It all started with receiving a Lego robotics kit for Christmas as a child, ever since then I have become enthralled by everything robotic. I chose my studies to turn my passion into a research career. During my bachelors and masters I have been involved in various internships and research projects around the world which I believe has given myself a solid research foundation as well as an international perspective. My research interest is incorporating machine learning techniques into the area of robotics research to improve upon classical control theory approaches. I am particularly interested in deep reinforcement learning, imitation learning and computer vision. I would like to further myself researcher career and hopefully pursue a doctoral course in the future.
To investigate whether an agent trained in simple 2D environment can be transferred to a real simulated robotic agent. The agent will only have access to a local egocentric occupancy map and thus the system will be partially observed MDP. The intermediate reward will be the reduction in map entropy at each step. The simulated robotic agent will be equipped with laser range scanner and will use graph-SLAM with the back end producing an occupancy map for which the trained agent has access to.
Many driving systems still require heuristic parameter tuning for control, and it is desired to automate this parameter tuning process. The difficulty of this problem is that there often exist multiple solutions that give high returns. Since a designed reward function is often not optimal in practice, the solution that gives the highest return may not be the optimal solution. For this reason, it is often necessary to check the multiple candidates of the solution and see which one is the best in the actual system. To address this issue, we are developing a hierarchical gradient-free optimization method that can find multiple modes of the return function. This framework is capable of proposing multiple solutions to users and letting users to choose the one that satisfies users' requirements.
Supervisor Website
Report
Leader-follower formation control is one of the most versatile and scalable formation
control architectures. This study iterates through the necessary steps to achieve
leader-follower formation control on a pair iRobotCreate robotic platforms through
the use of visual feedback provided by a Microsoft Kinect. Image identification and
measuring algorithms are developed to achieve the visual feedback requirements
for the selected formation controller. The entire system is accurately modelled
including the iRobot’s motion models and the sensor noise models. An EKF state
estimator is designed and shown to produce accurate states from noisy measurements
on the implemented system. A Lyapunov proven error based controller is selected
from literature and shown to drive the relative errors to zero. The entire system
is simulated and shown to accurately keep the formation amidst sensor noise. The
implementation results show the validity of the controller and state estimators with
the exception of a final error in the leader’s state in a certain formation scenario.
Undergraduate Thesis
GPA: B+
First Class Honours Pass for Thesis (>75%)
Dean's list for all undergraduate years (GPA>70%)
GPA: 72%
Math: 89% Physics: 72% IT: 82%
The Chemical Eng. Dept. have developed an electrical tomography unit which can analayse the flowrate, consisteny etc. of a mixtuxe flowing though a pipe line by attached a circular ring of electrodes. At that moment the system was running on an expensive labview microprocessor so the task of the project was to implement the complicated coding onto a simple inexpensive and portable Raspberry Pi 3. The program was implemented in C, the main challenge was the timing of the signals as they had to be in the order of 1e-3 and since the Pi 3 does not have a timing module an extrenal timing module must be incorporated. However, I managed to set up the code to use the Pi 3's internal cpu clock with interupts to run the very complicated timing sequence coded in labview.
I was a tutor to a third year Microprocessors course. My role was to assist and grade student's C code. After being a tutor for a year I was promoted to a teaching assist for a fourth year signals and systems course. My role was to lecture in the event the professor could not, grade tests and help conduct the exam.
A company based in Cape Town called Club Electron teaches children between grades six and eight the incredible world of robotics through Arduino based robots. During my undergraduate I was a tutor to grade eights and simply taught Arduino C and basic electronics. After one year I was promoted to help design the lessons and thus the circuitry needed. I was responsible for designing and printing circuit boards as well as teaching the lessons for said boards.
Python
C
Labview
Robotic Operating Software (ROS)
Matlab
Apart from my passion for robotics, I enjoy most of my time being outdoors. In the winter, I am an avid snowboarding whilst during the warmer months I enjoy surfboarding and hiking. I went hiking in Nepal in 2018 for three weeks and even though it was tough I would love to return.
When forced indoors, I would say my biggest hobby is reading. I tend to enjoy non-fiction with the occasional fantasy series to get yourself lost in another world.
Favorite books: 1984, George Orwell. Guns Germs and Steel, Jared Diamond. Mirror of the gods, Malcolm Bull.
I think I am an aspiring chef and I spend a large amount of my free time exploring the latest technology advancements in the machine learning and robotics world.