NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 1

RoboSub 2022 Technical Design Report

National University of Singapore (Bumblebee Autonomous Systems)

Amadeus Aristo Winarto, Ananya Agarwal, Chen Sirui, Chew Zhi En Samuel Joshua, Chin Zheng Hao,
Goh Jie Xuan Delvin, Gokul Rajiv, Gowthaman Aravindan, Hashir Zahir, Ho Wei Zong Jasper, Hou Lin Xin,
Justin Foo Guang En, Kaitlyn Ng Ke Yi, Lee Chan Wai, Lee Shi-An Matthew, Lee Tze Han, Li Po Hsien, Lim Sheng Wei,
Lu Sicheng Isabella, Lum Chang Xin Shawn, Manzel Joseph Seet, Marvin Pranajaya, Michael Jervoso,
Ng Cheng Yang Titus, Ng Xing Yu, Ng Yong Jie, Ng Zhia Yang, Nguyen Minh Nguyen, Niu Xinyuan, Png Qun Shen,
Quek Wei, Rani Karthigeyan Rajendrakumar, Seow Alex, 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 2022, Team Bumblebee’s strategy

involves deploying the BBAUV 4.0 to efficiently complete
all tasks. The mechanical design of the BBAUV 4.0
optimises for space and weight, yielding vastly superior
manoeuvrability while eliminating the weight penalty of
the BBAUV 3.99. Electrical work centred on testing and
integrating new sensors, as well as improving overall
system reliability. Software development focused on an
updated behaviour-tree mission planner, improvements to
the controls system, and driver development for our new
sensors. Physical testing was conducted at a steady pace,
and was supplemented using hydrodynamic simulations to
tune the AUV’s control systems.

I. C OMPETITION S TRATEGY Fig. 1: 3D model of the BBAUV 4.0.

For RoboSub 2022, we plan to deploy our BBAUV
4.0 (Fig.1) to complete all competition tasks. Taking
advantage of the lack of physical RoboSub compe- intensive upgrading and tuning, we are confident
titions these past 2 years, we have iterated tirelessly that the BBAUV 4.0 can complete all competition
on our newest vehicle, and have achieved feature tasks faster than BBAUV 3.99 ever could.
parity with our older BBAUV 3.99.

A. Competition Vehicles B. Course Strategy

During RoboSub 2021, our team strategy dictated
For our task strategy, we employ a sensor fusion
that both the BBAUV 3.99 and BBAUV 4.0 be
approach; our vision pipeline combines sonar im-
deployed. However, after extensive testing, the team
ages with machine learning (ML) object detection to
has opted to deploy only BBAUV 4.0 this year in
accurately localise and identify task objects, like the
light of its superior overall performance.
buoys in Make the Grade or the torpedo openings
A much smaller and lighter vehicle, BBAUV 4.0 in Survive the Shootout. Thus our strategies for
was purpose-built for tackling RoboNation compe- these, plus With Moxy and Choose Your Side, are
titions unlike the BBAUV 3.99, which falls behind similar: approach the targets using localisation from
in both speed and weight. In addition, limitations for our vision pipeline before performing task-specific
further improvement due to legacy issues inherent actions. For final alignment and adjustments, we ex-
to BBAUV 3.99 led us to shift our focus to BBAUV clusively use ML object detection, as our sonar can-
4.0 as our primary competition-ready vehicle. After not reliably detect objects at close proximity.
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 2

The Collecting and Cash or Smash tasks are also

handled exclusively by ML object detection, since
we are unable to use the sonar for bottom-facing
tasks. We track the centroids of the detected ob-
jects, precisely adjusting the AUV’s alignment using
our highly accurate Doppler Velocity Log, together
with a revamped controls system. Since both tasks
require grabbing actuation, we also designed a ver-
satile grabber capable of performing both tasks, yet
small enough to fit within our size constraints (see
Design Creativity).
Fig. 2: Internal layout of the main hull.
Bringing together these task-specific strategies is
our mission planner. Previously, we used a Finite
State Machine (FSM)-based mission planner due to The new octagonal mounting frame of BBAUV 4.0
the abundance of existing resources and its relative allows for more actuation modules to be mounted
ease of implementation. However, we found that while still being protected. Carrying handles were
modifying states in an FSM-based planner is error- also added for ease of operation and handling.
prone due to the complex transitions and internal
states of the FSM. This is further exacerbated by
high-stress conditions like the 15-minute window 2) Design of Battery Hull
of a run. As mentioned in last year’s paper, we Another creative aspect of our new AUV is the
have since moved to a Behaviour Tree (BT)-based battery hull, manufactured with novel 3D metal-
planner. After a year of extensive usage, we have printing technology; we also increased the rigidity-
identified its flaws and developed a second, im- to-weight ratio by embedding lattices in the walls
proved version which will be discussed below. and base (Fig.3). Unlike traditional methods, 3D
printing allows for tight corners, giving our battery a
snug fit. Together, these factors significantly reduce
II. D ESIGN C REATIVITY
the vehicle’s weight and volume.
A. Mechanical Sub-System
1) Design of Main Hull
We chose a rectangular hull for BBAUV 4.0 to get
efficient packing for internal components and elec-
tronics. Finite Element Analysis was used to ensure
our design could withstand the 3 bars of pressure
expected during operation. A centre divider wall Fig. 3: Isogrid layer of the battery hulls.
within the hull gives rigidity while isolating elec-
trically noisy components from the sensors. The main and battery hulls are directly connected
with right-angled SubConn Low Profile connectors,
A fibreglass float fits on the outside of the hull, reducing the length of cabling required.
making it much more resilient to external impacts
while maintaining superior manoeuvrability. A 3D-
printed shell was designed to fit below the fibreglass 3) Design of Actuation Systems
floats (Fig.1), including cavities for buoyancy tuning The bulky newton gripper of years past was replaced
via float insertions. More floats on the top cover pro- with a smaller claw, utilising bevel gears and a
tect the sonar while adding buoyancy. These exterior stepper motor (Fig.4); the claws tuck neatly under
coverings are designed not to interfere with opening the hull to reduce their footprint. We can grab
the hull, enabling faster access during testing. This both vertical and horizontal objects with the same
protective layer gave us the confidence to bump into claw, completing Collecting and Cash and Smash
the buoy for Make the Grade. by picking up the lid and bottles respectively.
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 3

under-voltage protection system is also integrated;

electronics are disabled to prevent damaging the
batteries if their voltage drops critically.
Our batteries are hot-swappable by means of a load-
balancer between them; the vehicle can stay running
Fig. 4: BBAUV 4.0 Grabber and Torpedo Launcher. during swaps, minimising operational downtime.
Sensitive components are also protected from elec-
trical noise via galvanic isolation between internal
Our ball dropper and torpedo launcher use the electronics and inductive loads.
Bluetrail underwater servo (Fig.4); while larger than
our old custom servos, it markedly enhances the
reliability and depth rating of both actuators. 2) Backplane System
Our backplane system lets the electronics sit in
B. Electrical Sub-System the hull, instead of being mounted on the end-cap.
1) Design of Electrical Architecture This gives the electrical system flexibility to use
multiple backplanes, separating the high- and low-
There are two main communication channels used in level circuitry. These backplanes are easily accessed
our electrical architecture: Controller Area Network by opening the top lid, and individual boards can
(CAN) and Ethernet. Ethernet is used for systems be replaced in a plug-and-play fashion. This lets us
requiring high bandwidth, while CAN is used for perform maintenance without being obstructed by
communication between embedded systems. other components.
Another benefit of separate backplanes is that each
system can function independently; this allows any
failures to be quickly traced, easing debugging. It
also isolates electrically sensitive equipment, such
as our IMU and acoustics boards, from the elec-
trical noise generated by the other electronics, thus
increasing sensor measurement accuracy.

Fig. 5: Communication architecture block diagram.

3) Acoustic Signal Processing
The BBAUV 4.0’s acoustic subsystem uses an au-
tomated programmable gain amplifier that gives us
a uniform amplitude of the incoming ping, enabling
more reliable measurements. This results in both a
reduction in signal clipping, and consistent output
regardless of distance from the pinger. Scaling of
the amplifier’s gain factor is done by comparing the
ratio between the optimal and the current amplitude
of the ping, which is done on the Data Acquisition
Fig. 6: Power architecture block diagram. (DAQ) board. To increase reliability and reduce
false positives, we check the signal-to-noise ratio
Our custom-designed Power Monitoring Board on the extracted ping, weighted towards known
reads the battery charge via a fuel gauge IC, letting sources of noise. Furthermore, ping extraction is
us show the estimated operational time on the vehi- performed using short-time Fourier transformation
cle’s display. Additionally, this monitoring enables with a dynamic thresholding method, allowing the
our control board to selectively prioritise and disable ping to be accurately extracted with low latency
systems during low-power scenarios. Finally, an even in noisy environments.
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 4

instead of repeatedly defining them in configuration

files. In addition, nodes interfacing with ROS topics,
services, etc. were rewritten to better fit the ROS 1
Fig. 7: Acoustics flow diagram.
paradigm.

2) Machine Learning Pipeline

4) Fibre-optic Gyroscope We completely revamped our machine learning
A new fibre-optic gyroscope (FOG) was installed pipeline this year, making it faster and more modu-
to complement our IMU, which has trouble when lar. Instead of a mixture of Python 2 and Cython, we
around ferrous objects, such as pool walls. The now use Python 3 for easier maintainability without
new sensor has a low rate of drift, and is much sacrificing speed. The new modular design also
more impervious to external electromagnetic inter- allows additional models to be added in the future,
ference than our IMU. Using our existing sensor- eg. models with rotated bounding boxes, which
fusion pipeline, we are able to combine the absolute would be useful for Cash or Smash in detecting
heading given by the IMU together with the relative the lid orientation. We also refactored our existing
angular velocity reported by the FOG. YOLOv5-based architecture to allow the use of
TensorRT, letting us leverage device-specific opti-
C. Software Sub-System misations, maximising our Nvidia Jetson Xavier’s
potential and gaining a 2x FPS increase.
1) Mission Planner
As previously mentioned, we switched to a BT- 3) Control System
based mission planner last year for its flexibility and
maintenance ease. Based on the BehaviorTree.CPP
package, it has many advantages over a traditional
FSM planner, like being able to quickly interpret
the AUV’s behaviour. The linear structure and well-
defined node transitions also make it easier to Fig. 8: Control architecture.
configure robot behaviours simply by repositioning
We implemented a trajectory generator to create
nodes, without accounting for innumerable state
smooth continuous paths for the AUV to follow.
transitions. This lets our planner accommodate more
The trajectories are calculated using linear segments
situations, whereas an FSM would require exponen-
with polynomic blend, with limits on the maximum
tially more states and transitions to compete.
velocity, acceleration, and jerk of the vehicle. This
BT planners also give us a high level of composabil- improved performance for distant setpoint goals by
ity and abstraction: both nodes and subtrees can be avoiding controller saturation.
reused, allowing us to define increasingly complex
The control system makes use of control law par-
robot behaviours easily. Furthermore, our planner
titioning, including a full state feedback controller
allows for tree structure and parameter configuration
which enables positional and velocity tracking, and
at runtime without recompilation, saving precious
a feed-forward controller to compensate for non-
time in a competition setting.
linear terms in the vehicle’s motion dynamics.
After a year of testing by our software team, we
Our thrust allocator uses quadratic programming
discovered several flaws with our initial implemen-
to optimise each thruster’s command based on the
tation, which we hope to rectify with our new
required forces, and maintains control along each
version. The first version was based on the ROS
axis of motion even during thruster saturation.
2 navigation stack, and suffered from being overly
generic. Adding new nodes involved many error-
III. E XPERIMENTAL R ESULTS
prone steps, while the extensive use of configuration
files proved cumbersome since many options were A. Thruster ESC
not relevant for us. Our new planner greatly simpli- During initial testing, the vehicle’s heading was
fies this by loading nodes on a per-package basis, found to oscillate heavily during forward motion.
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 5

The issue was found to be caused by the latency of thus activating the switch by pulling it to ground.
the ESCs used to control the T200 thrusters. To increase reliability, the kill switch circuitry was
changed to use an NMOS gate that needs to be
Three models of ESCs were tested: Tekin RX8
pulled up to 5V to activate.
Gen2, Blue Robotics Basic ESC, and Flipsky Mini
FSESC4. The Tekin ESC initially designed for the
BBAUV 4.0 had the largest latency at 300 ms, C. FOG Calibration
the Blue Robotics ESC yielded 130 ms, while the Calibration data was not available to us for the
Flipsky had a latency of only 5.3 ms. Since our FOG sensor, so we had to perform our own in-lab
controls system runs at 20 Hz, the delay of the Tekin calibration; we gathered readings from the sensor
and Blue Robotics ESCs were very significant at over a range of rotations and temperatures, then
3-5 timesteps, greatly degrading the AUV’s perfor- plotted them in MATLAB. We then performed a
mance. surface fit on the data to get a polynomial equation
that allows our software to convert the voltage
The modules for each ESC model were tested with
output from the sensor into angular velocity.
an Arduino MEGA and a CAN Shield. With an
oscilloscope, the latency was measured from the
trigger signal edge till the time when the ESC output
changed.
For the Blue Robotics ESC, a delay of approx-
imately 130 ms was observed. This large delay
between the input signal (CH1 in yellow) and the
thruster output changing (CH2 in blue) causes os-
cillations when tuning the controls system.
Fig. 10: MATLAB plot of FOG sensor reading data

ACKNOWLEDGEMENTS
Team Bumblebee’s development and achievements
would not be possible without the help from various
organisations and people. The team would like to
express their deepest gratitude to the sponsors (Re-
Fig. 9: Oscilloscope data for Blue Robotics ESC (left) and Flipsky
Mini FSESC4 (right); note 25 ms vs 100 µs time division. fer to Appendix C), including the Title Sponsors
— National University of Singapore (NUS), and
Through our testing, we settled on the Flipsky Platinum Sponsors — DSO National Laboratories,
Mini FSESC4 with its relatively quick response of and Future Systems and Technology Directorate
500 µs. After switching over to the Flipsky Mini FS- (FSTD). In addition, the team would also like to
ESC4, the performance of the BBAUV 4.0 greatly thank Sport Singapore and the Republic of Singa-
improved, reducing the heading oscillation during pore Yacht Club for their continuous support.
forward motion.
R EFERENCES
B. Kill Switch Interference [1] T. I. Fossen, Handbook of marine craft hydrodynamics and
motion control. Hoboken N.J.: Wiley, 2021.
We realised that there was some interference with
the kill switch in certain pool environments, due
to the metallic body of the AUV 4.0 chassis being
grounded. Initially, we were attaching a PMOS
switch to the kill switch; in some cases, the con-
ductivity of the pool water was high enough to
connect the kill switch contacts to chassis ground,
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 6

A PPENDIX A
C OMPONENT S PECIFICATIONS

Component Vendor Model/Type Specifications Custom/ Cost Year

Pur- of
chased Pur-
chase

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

Enterprise, Milling
Feimus
Engineering
Frame Cititech Custom Aluminium — Custom Sponsored 2021
Industrial Laser-cut
Battery Hull SLM Custom Aluminium — Custom Sponsored 2020
Solutions Selective Laser Melting
Floats Admiralty Diab HCP30 — Custom $4,650 2022
International
Nylon Shell 3D Print HP MJF — Custom $1,000 2022
Singapore
Waterproof SubConn Assorted Micro and Peak Depth: 300 bar Purchased Sponsored 2019
Connectors Inc., Low-profile Series
MacArtney
Waterproof Blue Trail SER110X Peak Depth: 10 bar Purchased $380 ea 2021
Servos Engineering
Thrusters Blue T200 — Purchased $176 ea 2021
Robotics
Motor Flipsky Mini FSESC4.20 50A — Purchased $145 ea 2021
Control
High-level Raspberry Pi RPi 3 Model B+ 1.4GHz 64-bit quad- Purchased $39 2019
Control core processor
Actuators/ In-house ABS/HP MJF — Custom Sponsored 2022
Manipulators
Battery Tattu Custom-made 4-cell 16000 mAh Purchased $120 ea 2019
battery
Battery In-house Custom-made circuit — Custom Sponsored 2022
Monitoring board
System
Power Murata UWQ-12/17-Q48PB-C 204W Isolated 24V- Purchased $52 2019
Isolator 12V
UVQ-24/4.5-D24P-C 108W Isolated 24V- Purchased $67
12V
Single Board AAEON GENE-KBU6 Intel Core i7-7600U Purchased Sponsored 2019
Computer BIO-ST03-P2U1 Intel i210
GPU Nvidia Jetson Xavier AGX Module Purchased $999 2019
Internal In-house CAN/Ethernet 1000kbps/1000Mbps Custom Sponsored —
Comm
Network
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 7

External In-house Ethernet 1000Mbps Custom Sponsored —

Comm
Interface
FOG Fizoptika VG103S-2LND — Purchased $3,060 2021
IMU Sparton AHRS-8P ±1.2 Gauss Purchased Sponsored 2019
Doppler Teledyne Pathfinder DVL 600kHz Phased Array Purchased $16,000 2019
Velocity Log Marine
Camera(s) BlackFly S BFS-PGE-31S4C-C 2448 × 2048 at 22 FPS Purchased $594 2019
PoE Gigabit
Camera
Hydrophones Teledyne TC4013 Acoustic transducers Purchased Legacy 2017
Reson
Sonar Oculus M750d Dual-Frequency Purchased $21,300 2019
Multibeam Sonar
(750KHz/1.2MHz)
Algorithm: — — Thresholding, Particle — — —
vision filter, Machine learning
Algorithm: — — Multiple Signal — — —
acoustics Classification (MUSIC),
localisation Short-Time Fourier
Transform (STFT)
based Ping Extraction
Algorithm: — — Short-Time Fourier — — —
acoustics Transform (STFT),
communica- Quadrature Phase Shift
tion Keying (QPSK)
Algorithm: — — Error State Kalman — — —
localisation Filter
& mapping
Algorithm: — — BehaviorTree.CPP — — —
autonomy
Open source — — OpenCV, ROS, PyTorch — — —
software
Team size — — 35 — — —
Hardware/ — — 3:1 — — —
Software
expertise
ratio
Testing time: — — 100 hours — — —
simulation
Testing time: — — 200 hours — — —
in-water
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 8

A PPENDIX B
O UTREACH ACTIVITIES
Ever since our humble beginnings in 2012, Team
Bumblebee has continued to grow, and we have be-
come one of the most accomplished student teams in
the maritime robotics scene. Despite this, we remain
grateful to the community and our sponsors, who
have supported us throughout the years. In order to
bolster our relationship with the community, Team
Bumblebee strongly believes in sharing our knowl- Fig. 12: Industrial visit to SLM Solutions.
edge and experiences with the community.

A. Lab Tour and Sharing Sessions

As part of Team Bumblebee’s public relations cam-
paign, the team extended invitations to various in-
ternational teams for visits to our lab, to exchange
knowledge and build lasting friendships.

Fig. 13: Collaboration with a local secondary school.

C. Collaboration with Local Schools

Team Bumblebee is collaborating with a local sec-
ondary school to conduct robotics lessons to inspire
the students between from ages 13–16. The program
aims to teach the students the basics of AUVs, while
Fig. 11: Lab visit by a professor from Florida.
providing guidance for them to design and build
Despite the COVID-19 pandemic, we have received their very own AUV.
multiple emails from groups interested in starting
their own robotics team. We have engaged them D. Recruitment of New Members
enthusiastically, and hope to meet them in the future
at competitions.

B. Industrial Partnership and Appreciation

Industrial partners are essential for the sustainability
of Team Bumblebee. Without their support, our
team would not have been able to achieve or
sustain excellence. Therefore, industrial visits are Fig. 14: Online recruitment session.
organised regularly with industrial partners to gain
In order to engage new students starting their uni-
experience and first-hand exposure to real-world
versity journey, an online recruitment drive was held
challenges.
as part of the NUS Engineering Life fair, E-genium.
SLM Solutions is one of our latest sponsors, who This allowed Team Bumblebee the opportunity to
have assisted us in manufacturing the battery hulls reach out to a wide audience of potentially inter-
for our new BBAUV 4.0 using metal 3D-printing as ested freshmen, giving us the chance to entice them
discussed above. to join the team.
NATIONAL UNIVERSITY OF SINGAPORE (BUMBLEBEE AUTONOMOUS SYSTEMS) 9

E. Hornet Training Program A PPENDIX C

S PONSORS
A. Title Sponsors
NUS (College of Design and Engineering, Innova-
tion & Design Programme and School of Comput-
ing) — For their cash support, equipment procure-
ment, and academic support of our project.

B. Platinum Sponsors
Future Systems Technology Directorate (FSTD) —
Fig. 15: Team Hornet working on their Hornet 7.0 AUV. For cash support.
DSO National Laboratories — For cash support and
technical guidance.

C. Gold Sponsors
Fugro, Festo, Cititech Industrial Engineering, Ken-
tronics Engineering, Würth Electronik, AAEON
Technology, SLM Solutions, SBG Systems, and
Avetics.

D. Silver Sponsors
Fig. 16: Team Hornet testing their Hornet 7.0 AUV.
Bossard, SolidWorks, MathWorks, Southco,
Since its inception 7 years ago, the Hornet Train- Samtec, and Sparton.
ing Program has evolved into a staple element of
training for the freshmen in our team. Through this E. Bronze Sponsors
program, we provide new members a platform to Edmund Optics.
build an AUV to compete in the Singapore AUV
Challenge. Our main objective is to challenge the
freshmen to explore and implement bold designs F. Supporting Organisations
instead of replicating what others and their prede- Republic of Singapore Yacht Club and Sports Sin-
cessors have done. gapore.