NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 1

RoboSub 2023 Technical Design Report

National University of Singapore (Bumblebee Autonomous Systems)

Amadeus Aristo Winarto, Amanda Koh Jing Ling, Ananya Agarwal, Benjamin Koh Zhao Hui, Chai Zi Yang, Chen Jiawei,
Chew Zhi En Samuel Joshua, Chin Zheng Hao, Drustan Yeo, Gokul Rajiv, Guk Yi Siong, He Shaoliang,
Ho Wei Zong Jasper, Irwin Kong Xunmeng, Justin Foo Guang En, Kaitlyn Ng Ke Yi, Kristoffer Videl Wijono,
Lee Chan Wai, Lee Shi-An Matthew, Lee Tze Han, Leong Deng Jun, Leong Xin Lei, Li Po Hsien, Lim Sheng Wei,
Lu Sicheng Isabella, Marvin Pranajaya, Michael Jervoso, Ng Cheng Yang Titus, Ng Xing Yu, Ng Zhia Yang,
Nguyen Minh Nguyen, Niu Xinyuan, Patrick Joy Surbakti, Png Qun Shen, Quek Wei, Rani Karthigeyan Rajendrakumar,
Sun Qifan, Seow Alex, Sim Justin, Stevanus Williem, Tan Chern Lin Justin, Tan Chew Miang Edwin, Teoh Xu En,
Tran Phuoc Huy Khang, Yam Jin Ee Dmitri, and Zhu Tianqi

Abstract—For RoboSub 2023, Team Bumblebee’s strategy

involves deploying the BBAUV 4.1 to efficiently complete
all tasks. Already optimized for space, weight and ease of
maintenance, mechanical work centered on the reliability
of our manipulators to complete all tasks with maximal
success. Electrical work has been centered on maintaining
current systems and researching new technologies and
development processes. Software focus was on ease-of-
development improvements and an overhaul of our per-
ception pipeline. New parts were tested thoroughly in
simulation and after integration with the vehicle.

I. C OMPETITION S TRATEGY Fig. 1: 3D model of the BBAUV 4.1.

For RoboSub 2023, we plan to deploy our BBAUV

4.1 vehicle (Fig. 1) to complete all competition B. Course Strategy
tasks. Focusing on speed and accuracy, we aim
to achieve a 100% success rate for all navigation We approach the course using a sensor fusion ap-
and actuated tasks. By optimizing our vehicle for proach, where we combine readings from various
competition tasks and validating its performance sensors such as our sonar, object detections from
through extensive testing, we believe we possess machine learning (ML), and acoustic signals from
a strong, stable platform for success at this year’s the pingers, to accurately locate and identify objects
RoboSub. relevant to the task, such as buoys in the Start Di-
aling challenge or torpedo openings in the Goa’uld
Attack challenge.
A. Competition Vehicles
Aligning to targets is a significant challenge, requir-
We will be competing with a single vehicle, the ing a robust control system and accurate estimations
BBAUV 4.1. Despite the physical similarity to its for the positions of the target, vehicle, and periph-
predecessor, much work has been done to upgrade erals such as grabbers and torpedo launchers.
its features. While retaining its size and weight ad-
vantages over the competition, substantial improve- Previously, we achieved this by combining ML
ments have been made to the vehicle’s accuracy, object detection results with 3D computer vision
reliability and maintainability. With this, we are techniques, such as homography estimation. This
confident of our ability to adapt to challenges during year, we have expanded on our use of computer
the competition preparation. vision to enable even faster and more accurate
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 2

alignments (see Section II-C2). Apart from adding buoyancy, fibreglass floats at-
tached to the main hull’s exterior also protect the
To better ascertain vehicle position, our Inertial
vehicle from impacts in transit or during testing.
Measurement Unit (IMU) has been upgraded to en-
Notably, floats of the top cover encase and shield the
able our localization pipeline to give more accurate
externally mounted sonar. The vehicle’s buoyancy
readings. Several tests were conducted to compare
can be fine-tuned by inserting floats in the cavities
different IMUs and presented in more detail in Sec-
of the 3D-printed shell affixed to the underside of
tion III-B. For knowledge of the externally mounted
the floats.
peripherals such as the actuation modules, our vehi-
cle’s Unified Robotics Description Format (URDF) An octagonal frame facilitates the mounting of ac-
was extensively used to model the relationships in tuation modules, allowing for rapid prototyping and
position and orientation between different parts of testing of our various actuators (see Section II-A3).
our vehicle. This allows for much better delivery Carrying handles on either side of the hull allow for
of autonomy compared to previous years, especially ease of handling.
with regard to aligning to obstacles.
To coordinate task-specific strategies, we employ 2) Design of Battery Hull
a Behaviour Tree-based mission planner. Based on Another creative aspect of the BBAUV 4.1 is the
our experiences using it for nearly two years, im- battery hull, manufactured with novel 3D metal-
provements and refinements were made in the run- printing technology; we also increased the rigidity-
up to this year’s competition. to-weight ratio by embedding lattices in the walls
and base (Fig. 3). The main and battery hulls are
directly connected with right-angled SubConn Low
II. D ESIGN C REATIVITY
Profile connectors, doing away with messy cables
A. Mechanical Sub-System and making battery changes quick and simple.
1) Design of Main Hull

Fig. 3: Isogrid layer of the battery hulls.

3) Design of Actuation Systems

Facing the novel challenge of picking up a large, flat
Fig. 2: Internal layout of the main hull chevron, two grippers utilizing different techniques
were designed and tested in parallel to maximize
Inherited from its predecessor, a rectangular hull our chances of success.
was chosen for efficient packing of internal com-
Our stepper motor-driven gripper (Fig. 4(a)) uses
ponents and electronics (Fig. 2). Finite Element
torsion springs in its compliant claw design. Capped
Analysis was used to verify this design’s ability to
with a layer of rubber for extra grip, this design
withstand the 3-bar pressure expected during opera-
grasps all course obstacles reliably, enabling us to
tion. A center divider adds rigidity and also isolates
tackle Location and Engaging Chevrons.
electrically noisy components from sensors. The
telemetry screen window in the hull cover allows Anticipating chevrons close to the DHD walls to be
us to monitor vehicle vitals during testing, and is challenging for the claw, inspiration was taken from
outfitted with improved status lights to track the industrial vacuum suction arms to create our own
vehicle’s progress through the competition course hydraulic vacuum gripper (Fig. 4(b)). A peristaltic
for easier troubleshooting. pump uses surrounding water as the hydraulic fluid
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 3

allowing for easier triaging and faster troubleshoot-

ing. This minimizes vehicle downtime when testing
and dedicates more time towards trial runs for the
actual tasks.
Apart from relying on upper layers of our software
stack, we have also introduced telemetry reporting
Fig. 4: (a) Stepper motor-driven claw gripper. (b) Pump-powered at the network layer. By modifying the SBC-CAN
suction gripper. board which bridges the CAN and Ethernet do-
mains, we are now able to also broadcast telemetry
data over UDP, providing redundancy and flexibility
to create a local vacuum inside the suction cup, in our monitoring.
holding the chevron tightly to the arm. This arm
is extended directly on top of a chevron using 2) Power Control and Monitoring
a telescopic pulley mechanism which will not be
obstructed by the DHD walls. Due to the flat and Our custom-designed Power Monitoring Board
smooth surface of the chevrons, a flat suction cup (PMB) reports battery statistics such as voltage,
was used instead of a bellows suction cup. charge and current consumption. Additionally, it is
capable of determining the battery’s state of health
Both our ball dropper and torpedo launcher use and charging cycle, allowing for better estimation of
the Bluetrail underwater servo for their excellent operational time in light of degrading battery life.
reliability and depth rating. Our 3D-printed torpedos The routing of power between different components
have also been refined to be more hydrodynamic is shown in Fig. 22.
(see Section III-C), allowing Goa’uld Attack to be
completed consistently. When the voltage drops appreciably, PMB fail-safes
ensure that the onboard under-voltage protection
system will trigger a warning and disable electronics
B. Electrical Sub-System to prevent damage to the batteries. Coupled with
1) Communications Architecture a centralized power control board, systems can be
The BBAUV 4.1 adopts a heterogenous architec- selectively prioritized to conserve power during low-
ture, with custom PCBs each serving a specialized power scenarios. More importantly, this affords the
purpose. Various communication protocols are used ability to independently power cycle components.
by these boards, sensors and our single-board com- Hot-swappable batteries also reduce the need for
puters (SBC) as detailed in Fig. 21. Inspired by disruptive restarts of the vehicle, keeping our op-
automotive applications, we utilize the Controller erational uptime as high as possible.
Area Network (CAN) bus for its robust, half-duplex
differential communication between embedded sys- 3) Backplane System
tems. Where high bandwidth data transfer is re- Our backplane system mounts the vehicle’s elec-
quired, SBCs and peripherals providing a networked tronic components to the main hull, allowing for
interface are connected to an Ethernet switch. The more versatility as opposed to mounting on the
CAN network is split into two unconnected buses end-cap. The backplanes (Fig. 15 and 18) provide
that rely on the Thruster and Actuation Board (TAB) power and communication lines that simplify the
to forward messages between them. development of custom daughter boards (Fig. 16, 17,
19 and 20) as connectors are internally standardized,
This year, we have upgraded our Operator Control
which also allow for boards to be easily redesigned
Station (OCS) that provides connectivity between
in the future.
the BBAUV 4.1 and the software subteam during
tethered operation. Quality-of-life improvements in- In preparation for RoboSub 2023, more power and
clude an integrated 4G modem and enhanced WiFi communication lines were added to the backplanes
connectivity. An in-built telemetry reporting screen to reduce the amount of wired connections in the
allows members on standby to monitor critical hull. This reduces crosstalk and electrical noise,
statistics (e.g., battery voltage and board statuses), and aids maintenance and debugging efforts. More
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 4

debugging ports were also added to aid monitoring. and evaluation (see Section III-A), a 9 LED config-
The new backplanes also have a thicker copper layer uration for the module was selected and integrated
that improves heat dissipation and increases the with our Sensor and Telemetry Board (STB). The
current limit that can be drawn by daughter boards. copper-cored PCBs are directly connected to the
This allows greater speeds to be achieved without exposure pad of the LEDs to facilitate heat dissi-
risk of the thruster Electronic Speed Controls (ESC) pation and keep temperatures sufficiently cool. For
overloading the power lines. brightness control of the LEDs, the LT3950 constant
current driver was selected for its flexibility and
In our experience with plug-and-play daughter
safety features such as overcurrent, overvoltage and
boards, as the connectors near their rated mating
open LED.
counts, the gradual oxidation and degradation of the
contacts result in slower transmission speeds and
occasional disconnections. As a preventive action, 6) Actuation
contact cleaners and lubricants are extensively used With our two gripper designs, our actuation system
in our maintenance routines to extend the lifes- now consists of 2 Bluetrail underwater servos and
pan of existing boards and drive down cost and 2 stepper motors. The latter utilizes the TMC2209
wastage. drivers for smooth and precise control of the grip-
pers. Using Trinamic’s StallGuard™, actuators can
4) Firmware Improvement + DevOps be calibrated to a reverse-EMF threshold in order
to automatically stop the stepper movement without
Control messages from the software stack are for-
additional sensors or external input. Using Cool-
warded through multiple boards before arriving to
Step™, the maximum current threshold is dynam-
the ESCs. Reducing the latency of our communi-
ically set based on the load of individual motors,
cations and making the firmware generally more
minimizing the overall power consumption and heat
performant thus had a great impact on the vehi-
generation. The newly added suction gripper also
cle’s responsiveness. This was primarily achieved
uses SpreadCycle™ in its telescopic pulley mecha-
by moving from polling loops towards interrupt-
nism for dynamic motor control.
driven, Real-Time Operating System (RTOS)-like
code structure. Upstream software are also posi-
tively affected; less oscillatory behaviour was ob- 7) Acoustic Signal Processing
served in our Proportional-Integral-Derivative (PID) Due to its stability, our acoustic subsystem has
controller due to the faster feedback loop between undergone minimal changes. An automated pro-
controller and ESCs. grammable gain amplifier on the Data Acquisition
Board (DAQ) normalizes incoming pings to reduce
This is complemented by our migration to the Plat-
signal clipping, providing consistent measurements
formIO IDE, which supports multiple architectures
at all distances from the pinger. Pings are extracted
and platforms in a single codebase, simplifying
using short-time Fourier transforms with dynamic
code organization and sharing of common code.
thresholding, and pings with a low signal-to-noise
This has facilitated standardization and reduction
ratio are discarded, allowing the acoustic subsystem
of our technical debt, and alongside our increased
to perform even in noisy environments.
usage of static analyzers and linters, has greatly
reduced our chances of introducing bugs. Adoption
of proper workflows for version control and using C. Software Sub-System
a self-hosted Continuous Integration build pipeline 1) Mission Planner
has also made collaboration easier and safer.
Adopted last year, our mission planner utilizes Be-
haviour Trees (BT) and has proven to be highly
5) Status Light Module effective in defining complex behaviours for our
Previously, only one RGB LED was used as a status vehicle. We have enhanced our mission planner
indicator to provide feedback to operators outside with a Graphical User Interface (GUI) to facilitate
the pool, but was found to have limited visibility easy designing and modification of mission plans.
under daylight and required a redesign. After testing Furthermore, harnessing the abstracted nature of
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 5

BTs, we have refactored our mission plans to en- compensate for non-linear terms in the vehicle’s
able different tasks to reuse the same high-level motion dynamics.
logic, thereby simplifying the mission planning pro-
Our thrust allocator uses quadratic programming
cess.
to optimise each thruster’s command based on the
required forces, and maintains control along each
2) Machine Learning Pipeline axis of motion even during thruster saturation.
We have upgraded our perception pipeline to lever-
age the benefits of both deep-learning and tradi- III. T EST S TRATEGY
tional computer vision (CV) approaches in order A. Custom Light Module
to improve our chances of localizing the different Key areas that were considered during the current
obstacles in the TRANSDEC environment. iteration were the LED’s brightness, heat generated,
The upgraded pipeline incorporates deep-learning and meeting the power budget. The circuit for the
models such as SuperPoint [1] and SuperGlue [2] constant current driver was simulated using LTSpice
to directly estimate obstacle poses by accurately to verify the behaviour of the driver and LEDs
matching image features from our camera against (Fig. 6). Using the simulation, the values of capac-
template images provided in training. These are itors and inductors in the buck-boost circuit were
used alongside object detection and segmentation also optimized for minimal ripple current.
models such as YOLOv8 [3], and supplemented
by traditional CV algorithms like Scale-Invariant
Feature Transform (SIFT) or Perspective-n-Point
(PnP).
An advantage of this hybrid pipeline is the abil-
ity to dynamically adjust our approach based on
environmental conditions and/or the task at hand,
allowing us to optimize the perception system’s
performance for various scenarios. In general, these
improvements have enabled us to achieve more Fig. 6: LTSpice simulation of current (pink) and
precise and robust perception capabilities. voltage (green) through driver over time.

The panel of the light module is designed to ac-

3) Control System comodate up to 12 LEDs. Visibility testing was
conducted in daylight from a 30-m distance in
batches of 4, 8 and 12 LEDs. The tests were not
conducted within the main hull as the module was
in the prototype phase, but were deemed sufficient
for a first estimation. Although tests revealed that 8
Fig. 5: Control system block diagram. LEDs were sufficient, the schematic was ultimately
designed using 9 LEDs in a 3-3-3 configuration
Our tried-and-tested control system is largely un- (Fig. 7). This made the circuit layout simpler and
changed. The trajectory generator uses linear seg- allowed for more efficient buck converters to be
ments with polynomic blend to generate smooth used.
continuous paths. These trajectories have limits im-
Thermal tests were carried out using a thermocouple
posed on the velocity, acceleration and jerk of the
to measure the temperature of the LEDs at 1-
vehicle to avoid controller saturation, improving
minute intervals. At an ambient temperature of 26◦
performance for distant setpoint goals.
C, temperatures remained below 60◦ C even after
We utilize a control law partitioning scheme – a continuous operation at maximum power for 20 min.
full state feedback controller enables positional and Although the LEDs have a rated maximum operat-
velocity tracking, and a feedforward controller to ing temperature of 155◦ C, the primary concern was
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 6

Fig. 7: 3D Model for light module, LEDs in red.

Fig. 8: Measured yaw values of old (blue) and
new IMU (orange) over time.

this new heat source affecting the operation of other

thermally-sensitive sensors. C. Torpedo Improvements
The torpedoes used last year have been redesigned
On the electrical side, power consumption was mea- to be positively buoyant with reduced drag. Assisted
sured using a bench power supply and oscilloscope by Computational Fluid Dynamics (CFD) analysis
by probing the current and voltage across the LEDs in SolidWorks, our current torpedoes are redesigned
to calculate transient power draw. The ripple current for minimal drag and trajectory disturbance. CFD
was verified to lie within the margins of our power results are given in Appendix A.
budget, which is especially important as the LEDs
draw power directly from a 12-V DC converter that The torpedo launcher was also redesigned, as torpe-
only supports up to 120 W. Using a load tester does were loose and would fall out during certain
to simulate the LEDs, the current driver was also manoeuvres. This resulted in inconsistent loading
benchmarked against various resistance values to and large variances in travel distance over several
ensure its stability under various conditions. These attempts. Several designs were tested by firing at a
resistances were derived from the datasheet of the styrofoam board underwater. The results of torpedo
LED. spread are given in Appendix B. A cross profile
is also used to accommodate the new torpedoes
with round cross-section to provent rotation when
loaded.
B. IMU Comparison
It was also suspected that the moving metal compo-
In previous years, we found that the IMU was nents were causing magnetic interference, contribut-
highly susceptible to external magnetic interference. ing to IMU drift, and have thus been replaced with
In order to investigate and understand the issue in plastic parts.
detail, a test mount was printed to allow for testing
of several IMUs at the same time under the same ACKNOWLEDGEMENTS
conditions. Team Bumblebee’s development and achievements
would not be possible without the help from various
One test involves conducting a 180-degree yaw on
organisations and people. We are deeply grateful
the test mount and measuring the resultant vehicle
to Tan Annie and Cheah Kok Keong for their
yaw over time (Fig. 8). It was observed that the old
unwavering support. We also thank Hou Linxin,
IMU has a delayed response which is especially
Koh Yang Kai, Lee Shi-Kang Matthias and Ng
obvious after a large difference in yaw. This was
Cheng Yee Trina for their valuable contributions to
ascertained to be an issue with the sensor charac-
the team.
teristics and was resolved by switching to another
IMU. With this newer IMU, we note a reduction Team Bumblebee would like to express their deepest
in position error by up to 50% over a 2-minute gratitude to the sponsors (Refer to Appendix G), in-
period. cluding the Title Sponsors — National University of
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 7

Singapore (NUS), and Platinum Sponsors — Future

Systems and Technology Directorate (FSTD), DSO
National Laboratories, ST Engineering, Altium, Re-
public of Singapore Yacht Club and Wartsila.

R EFERENCES
[1] D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superpoint:
Self-supervised interest point detection and description,”
CoRR, vol. abs/1712.07629, 2017. [Online]. Available: Fig. 11: CFD results of alternative model.
http://arxiv.org/abs/1712.07629
[2] P. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich,
“Superglue: Learning feature matching with graph neural
networks,” CoRR, vol. abs/1911.11763, 2019. [Online]. situated at the front move into place, and suc-
Available: http://arxiv.org/abs/1911.11763 cessfully prevented torpedoes from falling out in
[3] G. Jocher, A. Chaurasia, and J. Qiu, “YOLO tests.
by Ultralytics,” Jan. 2023. [Online]. Available:
https://github.com/ultralytics/ultralytics

A PPENDIX A
T ORPEDO CFD A NALYSIS
The CFD simulation results of the old model
(Fig. 9) and two proposed designs (Fig. 10 and 11)
are presented here. Note that the colour scale for
fluid velocities are normalized for ease of compar-
ison.
Fig. 12: Spring-loaded slot blocking launcher.

An alternative design uses holding pins to grip the

torpedoes (Fig. 13), inspired by the alignment pins
found in NERF guns. This design requires minor
edits to the torpedo model to accommodate the
pin.
Fig. 9: CFD results of old torpedo.

Fig. 13: Holding pin launcher.

Fig. 10: CFD results of current torpedo.

Both designs have a comparable spread as shown

A PPENDIX B in Fig. 14. During tests in the pool however, it
T ORPEDO L AUNCHER S PREAD was noted the spring-loaded slot blocking launcher
The first new launcher design utilizes a spring- causes the torpedoes to veer, and so the holding pin
loaded slot blocking mechanism (Fig. 12). Blockers launcher was used instead.
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 8

Fig. 14: Spread of torpedoes on styrofoam board. (a) Holding Fig. 16: 3D model of Actuation daughter board.
pin launcher. (b) Spring-loaded slot blocking launcher.

A PPENDIX C
BACKPLANE S YSTEM
Our backplane system makes extensive use of the
Samtec PowerStrip™ connectors. Where connec-
tions to other third-party components are required,
adapter daughter boards (Fig. 17) are used instead of
designing with proprietary connectors for flexibility Fig. 17: 3D model of ESC adapter daughter board.
and preventing vendor lock-in.
Such a design also simplifies power control from a
centralized power control board (Fig. 18).

A PPENDIX D
A RCHITECTURE B LOCK D IAGRAMS
The following initialisms are specific to our internal
usage and may be useful for elucidation of our
architecture: Fig. 18: 3D model of power control backplane.

• DAQ: Data Acquisition Board (Acoustics)

• DVL: Doppler Velocity Log
• FOG: Fiber-Optic Gyroscope
• PMB: Power Monitoring Board
• SBC-CAN: Custom board bridging our CAN
network and our main SBC
• STB: Sensor and Telemetry Board
• TAB: Thruster and Actuation Board
Fig. 19: 3D model of Sensor and Telemetry daughter board.

Fig. 15: 3D model of ESC backplane.

Fig. 20: 3D model of Thruster and Actuation daughter board.

NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 9

Fig. 21: Inter-board communication architecture block diagram.

NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 10

Fig. 22: Power architecture block diagram.

NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 11

A PPENDIX E
C OMPONENT S PECIFICATIONS

Component Vendor Model / Type Specifications Custom Cost Year

/ Pur- of
chased Pur-
chase

Main Hull Samco Enterprise, Custom Aluminium — Custom $2,700 2019

Feimus Engineering Milling
Frame Cititech Industrial Custom Aluminium — Custom Sponsored 2021
Laser-cut
Battery Hull SLM Solutions Custom Aluminium — Custom Sponsored 2020
Selective Laser Melting
Floats Admiralty Diab HCP30 — Custom $4,650 2022
International
Nylon Shell 3D Print Singapore HP MJF — Custom $1,000 2022
Waterproof SubConn Inc., Assorted Micro and Peak Depth: Purchased Sponsored 2019
Connectors MacArtney Low-profile Series 300 bar
Waterproof Blue Trail SER110X Peak Depth: Purchased $380 ea 2021
Servos Engineering 10 bar
Thrusters Blue Robotics T200 — Purchased $176 ea 2021
Motor Flipsky Mini FSESC4.20 50A — Purchased $145 ea 2021
Control
High-level Raspberry Pi RPi 3 Model B+ 1.4GHz 64- Purchased $39 2019
Control bit quad-core
processor
Actuators/ In-house ABS/HP MJF — Custom Sponsored 2022
Manipulators
Battery Tattu Custom-made 4-cell 15000 mAh Purchased $200 ea 2023
battery
Battery In-house Custom-made circuit — Custom Sponsored 2022
Monitoring board
System
Power Murata UWQ-12/10-Q12PB-C 120W Purchased $52 2021
Isolator Wide-range
Isolated DC-
DC to 12V
UWE-24/3-Q12PB-C 72W Wide- Purchased $67
range
Isolated
DC-DC to
24V
Single Board Avalue ECM-TGU Intel Core Purchased $1,000 2023
Computer i7-1185GRE
GPU Nvidia Jetson Orin AGX Purchased $2,500 2022
Internal In-house CAN / Ethernet 1000kbps / Custom Sponsored —
Comm. 1000Mbps
Network
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 12

External In-house Ethernet 1000Mbps Custom Sponsored —

Comm.
Interface
FOG Fizoptika VG103S-2LND — Purchased $3,060 2021
IMU Sparton AHRS-8P — Purchased Sponsored 2019
VectorNav VN-100 — Purchased $1,500 2023
Doppler Teledyne Marine Pathfinder DVL 600kHz Purchased $16,000 2019
Velocity Log Phased Array
Camera(s) BlackFly S PoE BFS-PGE-31S4C-C 2448 × 2048 Purchased $594 2019
Gigabit Camera at 22 FPS
Hydrophones Teledyne Reson TC4013 Acoustic Purchased Legacy 2017
transducers
Sonar Oculus M750d Dual- Purchased $21,300 2019
Frequency
Multibeam
Sonar
(750KHz /
1.2MHz)
Algorithm: — — Thresholding, — — —
vision Particle filter,
Machine
learning
Algorithm: — — Multiple — — —
acoustics Signal
localisation Classification
(MUSIC),
Short-Time
Fourier
Transform
(STFT)
based Ping
Extraction
Algorithm: — — Short-Time — — —
acoustics Fourier
communica- Transform
tion (STFT),
Quadrature
Phase Shift
Keying
(QPSK)
Algorithm: — — Error State — — —
localisation Kalman Filter
& mapping
Algorithm: — — Behavior- — — —
autonomy Tree.CPP
Open source — — OpenCV, — — —
software ROS,
PyTorch
Team size — — 38 — — —
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 13

Hardware/ — — 3:1 — — —
Software
expertise
ratio
Testing time: — — 100 hours — — —
simulation
Testing time: — — 200 hours — — —
in-water
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 14

A PPENDIX F
O UTREACH ACTIVITIES
Ever since our humble beginnings in 2012, Team
Bumblebee has continued to grow, and we have
become one of the most accomplished student teams
in the maritime robotics scene. Despite this, we
remain grateful to the community and our sponsors
for their support throughout the years. We believe in
the importance of fostering new relationships, and
strive to share our knowledge and experiences as a
form of giving back to the community.
Fig. 25: A recent sponsor appreciation event.

A. Lab Visits
As part of Team Bumblebee’s public relations cam- In order to gain experience and understanding of
paign, we regularly conduct lab visits for fellow real-world challenges, our team also regularly orga-
robotics teams and enthusiasts in the field of marine nizes visits with industrial partners. ST Engineering
robotics. Through these visits, we hope to exchange is one of our major sponsors, who have graciously
knowledge and build lasting friendships. loaned us sensors and other equipment to assist in
our testing and development.

Fig. 23: Lab visit by a local pre-university institution.

Fig. 26: Industrial visit to ST Engineering.

C. Hornet Training Programme

Team Bumblebee is dedicated to fostering students’
passion for maritime robotics. This objective is
accomplished through the implementation of the
Hornet Training Program and its recruitment drive.
Our team actively engages new students by conduct-
ing sharing sessions during orientation camps and
Fig. 24: Lab visit by an international team.
setting up booths at freshman welcome talks.
The Hornet Training Program serves as an introduc-
B. Industrial Partnership and Appreciation tion to engineering and robotics. In this program,
Team Bumblebee is grateful to our industrial part- students are entrusted with the task of designing,
ners, without whom our team would not be able to building, and testing an Autonomous Underwater
achieve continued excellence. Vehicle (AUV) for the Singapore AUV Challenge.
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 15

Through this program, students are encouraged to A PPENDIX G

explore and experiment with novel designs, foster- S PONSORS
ing a spirit of innovation and creativity. We have A. Title Sponsors
recently concluded the eight iteration of the Hornet
NUS (College of Design and Engineering, Innova-
Training Program (Hornet 8.0), and have welcomed
tion & Design Programme and School of Comput-
the new members into Team Bumblebee for the
ing) — For their cash support, equipment procure-
development of BBAUV 4.1.
ment, and academic support of our project.

B. Platinum Sponsors
Future Systems Technology Directorate (FSTD) —
For cash support.
DSO National Laboratories — For cash support and
technical guidance.
ST Engineering — For loaning of equipment.
Altium — For providing software licenses.
Fig. 27: Some Hornet 8.0 members conducting a pool test. Republic of Singapore Yacht Club — For providing
a testing location.
D. Collaboration with Local Schools Wartsila — For providing equipment.
Team Bumblebee have conducted sharing sessions
at a local high school to inspire students to pursue C. Gold Sponsors
engineering as a career. The team shared about their Avetics, Fugro, Kentronics, MacArtney, SBG Sys-
experiences at RoboNation’s competitions (Robo- tems, SLM Solutions and Würth Electronik.
Sub, RobotX) and the development and testing of
vehicles leading up to it.
D. Silver Sponsors
Bossard, Festo, Samtec, Solidworks, Southco and
Sparton.

E. Bronze Sponsors
Blue Trail Engineering, Edmund Optics, Lionsforge,
Pololu and TGN Technology.

Fig. 28: Collaboration with a local high school.