1

Design and Implementation of BumbleBee AUV

Hoang The Huan (Team Captain), Goh Eng Wei, Grace Chia, Alex John, Tey KeeYeow, Tan Soon Jin, Yaadhav Raaj, Vanessa Cassandra, Gao Bo, Steven Harta

Prawira, Foo Rong Xuan, Lim Qi Xiang, Esther Tan, Louis Goh, Joshua Yow, LiuShih-Chiang, Akshat Dubey, Pyae Phyo Tun, Lim Junjie

Abstract—BumbleBee Autonomous Underwater Ve-hicle (BBAUV) is the product of a team of undergrad-uates from National University of Singapore (NUS).This vehicle is designed for two competitions: theRoboSub Competition and the Singapore AUV Chal-lenge. The BumbleBee vehicle was fully modelled inCAD and fabricated with CNC machining, laser cuttingand 3D printing. This year, Bumblebee has beenimproved with better machine vision cameras, moreaccurate navigation and imaging sonar integration formore robust computer vision. BumbleBees sensor suiteincludes a Doppler Velocity Log (DVL), an imagingsonar, a hydrophone array, an Inertial MeasurementUnit, two machine vision cameras and a depth sensor.Its software architecture is built upon Robot OperatingSystem (ROS) and the complex vision algorithms havebeen implemented in OpenCV.

I. INTRODUCTION

For the third year, Team BumbleBee de-signed and built an Autonomous UnderwaterVehicle (AUV) for two annual competitions:the AUVSI International RoboSub competitionand the Singapore AUV Challenge. RoboSubis held in July in California; while the Sin-gapore AUV Challenge was held in Marchin Singapore. Both competitions are designedwith challenges that mirror industrial missions:visual recognition of objects, manipulation andacoustic localisation tasks. The Bumblebee de-velopment cycle this year has two paralleltracks: (1) to design and development the nextgeneration Bumblebee 3.0 vehicle to compete inRobosub 2016 and (2) upgrading of Bumblebee2.0 to Bumblebee 2.5 to patch the issues thevehicle faced from Robosub 2014.

Team BumbleBee is divided into Mechanical,Electrical and Software subteams. The teamcomprises students from mechanical, electrical,

computer engineering and computer science ofall years of studies.

Fig. 1: BumbleBee AUV 2.5

II. SPECIFICATION

Weight 50 kgDimensions 0.7m X 1.1m X 0.5m

Single Board Computer

Core i7 - 3610QEAaeon EMB-QM778GB DDR3 RAM512GB SATA3 SSD

Embedded SystemArduino Mega 2560Xilinx Spartan-3 on NI sbRIO 9602

Propulsion6 SeaBotix BTD1502 VideoRay Surge Thrusters

NavigationTeledyne RDI Explorer DVLSparton GEDC-6 IMUUS300 Pressure/Depth Sensor

Vision SensorsAVT Guppy ProAVT Guppy

SonarBlueView P450 Imaging Sonar4 Teledyne Reson TC4013 Hydrophones

Manipulators Festo Pneumatics SystemsPower Supply 22.2V 10000mAh LiPo Battery (x2)Underwater Connectors SubConn Micro and Low Profile Series

Software ArchitectureRobot Operating System (ROS)Debian GNU/Linux x64

TABLE I: BumbleBee AUV 2.5 Specifications

2

III. MECHANICAL SUB-SYSTEM

BumbleBee 2.5 features several upgradesfrom its original design. While maintaining itsindustrial Remote Operated Vehicles (ROVs)shape, enclosures for pneumatic pod and IMUhave been changed.

A. Frame

Fig. 2: Frame

The vehicle’s frame is made by laser cuttingaluminium sheet at Cititech Industrial Engi-neering and is bolted together using fasten-ers from Bossard. This robust and rigid frameencompasses all components while protectingthem from direct impact. It also allows usto flexibly reposition components for optimalhardware stability through its unique regulardrill pattern across its side.

B. External Hull and Enclosures

Fig. 3: Hull Assembly

The main electrical hull is made of standardsize acrylic tube (250mm diameter) and it fea-tures end caps that are fabricated by Com-puter Numerical Control (CNC) machining ofaluminium in the university’s workshop. The

main hull has a rapid disassembly feature thatis achieved by using six draw latches to com-press a face seal o-ring between the end caps.This improvement is a major breakthroughover the previous bolted flange design as iteases access to electronics within the hull.

A valve on the hull allows it to be pressur-ized and also monitors any drops in pressure.This feature allows the hull to be tested for anyleaks before the vehicle is submerged in water.By pressurizing the hull to around 120kPa dur-ing operations, any leaks in the water wouldbe made visible by escaping bubbles. Apartfrom that, a pressure sensor in the hull allowsus to constantly monitor fluctuations in hull,which is a reflection of hull temperature. Thehull also has a leak sensor in the case of fail-ing pressure sensor. Leak alert is propagatedto Bumblebee Control Panel which displaysa warning. Through Hallin Marine’s pressuretesting chamber, the hull is rated to a depth of40 meters.

Bumblebee’s external enclosures house theelectrical systems located outside the main hull.These comprises of the DVL housing, camerahousing, acoustics housing, IMU housing andbattery pods.

• DVL Housing:The DVL housing is fabricated by weld-ing 6061 aluminium tubes into a T struc-ture. The vertical tube houses the sensorhead while the horizontal tube houses theelectronic box. A Subconn micro circularbulkhead is used to provide an electricalconnection which is linked to the mainelectrical hull.

• Camera Housing:The camera housing is fully machined in-house and an adapter that utilizes Festo’spneumatic push-in fitting waterproof fea-ture is used to pass through the firewirecable for minimum data loss.

3

Fig. 4: DVL Housing Assembly

C. Thrusters Configuration

Bumblebee’s propulsion system consistsof six Seabotix BTD150 and two Videoraythrusters. Four Seabotix thrusters are used toprovide movement in the z-direction (heave).The other two horizontal Seabotix thrusters areused for movement in the y-direction (sway).On top of that, the two Videoray thrusters areused to provide movement in the x-direction(surge). By controlling the thrust of variousthrusters in a certain manner, we are able toachieve 6 DOF.

Fig. 5: Thrusters Configuration

D. Manipulators

Bumblebee is equipped with a set of 3Dprinted manipulators on board, consisting of apair of grabber arms, a pair of torpedo launch-ers and a ball-dropper that are actuated using

a pneumatics system. The compressed air forthe pneumatics system is stored in a 13 cubicinch Ninja Paintball Tank. This year, some ofthe manipulator designs are revised. Firstly, thedesign of the grabber arms is improved so asto enhance their versatility in grabbing objectsof different shapes and sizes.

Fig. 6: Grabber

Simultaneously, compactness is incorporatedin the design to ensure better integrationinto the vehicle. Secondly, the torpedo de-sign involves the use of an O-ring to buildenough pressure to ensure sufficient pressurefor launching the torpedo. Thirdly, the ball-dropping mechanism has undergone a re-design to become more compact. All of Bum-blebee’s pneumatics system is supported byFesto Pneumatics Suppliers in Singapore.

Fig. 7: Dropper

4

IV. ELECTRICAL SUB-SYSTEM

The electrical system comprises the power,sensors and actuators, and the computer sub-systems. The previous OpenUPS power sub-system has been replaced with the integrationof Battery Management Boards. The sensorsand actuators have been redesigned and ad-ditional ones are integrated. The main com-puter board hosts a quad core i7 processor forfast multicore software processing. The compo-nents are integrated on a single multi-level rackfitted into a single hull.

Fig. 8: Electrical System Block Diagram

A. Camera System

This year, our vehicle features a major up-grade to the system with the incorporation ofnew machine vision cameras interfaced overfirewire as follows:

• Front Camera: Guppy Pro F046C with Ed-mund Optics 4.5mm fixed focal length lens

• Bottom Camera: Guppy F146C with Ed-mund Optics 4.5mm fixed focal length lens

• Firewire interface: PCIe 4 port1394a(Firewire) from Allied VisionTechnologies

The previous vision system faced constantlychanging colour reproduction and automaticexposure which caused major issues for ourthresholding of the various colours in objectdetection and identification. Our previous cam-eras were based on webcams and paled in com-parison to the Guppy cameras which boastedfar superior performance due to their camerasensors with better field of view (with the4.5mm focal length lens), lower image noiseand higher dynamic range. The Guppy cam-eras are also more readily configurable al-lowing full access to the camera parameters(such as shutter, exposure, white balance) to betuned.

B. Power System

The vehicle is powered by two 10000 mAhLiPo (lithium polymer) batteries extending test-ing time to approximately 210 minutes beforerequiring a recharge.

i. Power Monitoring and Management:Each LiPo battery is installed into a batterypod which allows for both charging anddischarging. Within the pod, the batteriesare connected to PMB (Power Monitor-ing Boards) which monitor vital powerstatistics such as current, cell voltage andcapacity. The custom fabricated PMB hasbeen designed to withstand a maximumcurrent of 30 A. This system allows track-ing of power statistics of the batteries andis more reliable than the previous off-the-shelf OpenUPS system. A battery chargingbox has been designed for quick deploy-ment of mobile charging stations. This bat-tery charging box supports parallel charg-ing of two battery pods of up to 25 A perchannel at one go.

ii. Power Distribution:The power system utilizes a M4 ATXpower supply to generate three voltagerails to power the SBC and the sensors andactuators. A Y-PWR provides load balanc-ing hot-swap capabilities for the batteries.

5

Batteries can be swapped out one by onewithout powering down the SBC.

C. Sensors and Actuators

Fig. 9: Sensors and Actuator Board Design

BumbleBee 2.5 presents a redesigned andcustom fabricated Printed Circuit Board (PCB)for the sensors and actuators shield. The Ar-duino Mega 2560 microcontroller developmen-tal board is based on the previously used AT-Mega 2560 microcontroller. The previous shielddesign was continued because it provides quickswap of the microcontroller which is crucial formodular debugging during port or componentdamages.

New sensors such as internal pressure andhumidity sensors are added onto the sensorsand actuators board, to the existing suite ofthe GEDC-6 IMU, depth and temperature sen-sors. This board also interfaces the six SeaBotixthrusters and the pneumatic manipulators. Thecurrent thrusters are enhanced by includingactuator controls to the electronic speed con-trollers operating on the two VideoRay surgethrusters.

In addition, LED strips are integrated as stateindicators, while a TFT LCD screen displayprovide visual feedback on the system sensorstatus. These indicators are especially usefulfor understanding the vehicle’s current stateduring autonomous runs.

D. Navigational Sensors

i. Sparton GEDC-6 IMU:The Sparton GEDC-6 IMU (Inertial Mea-surement Unit) provides critical inertialdata at a rapid rate of 100 Hz. The IMU’sproprietary algorithms ensure the output

of correct data despite the presence ofelectromagnetic interference generated bythe BumbleBee’s suite of electronics andthrusters.

ii. Teledyne RDI Explorer DVL:The DVL (Doppler Velocity Log) is an ac-tive sonar system that tracks the velocityof the instrument via a four-beam solutiondirected at 30 degrees nominal from thesensor’s ceramic head. The velocity read-ings obtained are combined with tilt andaltitude measurements, then resolved intothe three orthogonal x, y and z axes viaa least squares fit solution. These resolvedreadings are further filtered through a di-rect three-degree of freedom Kalman filter,which serves to attenuate noise. The calcu-lations output a more accurate positionalcoordinate of the vehicle.

iii. Navigation:The data provided by the navigational sen-sors are fused together via trigonometricequations to generate a global position vec-tor. The odometric data obtained from thisvector allows precise navigation to a givenspatial coordinate.

E. Computer System

BumbleBee’s software system is powered byan Intel Core i7-3610QE quad core processoron an Aaeon PCM-QM77 motherboard alongwith a 512 GB SATA SSD (Solid State Drive).This upgrade from the previous 16 GB SSDaccommodates more software data and rosbagdata collected.

A USB hub interfaces the embedded sensorsand actuators as well as other serial devices, i.e.two PMBs and the GEDC-6. The imaging sonaris connected via Ethernet with a PoE (Powerover Ethernet) connection to the SBC, dedi-cating imaging sonar bandwidth to the maincomputer. VGA and USB ports are exposedto allow external debugging of the softwaresystems.

The computer is connected to dockside viaa 1000 Mbps Ethernet tether. The vehicle isnetworked to a Gigabit switch that connects toNI sbRIO 9602 and surface router.

6

F. Control System

Six PID (Proportional Integral Derivative)control loops are used to control the vehicle’ssix degrees of freedoms. The PID controllers aredesigned with the following considerations:

• Low pass filter for the derivative compo-nent to reduce the exponential effects onsensor noise

• Variable period time sampling for moreaccurate integral and differential compu-tation

• Weighted set points to reduce transienteffects in set point changes

• Integrator windup protection for when ac-tuators are unable to fulfill the PID Con-troller requirements

The PID control loops have been improvedfor dynamic allocation of actuator limits, al-lowing greater output in specific degree offreedoms. Velocity controllers for the surgeand sway domains have been implemented formore precise maneuvering of the vehicle dur-ing mission runs. A control system tuning UserInterface was designed to assist in softwaretuning of the control parameters.

Fig. 10: PID tuning UI

V. ACOUSTIC SUB-SYSTEM

BumbleBee’s acoustic sub-system uses fourTeledyne hydrophones, integrated with customanalog and digital boards. The MUSIC (MUlti-ple SIgnal Classification) algorithm is used tolocalize the acoustic pinger. This year’s acousticsub-system features a redesign of the software

to improve the performance and robustness ofthe algorithm.

A. Hardware

The hardware setup for the acoustics subsys-tem on BBAUV consists of 4 Teledyne TC4013hydrophones, integrated with custom analogand digital boards. The setup is shown in thefigure below and each component is furtherelaborated.

Fig. 11: Hardware Flow Diagram on AcousticSub System

i. Teledyne TC4013 Hydrophones4 compact 9.5 mm Teledyne TC 4013 hy-drophone are arranged in a square arraywith inter-element spacing of 1.5 cm. Thisprevents spatial aliasing associated withphase-difference based algorithms. The hy-drophone mount is designed on Solid-Works and fabricated precisely using laser-cutting technology.

ii. Custom Preamplifier and Bandpass Filter boardA Custom Preamplifier board is used toamplify the signal from the hydrophonesto a suitable voltage for Analog-to-Digitalconversion. The signal is low-pass filteredto remove the DC components to allow thesignal of interest to be amplified withoutclipping. It is then high-pass filtered toremove high frequency noise above theNyquist rate as these may cause aliasingwhen converted to digital form. The sig-nal is then passed through a Low-NoiseAmplifier to amplify the signal to suitableamplitude so that the resolution of theAnalog to Digital Converter is maximized.

iii. NI9223 Analog Input ModuleThe NI9223 Analog Input module thenconverts the signal to digital form. TheNI9223 can sample at a rate of 1MS/ch/s

7

with 16 bit resolution (±10V) on up to 4different channels.

iv. NI sbRIO 9602Lastly, the signal is processed by the Na-tional Instruments sbRIO 9602 to com-pute the Direction of Arrival (DOA) ofthe signal. The sbRIO 9602 has a 400MHzPowerPC processor paired with a XilinxSpartan-3 Field Programmable Gate Array(FPGA) with 2M gates.

B. Software

The Acoustic subsystem this year features aredesign of the software architecture to fullyutilize the software resources available. Thisresults in an improvement in the speed of thealgorithm. Also, by pushing the MUSIC algo-rithm down to a lower level processor, CPUresources on the main computer of BBAUV canbe dedicated to other tasks.

Fig. 12: Acoustic Sub-System SoftwareArchitecture

The usability and robustness has been im-proved of the acoustic sub-system. Error codesallow the main computer to gain insight on thefunctioning of the acoustic sub-system withouthaving to access the program directly, allowingfor easier debugging. The system can also auto-matically detect faults in the hydrophones andcompensate accordingly, allowing the systemto operate even in the event of 1 hydrophonefailing, making it more robust.

The software architecture this year dividesthe tasks to be completed between the pro-cessor and the FPGA. The processor handlesless computationally intensive tasks like band-pass filtering, Fast Fourier Transform (FFT) andcommunications with the main computer. TheFPGA handles computationally intensive tasksor time-critical tasks like the computation of

the MUSIC spectrum and the sampling of theanalog signal from the hydrophones.

The division of the software tasks betweenthe processor and FPGA allows the algorithmto complete each cycle of processing in un-der 800ms, while not compromising the ac-curacy and robustness of the system. This al-lows BBAUV to make movements much morequickly when localizing to the pinger.

Statistical analysis is done on the hy-drophones to determine if the readings arefaulty. The mean of readings is checked toensure the signals are grounded properly. Thevariance is also checked to ensure that theminimum background noise is being receivedat the hydrophones. If the readings from onehydrophone are determined to be faulty, itsidentity is sent to the main computer and thereadings from that hydrophone are ignored inthe computation of the Direction of Arrival.This allows the acoustic system to continueoperating even in the case of a failure in onehydrophone channel.

VI. SOFTWARE SUB-SYSTEM

Fig. 13: Software Block Diagram

Bumblebee AUV’s software stack runs on theopen source message passing interface, RobotOperating System (ROS) which in turn runson top of Linux. ROS provides a standardizedmedium for package management, serializa-tion and over the network message passingbetween processes by utilizing a graph archi-tecture. One of the biggest advantages that thisoffers is that we can run processes outside the

8

AUV that can control the system when its inwater as well as connect to processes runninginside a deployed AUV which makes real timedebugging feasible.

The various processes that need to be run arehence written as nodes of a graph. The systemis headed by the mission planner which has theauthority to start, stop, monitor the progress ofand time each individual task.

Each task node contains its own internalstate machine to perform its various tasks. Thisallows easy flow control and avoids the useof jump statements and other convolutions forcomplex logical sequences.

A. Mission Planner

The mission planner is implemented usingFinite State Machines and a graph walkingalgorithm. Finite State Machines are used be-cause each task in the mission sequence canbe represented as its own individual stateswith known inputs and a deterministic set ofoutcomes. The mission planner is a high levelprocess and therefore written in Python forease of modification and compatibility withvarious libraries that may be used.

The mission planner has the capability todynamically load modules and generate linkedstate machines on the fly.

B. Navigation

The navigational sensor suite consists of a 9axis IMU, a DVL and a barometric depth sen-sor. Provisions have been made for the integra-tion of a GPS receiver as well. The sensor datafrom the Doppler Velocity Log (DVL) is fusedwith the data from the 9 axis IMU and depthsensor is fused to obtain independent statedata. An error state Kalman Filter is used toobtain more accuracy than each sensor is ableto provide independently. It is notably morerobust and suitable for dynamically changingstates than the traditional full-state KF whichhas inherent assumptions of vehicle motion be-haviours in the state equation setup. Since errorvariables are used as the state vector, nonlin-earities can be canceled. In addition, motion as-sumptions are not necessary in formulating thestate equations. The absence of these motion

assumptions greatly enhances the robustness ofthe filter state equations in handling variousvehicle manoeuvres.

Fig. 14: Navigation Block Diagram

C. Computer Vision

Our basic vision processing framework con-sists of color thresholding, contour detection,morphological transformation and histogramequalization. Though color thresholding is sus-ceptible to underwater perturbations, selectingoptimal color space such as HSV, LUV or LABfor thresholding may yield better result de-pending on the object of interest. In addition,dynamic selection of threshold based on infor-mation such as contrast, brightness or entropyof certain channel. For image classification,contour features such as convexity, eccentricityand ellipticity are particularly useful to distin-guish various shapes.

This year we have also explored severalimage processing techniques to achieve betterimage segmentation and object tracking in Ro-bosub water condition. More focus is given onpreprocessing of raw image captured from thecamera to rectify problems such as specularreflections, poor visibility and color cast. Forinstance, gamma correction and dark chan-nel prior are employed to handle hazy watercondition that causes loss of details. Havinglearned from previous Robosub experience, in-stead of relying on a single method to identifyobject of interest, current vision algorithm uti-lizes ensemble of classifiers to achieve success-ful segmentation in varying water condition. Acombination of edge detection, color threshold-ing and saliency detection provides consistenttracking of delorean and train despite loss ofcolor intensity by light attenuation.

9

As most vision algorithms call for constantfine tuning of parameters to obtain maximumefficiency, utility tools such as ROS dynamicreconfigure client and rosbag tool to makepossible rapid development of our vision al-gorithms. Besides that, we also write utilitylibrary for more comprehensive analysis of bagfiles obtained from daily practice run. With theutility library, we are able to possess a bettergrasp of minute informations such as effect ofeach vision filter.

Fig. 15: Processed Image on ROS

Fig. 16: Analysis of different channels ofimage frame after preprocessing

D. Imaging Sonar

This year, a new framework has been de-vised for better localization and tracking of tar-gets in the forward space. A new multi-sensoryfusion approach is used, involving the camera,sonar and vehicular POSE and velocity. Previ-ous approaches for tracking targets involvedHSV thresholding and contour detection in thecamera space, but were susceptible to noise dueto poor visibility.

In order for this fusion to work, a pre-calibration step is performed to find the rela-tionship between the sonar, camera and vehicu-lar dynamics. This step solves for unknowns aspart of a model involving the vehicle’s POSE,camera’s intrinsic matrix and 3D sonar coordi-nates derived from the sonar image and depthsensor of a fixed calibration object.

With the above unknowns solved, we canproject the sonar coordinates of any objectinto the camera space. If the object depth isunknown, the sonar’s 20 degree vertical fieldof view can be used. Since the object’s beingtracked are stationary, it’s position is a functionof the Angular (Yaw) and Tangential (Vehicle)velocity as seen in Figure 16, hence the correctobject can be selected. This is then projectedto the camera as seen in Figure 17. From there,the multiple spaces are fused using a RecursiveBayesian Inference filter, and the final targetpoint of interest is obtained.

Fig. 17: Optical velocity fused with Angular(Yaw) and Tangential (Vehicle) velocity.

Fig. 18: Sonar plane is projected onto thecamera plane and fused using a Recursive

Bayesian filter.

10

VII. VEHICLE STATUS AND TESTING

BumbleBee 2.5 has been undergoing exten-sive pool tests at Queenstown Swimming Com-plex in preparation for RoboSub 2015. Priorto the integration of the vehicle, the mechan-ical subsystem was thoroughly leak tested; theelectrical components were bench tested; thesoftware subsystem was constructed and testedon recorded data from previous runs.

VIII. CONCLUSION

Bumblebee 2.5 has undergone rigourous test-ing and evaluation to achieve the robust per-formance and capabilities it has today. Concur-rently, the many lessons learnt in the designof this vehicle will be carried over to the nextgeneration Bumblebee 3.0 which will see de-ployment for next year’s Robosub. Our teamhas evolved considerably also with better or-ganisational structures and greater capacity fordevelopment. We are confident that Bumblebee2.5 is ready to meet with the challenges ofRobosub 2015.

Fig. 19: Team Bumblebee 2015

IX. ACKNOWLEDGEMENT

Team BumbleBee would not have beenwhere we are today without the followingpeople and we would like to thank them forgiving us the chance to compete in RoboSub:

Title SponsorNUS: For their cash sponsorship, equipmentprocurement, and academic support in ourproject.

Platinum SponsorSuperior Energy Services(Hallin Marine): Forsupporting the early developments in ourproject with cash sponsorship and air ticketsfrom 2012 onwards.Seatronics: For the continued loan of ExplorerDVL, BlueView Imaging Sonar and underwaterconnectors.Cititech: For fabrication support of ourmechanical parts and competition obstacles

Gold SponsorsTemasek Hall, National Instruments, Festo,IKM, KOMTech, Kentronics Engineering

Silver SponsorsTeledyne Reson, Macartney UnderwaterTechnology, Allied Vision Technologies,Edmund Optics, Sparton

Bronze SponsorsSolidWorks, JKS Offshore, MathWorks, SECOTools, Evernote, Sterling Comm Intl Pte Ltd,Aaeon Technologies, Deep Sea Power andLights, Southco, Bossard, Tekin, Digikey,Aztech, SeaBotix

Supporting OrganisationsDSO National Laboratories, Advanced Marineand Acoustic Research Lab, Singapore SportsCouncil, Hometeam NS