digital filter, the event of interest is easily detected by a spectral analysis, as seen in Fig. 9, allowing a subsequent analysis of its general characteristics. The robotic platform was also able to successfully build a 3D map of the plant in real time. Odometry, laser measurements and camera images are combined to build a 3D point cloud, where each point is associated to a color defined by RGB values. The 3D point cloud is processed based on probabilistic maps using Octomap ROS node (Hornung et al., 2013), which returns a representation of the environment. Figure 10 shows the panoramic view of the environment and the 3D map with relatively high

Figure 6: (a) Backpack and box, umbrella, and bottle reference frame and (b) detection. To develop and test the audio processing algorithms, a large database was devised emulating the adverse audio environment of an oil platform and the following events of interest: (i) A refrigeration pump, operated in two distinct modes, acted as the background noise; (ii) A fixed loudspeaker reproduced audio signals such as speech, whistling noise of a tea kettle, and 13 industrial machines, including the sound of rotating machines with different fundamental frequencies. A fixed microphone was set close to the background pump and another microphone was used on the moving platform to acquire the signal of interest (heavily corrupted by a reverberating version of the background signal), as illustrated in Fig. 7. Using this database, the following audio capabilities were devised for future integration on the DORIS system: Pitch detection: In this case, the moving micro-

3179

Anais do XX Congresso Brasileiro de Automática Belo Horizonte, MG, 20 a 24 de Setembro de 2014

(100 Hz)

Figure 9: Detection of an event of interest in 100 Hz after background noise filtering. precision. The map can be employed by a robot for task planning and execution, providing information for collision avoidance with obstacles in the environment. It is also possible to process the 3D colored point cloud looking for specific patterns in the environment, such as green and yellow pipes, or black valves. 5.2

(a)

DoriAna

DoriAna is a prototype developed to test the mechanical design of the traction module, the passive module, and the joint that couples them. A tubular track built using straight and curved segments was installed in the GSCAR laboratory, in COPPE/UFRJ. The track comprises all possible movements that the robot must make. The traction module consists of a wooden base, two sets of aluminum gimbals, polyurethane wheels, a machined prismatic mechanism that uses linear bearings to displace the traction wheels, and a clamping mechanism that uses a bicycle brake system to apply radial forces on the traction wheels. The passive module comprises only a base with two sets of gimbals. Two coupling joints are considered for the tests: one that uses a spring and a steel cable, and a double Cardan joint. The main objective of DoriAna (Fig. 11) is to test the following mechanical concepts:

(b)

Figure 11: (a) Traction and passive module of DoriAna prototype moving on a vertical curved section. (b) Rail installed in GSCAR/UFRJ. The force applied by the bicycle brake system was appropriate to hold and move the robot through vertical sections, showing that it is possible to achieve motion using only friction by applying a radial force. The joint with spring and steel cable performed better than the double Cardan joint. As for the traction system, the conclusion was that the prismatic mechanism is not satisfactory, given that it is prone to lock and the weight of the traction system led to loss of contact between the grooved wheels and the tube. This results suggest investigating an alternative concept for the traction system.

• The use of gimbals for guidance, stability, and weight support;

6

• The traction system mounted on a prismatic mechanism;

Conclusions

In this paper, we presented the DORIS project, which endeavors to develop an offshore facilities inspection and monitoring robot. The prototype is based on rail guided modules powered by a battery system and equipped with multiple sensors that enable detection of anomalies, such as abandoned objects and gas leakage. A prototype was built to validate anomaly detection under movement in a real environment similar to an offshore platform. Tests proved that the device is able to detect multiple objects in a video stream. Initial results with audio processing

• The clamping system, verifying whether the applied force is sufficient to support the robot in vertical sections; • The two joints used to couple the modules. Initial tests performed with the prototype show good performance of the gimbals in terms of stability. Even though the gimbals may shake slightly due to irregularities on the rail surface and asymmetrical positioning of the guide wheels, the base keeps a steady orientation while moving.

3180

Anais do XX Congresso Brasileiro de Automática Belo Horizonte, MG, 20 a 24 de Setembro de 2014

Figure 10: Panoramic image of the platform and the corresponding 3D map built by the prototype. algorithms indicate the possibility of detecting audio abnormalities in a noisy background scenario. The prototype was also capable of building a 3D map of the surrounding environment. Another prototype was built based on the mechanical design to test related concepts. Preliminary results show good overall performance of the guidance system using gimbals. It was proved the possibility of using just wheels on a tubular rail to achieve vertical motion by applying radial forces. A joint composed of a spring and a steel cable achieved good transmission of traction between the modules. The bad performance of the prismatic system led to the adoption of a different traction concept. Currently, a new mechanical concept is under development. In future works, all DORIS subsystems will be tested and integrated, and, finally, the complete robotic mobile monitoring system, composed of traction, sensing, battery, and manipulator modules, will operate in a real offshore platform environment.

Bengel, M. and Pfeiffer, K. (2007). Mimroex mobile maintenance and inspection robot for process plants. Fraunhofer Institute for Manufacturing Engineering and Automation IPA, pp. 1–2. Bengel, M., Pfeiffer, K., Graf, B., Bubeck, A. and Verl, A. (2009). Mobile robots for offshore inspection and manipulation, Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS)., pp. 3317–3322. Christensen, L., Fischer, N., Kroffke, S., Lemburg, J. and Ahlers, R. (2011). Cost-effective autonomous robots for ballast water tank inspection, Jour. of Ship Production and Design 27(3): 127–136. Fauroux, J. and Morillon, J. (2010). Design of a climbing robot for cylindro-conic poles based on rolling selflocking, Industrial Robot: An Int. Jour. 37(3): 287– 292. From, P. (2010). Off-Shore Robotics: Robust and Optimal Solutions for Autonomous Operation, PhD thesis, Norwegian University of Science and Technology. Graf, B. and Pfeiffer, K. (2007). Mobile robots for offshore inspection and manipulation, Proc. Int. Petroleum Technology Conf. (IPTC). Hartley, R. and Zisserman, A. (2003). Multiple View Geometry in Computer Vision, 2nd edn, Cambridge University Press, Cambridge, U.K. Hornung, A., Wurm, K., Bennewitz, M., Stachniss, C. and Burgard, W. (2013). Octomap: an efficient probabilistic 3d mapping framework based on octrees, Autonomous Robots pp. 1–18.

Acknowledgments This work is supported primarily by Petrobras S.A. and Statoil Brazil Oil & Gas Ltda under contract COPPETEC 0050.0079406.12.9 (ANP-Brazil R&D Program), and in part by CNPq and FAPERJ. The authors wish to thank all other members of DORIS project, including Alex Neves, Renan Freitas, Marcos Xaud, Igor Marcovistz, Gabriel Casulari, Thiago Braga, Fernando Coutinho, Allan da Silva, Lucas Thomaz, Gabriel Ramalho and Raphael da Silva from Federal University of Rio de Janeiro. We also wish to thank Auderi Santos, Pedro Panta, Felipe Noel and Jose Almir from ALIS Tecnologia for their mechanical consulting services.

Inuktun (2014). Inuktun versatrax 150TM vertical crawler. http://www.inuktun.com/crawler-vehicles/ . Accessed on February 11th , 2014. Kawasaki, H., Murakami, S., Kachi, H. and Ueki, S. (2008). Novel climbing method of pruning robot, Proc. SICE Annual Conf., an Int. Conf. on Instrumentation, Control and Information Technology, pp. 160–163. Kong, H., Audibert, J. and Ponce, J. (2010). Detecting abandoned objects with a moving camera, IEEE Transactions on Image Processing 19(8): 2201–2210. Kyrkjebø, E., Liljeb¨ ack, P. and Transeth, A. (2009). A robotic concept for remote inspection and maintenance on oil platforms, Proc. Int. Conf. on Ocean, Offshore and Arctic Engineering (OMAE).

References Anisi, D., Gunnar, J., Lillehagen, T. and Skourup, C. (2010). Robot automation in oil and gas facilities: Indoor and onsite demonstrations, Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), pp. 4729–4734. Baghani, A., Ahmadabadi, M. and Harati, A. (2005). Kinematics modeling of a wheel-based pole climbing robot (ut-pcr), Proc. IEEE Int. Conf on Robotics and Automation (ICRA), pp. 2099–2104. Bay, H., Ess, A., Tuytelaars, T. and Gool, L. V. (2008). Speeded-up robust features (SURF), Computer Vision and Image Understanding 110(3): 346–359.

NREC/CMU (2012). Sensabot: A safe and cost-effective inspection solution. Jour. of Petroleum Technology, pp. 32–34. Quigley, M., Conley, K., Gerkey, B., Faust, J., Foote, T., Leibs, J., Wheeler, R. and Ng, A. (2009). ROS: an open-source robot operating system, Proc. Int. Conf. on Robotics and Automation (ICRA). Skourup, C. and Pretlove, J. (2009). The robotized field operator, ABB Review (1): 68–73.

3181