Intro
Over the school term of Summer 2020, I teamed up with 3 other Engineering Physics students to design a robot that could retrieve empty soda cans and dump them into a bin. The twist was that we were in a global pandemic, which meant that we needed to build the robots entirely from home with our own tools.


What
The robot is built entirely from scratch. We designed the chassis to mount all our electronics while maintaining structural integrity, despite the use of plastic. I prototyped a mechanically actuated back door that doubles as a ramp to release the cans into a bin. After rounds of rapid prototyping with household items, I ended up with a reliable and elegant solution.


We painstakingly built H-bridges to enable both-way driving and to protect the microcontroller from noise. We also equipped our robot with stereo IR sensors and wrote control code based off of PID for autonomous line tracking.

Devpost Link
Github Link

Intro
This was a project made for NWHacks 2021 at UBC! It consists of a Discord bot as a front end for the user to interact with, a Flask API and a Tensorflow Lite CNN model backend hosted by a Raspberry Pi, and photo storage hosted in an instance of an AWS S3 storage.


What
I worked on the neural net and image processing for this project. Using the OpenCV library, I processed images of a person sitting to remove all the details except for the person's silhouette. These processed images would be labelled manually (good posture/bad posture) and then collected into a dataset.

Image processing to remove details for better CNN training data.


With Keras, I trained a CNN (convolutional neural network) model on half of my dataset. To verify the model, l tested the accuracy of my model's predictions using the other half of the dataset it hasn't seen before.

Finally, the model was compressed to be compatible with Tensorflow Lite, which is designed to be run on less powerful computers such as the Raspberry Pi. With the model now running in the Pi, the Pi could take pictures of your posture periodically and predict whether you have bad or good posture!

Why
With the continued presence of COVID-19 in 2021, many students and professionals are now forced to study/work/be productive at home. One problem that this working environment has created is the increased time spent sitting in front of a desk. We created this Discord bot to remind the growing population of people working from home to protect our backs against the damaging effects of sitting with bad posture and to maintain a healthy work-from-home environment.

Intro
In this project, my teammate and I explored the challenges of developing an autonomous driving agent in a timed competition. The agent was evaluated on its ability to obey traffic laws and recognize license plates of parked cars in a simulated Gazebo environment. Limited to a single front-facing camera, we applied both classical and machine learning techniques to use an image feed to navigate, avoid collisions and recognize license plates while keeping within the boundaries of the road.

After 3 months of hard work, my teammate and I scored 3rd place out of 20 in our class competition, with a perfect score on the license plate recognition!

What
I focused on the traffic detection and license plate reading for this project.

License Plate Recognition
The license plate recognition started off with image processing of the front camera output of our car. Using the OpenCV library, I detected the location and orientation of the license plates and used a perspective transform to extract the license plate.

View from the car's camera and the extracted license plate.
          

The license plate is then fed into CNN models (convolutional neural network) trained on thousands of license plates to predict the characters on the license plate.

After lots of tweaking, I achieved a 99.7% accuracy on our tests, and a 100% accuracy during the competition round.

Traffic Detection
The traffic detection algorithm was implemented with classical computer vision techniques. The algorithm takes 2 frames from the car's camera and takes the difference in light intensity between them. Then, some transforms were applied to clean up noise and isolate movement of objects, and it worked surprisingly well! The traffic detection will trigger if the object movement detected was above a threshold.

The image processing steps used in the traffic detection algorithm.

WIP!

WIP!

Intro
UBC Solar is a design team with a goal to build a 100% solar-powered car to race in the Formula Sun Grand Prix and the American Solar Challenge. As a member of the Suspension and Steering Team, I designed the gas and brake pedals for the 3rd generation car for the team.

What
The project included the design and FEA modelling for the gas pedal, brake pedal and the mounting system to the chassis. Additionally, I selected the hydraulic fittings for the brake lines that connect to the calipers on the wheels.

Why
The 2nd generation car used a simpler floor-mounted pedal design. We couldn't use this same design in the 3rd gen car because our new wire-frame chassis was much smaller in the front and had a different geometry. After preliminary design, I determined that mounting the pedals on a panel was the best solution.

This panel proved to be a challenge to design, as it needed to withstand bending and torsional forces caused by full force slamming on the brake pedals during an emergency. A lot of work went into FEA on the panel and redesigning it until a safety factor on the max stress levels were reached.

Brake Pedal Assembly


Gas Pedal Assembly


FEA of the Panel

WIP

Jmacro Rev 1
During the pandemic I was using a drawing tablet and OneNote to take notes for online classes. However, it was challenging to switch between different pens on the desktop UI, and the keyboard shortcuts were marginally better as I was getting finger cramps from repeatedly typing Alt-1 and Alt-2. What I needed was a dedicated macro pad, and since I had access to the Vancouver Hack Space for my co-op job, I decided to design and make my own!


Specs:
Plate and case: Laser-cut acrylic
Electronics: Hand-wired, Arduino Pro Micro
Switches: Gat Ink Black V2's


Dactyl Manuform 5x6 Imgur Post
After the success of my macro pad, I decided to do something bigger. I was experiencing some wrist pain, and was irritated by having to reach my right hand all the way to the left to type on my standard full-size keyboard. I then discovered the Dactyl Manuform and knew I needed to make it! I borrowed a 3D printer from the Hack Space to print the case, then diligently sanded and spray-painted the outside to get a smooth finish. The hand-wiring was also a fun but long journey.


Specs:
Plate and case: 3D printed PLA, primed and spray-painted, laser-cut acrylic bottom plate
Electronics: hand-wired, 2x Pro Micro, RJ9 cable, USB-C port
Switches: Boba U4
Keycaps: SA(R3) Ice Cap


Jmacro Rev 2
I wanted to dabble into PCB design, and also wanted to add some more features to my macro pad, and so I designed the Jmacro Rev2! The PCB comes equipped with 2 rotary encoder slots and even support for RGB backlighting.



Specs:
Plate and case: FR-4 PCB material, acrylic bottom
Electronics: Custom PCB, Pro Micro, 2x Rotary Encoders
Switches: Gateron Milk Yellow