Kelvin's Portfolio

Lab Automation

Low-cost, open-source hardware tools for self-driving labs [2024-Present]

At the Acceleration Consortium, we are pushing the technology of self-driving labs where AI and automation is used for accelerating materials discovery. The core concept of self-driving labs is using a closed-loop design-build-test-analyze cycle where AI predicts the next set of experiments based on prior results, and automation enables higher throughput and reproducible data collection.

The mission I care most about here is open science: developing low-cost, scientist-friendly hardware automation tools. Commercial lab automation equipment is often not ideal. First, it's out of the budget for many research labs. Second, they are often closed and require users to lock into their ecosystem of products.

Robot arms are strongly tied to lab automation, but that's not the only tool available for automating a lab. That being said robot arms are awesome and here are a few of my side projects with robot arms. In the slides following, I show other automation tools being developed.

This was an open-source kit from Annin Robotics that was built. It probably isn't the safest option for lab use so I taught a duck to quack instead.

The robotics team here works on projects like teleoperation (mini robot arm to control an actual robot arm) to teach general robot skills for lab automation. Here is a teleoperation leader arm I built by 3D printing a scaled down 1:2 version of a UR5 arm. This was designed based on FACTR for a different robot arm.

While making the teleoperation hardware, I thought about scaling it further to make something you could stick on your fridge. I put this design on MakerWorld

Overpriced Lab Automation Equipment

To give an idea of how overpriced lab automation equipment is, take this plate hotel. This is 10 storage shelves for well plates, used with a robot arm to pick and place plates.

This thing costs $2,490 CAD. WHY?!? This is nothing more than a structural frame. It's sold under lab automation equipment because that's its intended use, but is it not just 10 shelves?

With a BOM of $100 CAD and some 3D printed parts, we made our own version that's functionally equivalent and 25x cheaper. It takes a day of 3D printing, a day for parts to deliver from McMaster Carr, and 1 hour to assemble.

This plate hotel "design" is reconfigurable, which is necessary in workflows where deck layouts and labware needs vary greatly. The image above on the right shows the reconfigurability concept. This "plate hotel" is used for storing vials on top and doser heads on the bottom. The 3d printed trays are different to adapt for different labware, but the core concept of stacking labware vertically is preserved.

For even more well plate storage capacity, there are products like this plate carousel. In 2025, I got a quote for this for €25,000- €30,000. The additional hardware is adding a motorized turntable.

We made our own version of a plate carousel, which we call a plate RESORT (Robotic Engineered Standardized Organization of Rotating Trays). The reason why this works in tandem with a robotic arm is that only one of these hotels (columns) is accessible for the robot arm to interact with. In the image below, an 80-plate capacity plate RESORT is installed in one of the labs, located at the end of a SCARA robot-on-rail system.

CNC 4 Science + CV/ML Error Handling

The CNC exemplifies the open science initiative at the Acceleration Consortium. Many lab tasks can be automated by moving a tool on an XYZ gantry, and a $300 CNC machine from Amazon is one of the simplest, robust, easy-to-use gantry that can be purchased off-the-shelf. It's been used in 3 labs for different applications such as liquid handling, nanoindenter characterization, viscosity testing.

One application to highlight here is automated vial capping. Caps that are normally screwed on are a surprisingly difficult task to automate. The cheapest commercial system that we could find cost $7,500. Below, a CNC machine with a vacuum gripper is used to pick and place caps and vials for automated capping.

Going from a single working prototype to something you can rely upon 100% in an autonomous workflow is a jump. Especially with low-cost DIY tools, there's a tradeoff with robustness. To address this, we apply computer vision and machine learning for error handling.

Step 1: Hardcode a workflow

Many DIY tools we build operate like this where it works 80% of the time. Then, we would find ways to increment robustness such as improving the hardware design or adding sensors. Instead, I pointed a camera at it and learned how it failed.

Step 2: Image Classification

Using Label Studio, we annotated the images to identify the 4 slot positions as well as the objects (cap, vial, empty) in each slot. A RT-DETR image classification model is trained. For each image, we can count the number of each object in each slot.

Step 3: Error Handling

With the ability to classify and count objects, we can take an inspection photo after each run to compare the actual state from the image with the expected state. If there is a mismatch, our rules would determine how we handle this error (retry, continue, or abort).

Simple Actuators, Many Use Cases

Robot arms are so strongly associated with lab automation, when sometimes a module with an actuator with one degree of freedom can be simpler, more robust, more compact, yet achieve the same results with a robotic arm. Here are a couple of examples, but there are many more where actuators have been utilized to automate modules.

Autosealing OER Reactor

This existing project is highly automated, but there is a manual labour step to replace a substrate plate sandwiched between the top of the reactor (liquid chamber) and the bottom of the reactor (electrical contacts). To further automate this process, I added linear actuators to the top of the reactor, controllable by applying a voltage. By doing this, it removes the manual swapping of substrate. Instead, a robot arm can insert new substrates and seal it with the actuators.

Fumehood Sash Actuator

Humans work in fumehoods with the sash down. Robots need it up. After a robot arm damaged a fumehood by crashing into a lowered sash, I was asked to automate it. The actuator raises the sash when the robot needs access and lowers it when it doesn't — balancing safety, energy efficiency, and robot-human collaboration.

3D Printing and DIY tools for lab automation

Somehow at the Acceleration Consortium, I'm associated with 3D printing. In 2016 when I was in undergrad, I was disinterested with my career outlook. Then I had an internship where I got to play with a 3D printer for the first time. The experience of designing a CAD model, then staring at a 3D printer bringing my design to life just made me flip a switch. It was empowering, and gave me this sense of (likely false) confidence that I could design and build anything. I still feel this way today.

Here, I work with domain experts in science many of whom are new to 3D printing. When discussing 3D printing with them, I can see that spark and transformational mindset shift that we don't need to rely on the vendor to get a part replaced or customized. There's this youthful exuberant joy around 3D printing and DIY tools here.

The 3D printing community is all about sharing ideas and designs, a shared mission of the Acceleration Consortium. Here's one example. This Makerworld model are helping hands for soldering. I loved the 3D printing spring design so much, I wanted to implement it somewhere in the lab. It ended up being used as a vial gripper adaptor. Robotic grippers are 2 parallel jaws, which would be unstable to pick up round vials. This adaptor would be gripped by the parallel jaws, then further closed to grip a vial.

As much as I love the maker movement I constantly question whether it's right for lab automation. Industrial automation uses PLCs and precision-machined parts, which is the opposite of what's happening here. If we want robustness, the obvious direction is adopt industrial automation tools. However, too many automated labs elsewhere have gone this route, where building the lab was outsourced to engineers taking years to build a robust automated lab only to get abandoned by scientists who prefer their traditional tools. Robustness means nothing if there's no utility. DIY tools that get adopted with less robustness is likely more useful. The project a few slides ago with error handling is something I'm excited about, because it potentially is a way to enhance robustness, while using solutions that the scientists are comfortable using. Right now, my mindset is adoption over robustness, but I'm constantly second-guessing.

MY CONTRIBUTION

Organ-on-a-Chip

Recapitulating the human body environment on a chip [2017-2019]

STANDARD CELL CULTURE

The human body is dynamic, complex, and extremely challenging to study. A lot of what we know about how the human body functions is through studying biology on a cell level. Interpolating from cell function to human body function is a huge jump, which avoids the importance of tissue and organ-level functions. Studying cells in a petri dish is not an accurate model for the human body. Neither is animal testing, which has poor success translating to human clinical relevance.

NEWER DEVELOPMENTS

The purpose of organ-on-a-chip technology is to provide a better in-vitro model to study cells by recapturing relevant tissue and organ level functions. These custom engineered models introduce physiological-relevant dynamics. For example when the heart beats, it causes cells to undergo shear stresses from pulsatile blood flow and membrane straining from tissue contractions. By introducing dynamics, the goal is to define a new standard for disease and drug response modeling.

Collagen Tubes for Tissue Scaffolding

Collagen is the most abundant protein in our body, and many tissues are tubular in nature. In our lab, we had the unique ability to manufacture collagen tubes with physiologically-relevant properties. We believe these tubes are a good foundation to build around for organ models on a chip.

Collagen tubes floating in a petri dish.

The low stiffness allow us to expand & collapse the membrane, similar to some human tubular tissues.

Collagen tubes expand and contract under subtle pressure variations. Viewed under brightfield microscope.

With a tubular structure, we can duplicate the uniform shear stresses from fluid flow in tubular tissues.

Fluorescent tracer solution perfused inside collagen tubular channel. Viewed under fluorescence microscope.

Developing a Tube Hosting Device

These tubes are challenging to physically manipulate (we use the analogy of handling wet toilet paper) and we needed to come up with a hosting device to perform cell culture experiments. We developed a device with 4 access ports to the luminal (interior) and abluminal (exterior) space of a 2cm long collagen tube segment. The device was made using clear acrylic for imaging access. All components were autoclavable.

With such a small internal volume inside the tube, there was not enough cell culture media to sustain long term cell culture inside our tubes. To overcome this problem, we came up with a rocking platform to recirculate media from reservoirs attached to the ports of our device. To run multiple experiments at a time, we developed this seesaw to hold 36 chips inside an incubator, allowing for intermittent gravity-driven perfusion.

Utility for Barrier Tissue Studies

In human physiology, barrier tissues are a permeable membrane that regulate transport of gases, nutrients, drugs, and others. One method for studying tissue barriers is culturing cells on Transwell inserts in a well plate — a porous plastic insert used to create two compartments. A common experiment to check barrier integrity is a trans-epithelial electrical resistance (TEER) measurement, where two probes are inserted on both sides of the insert and the electrical resistance across the membrane is measured.

In our custom tube hosting device, we could not use commercial TEER equipment, which could only be done in traditional well plates. We adapted the TEER principle and developed our own electrode probes that insert into the device reservoirs. We supply a voltage from a function generator and measure the voltage drop across the tube membrane to measure TEER across our tubular membrane.

Human bronchial epithelial cells were seeded inside our tube lumen and TEER measurements were taken over 2 weeks. As cells proliferated to create a complete barrier, we see a gradual increase in membrane resistance, plateauing at day 11. We later demonstrated this TEER setup to study barrier function, exposing our tubular tissue barriers to various drugs and stimulants.

Collagen-based Quake Valve

Quake valves played an instrumental role in the growth of microfluidics research. Through the deflection of a soft, flexible, silicone (PDMS) membrane, it allowed for the integration of on-chip miniature valves and pumps for precise fluid manipulation. We attempted to recreate the Quake valve with our collagen tubular structures, a natural substrate more suitable for organ-on-a-chip modeling.

Two collagen tubes were hosted in an acrylic device in cross-configuration. We flow fluid through the red tube. Using a solenoid valve, we toggle the blue valve between a high and low transmural pressure. At low pressure, the blue tube is collapsed allowing unrestricted flow in the red tube. At high pressure, the blue tube is expanded to block flow.

(a) Collagen valve junction viewed under brightfield microscope. (b) Red tube perfused with fluorescent tracer solution. Imaging window (Q in far left render) viewed under fluorescence microscope and images taken. (c) Sequential images from fluorescence microscope analyzed to find instantaneous flowrate in channel using particle imaging velocimetry (PIV) software.

Collagen-based Quake Pump

The extension of the Quake valve is to use 3 in series to create a peristaltic pump to recirculate fluid and generate a net flow. A modified acrylic device was designed to host 3 control blue tubes to recirculate flow within the red tube. These tubes were actuated in a 6-phase pattern, 0.75 seconds apart.

The collagen pump device viewed under brightfield microscope.

Using the same PIV technique to measure instantaneous flow, we plot the pump profile here. The net recirculatory flow was about 10 µL/min, mimicking pulsatile flow patterns found in human organ systems.

MY CONTRIBUTION

Gesture Control Remote 1: Hardware Development

An inclusive gesture control device for individuals with cerebral palsy [2020]

THE PROBLEM: For most people, interacting with digital devices is easy and second nature. Input devices such as a phone touchscreen, a computer mouse/keyboard, or a TV remote require little thought or concentration to use. However, these devices are not designed for everyone. People with impaired motor control or impaired vision, have a very difficult time using these devices, which require fine motor control. Commercial assistive technology are often expensive, limited in options, and are not generally designed with these users in mind. Everyone is unique in how they move, making it challenging to design a user-friendly, all-encompassing solution.

THE GOAL: To design low-cost, inclusive assistive technology for individuals with cerebral palsy to help them interact with digital devices easily.

THE DREAM: To close the digital accessibility gap between the general population and the cerebral palsy population. With this device, we want to give individuals with cerebral palsy (and other disabilities, too) greater individuality and be more capable of making positive contributions to society.

This project was a 2 month sprint in collaboration with Hackaday and the non-profit organization, United Cerebral Palsy Los Angeles. More details can be found on the Hackaday project page.

Project Inspiration

One thing we learned about the UCPLA community is the emphasis on visual arts. The UCPLA art studio allows users to paint with different body extremities (head, arm, leg, etc.) with the help of 3D-printed adaptive painting tools (see right images).

Inspired by watching UCPLA artists paint, my thought was: could we capture these motions and apply it to use digital graphics software? Using 2 M5StickC microcontrollers as "floating joysticks", I used these "joysticks" to replace a computer mouse for drawing images in Microsoft Paint. More details can be found in this log.

I can use it, but can you?

Playing a maze game with hand and knee and demonstration of re-calibration routine.

User testing session: Same maze game with cerebral palsy user over Zoom.

If I can use this device, then it's kind of cool. If an individual with cerebral palsy can use this device, then it becomes useful. Step 1 was to find a simple application suitable for testing. For this, I chose a computer maze game requiring 4 directional inputs (keyboard arrow keys). This application should be familiar for users who operate motorized wheelchairs with the ubiquitous 4-direction joystick. The left video shows me playing this game with two extremities, demonstrating robustness of this device. The right animation is from the first remote user testing session within a UCPLA residence, where our first user was able to successfully use this device, too. It was amazing to see how engaged the user was and also provided valuable learning lessons for future device iterations.

Iterations After First User Testing Session

HARDWARE UPGRADES

(a) A new 3D-printed enclosure was designed to house the (b) PCB from the first prototype. The unnecessary components (small battery, LED screen, etc.) were stripped from the original board and new electronics were added. This included a (c) vibrating motor for haptic feedback, (d) increased battery capacity for 9 hour runtime, and (e) a large, raised, CP-friendly button.

FIRMWARE UPGRADES

An objective of this project was to design a "universal" remote control, meaning with one input device, someone could control many electronic devices. For an initial proof-of-concept demo, this gesture control remote was used to emulate a Roku remote through a smart hub (Home Assistant installed on a Raspberry Pi). More details can be found in this log.

Future Roadmap

To wrap up this project, the idea was presented to the UCPLA panel, which can be seen in the video on the left. In addition, 3 prototypes were shipped to UCPLA for further user testing.

This project is currently still under development, with the focus on gathering more user testing data with other users with cerebral palsy. The software portion of this project was rushed and left much to be desired. Hopefully in a future iteration, this can be improved upon.

MY CONTRIBUTION

Gesture Control Remote 2: Firmware Update

USING TINYML TO MAKE GESTURE CONTROL TRULY ADAPTIVE [2021]

Limitations of Initial Prototype: During the initial hardware development phase of the gesture control project, it was clear during the user testing sessions that the initial signal processing of the IMU data was not adaptive enough to meet the needs of each unique user in the cerebral palsy community. Seeing our device used by others showed us that we were still forcing gestures that seemed unnatural. There was also a limitation that with our initial approach, we would not be able to scale the number of output gestures, which would have been limited to about 10 gestures.

New Approach: There was a thought that implementing a machine learning model would have been complementary for our application. After skimming the TinyML textbook, we thought integrating this was feasible. We saw that TinyML could be the enabling technology that allowed for a high degree of user configuration, perfect for a community were user's abilities and needs are one-of-one.

Introducing flourish, a project aimed to democratize digital accessibility to help the cerebral palsy community gain individuality and the ability to flourish. This project was selected as a world finalist for Microsoft's Imagine Cup 2021.

Platform Map

This map shows the overview of our platform. Apologies for the oversaturation of Microsoft product ads, this was for a Microsoft competition.

To use our platform, a user wearing the gesture device would go to our gesture training web application where they would be asked to perform various gestures. The recorded motion data would then be sent to our training portal where we would train a convolutional neural network to identify these gestures.

We then output and deploy a fully trained model back to the gesture device through a firmware update.

This device would then be ready to be used offline as a digital control interface.

Data Collection through User Training Sessions

This is a walkthrough of the gesture training application. The initial proof-of-concept dataset collected through this application was 225 sets each of 31 unique gestures. Yes, this hurt my rotator cuff to collect in 2 weeks.

Training a Convolutional Neural Network Model

Once enough gesture training data was collected, we used this data to train a supervised machine learning model, which was a convolutional neural network. The input dataset we ended up using was just the accelerometer (X,Y, and X) time series data logged by the IMU on the device.

For the initial demo application, we finetuned the model to train 31 different gestures (26 letters of the alphabet + a few extra gestures). On the right, you can see the accuracy of the validation dataset for each individual letter. For transparency, we recognized early on that some letters were very similar (D&P, E&F, X&Y, etc.). Our workaround was training a gesture of mixing lower case and upper case letters to improve model accuracy.

This trained model is uploaded back onto the gesture control device for real-world use.

The Fun Stuff: Demos!

After 2 weeks of data collection, we obtained an overall accuracy of 96% for 31 gestures, making it likely that the number of gestures could be increased further.

Demo 1: Hello World

With a trained English alphabet on our wrist, the first thing any reasonable human being would do is spell out hello world. Not going to lie here, this worked much better than expected. Didn't even have to fake it.

Demo 2: Personalized Dictionary

Speech impairment is common for people with cerebral palsy, and one instance this platform can help is with a personalized dictionary of phrases communicated through gestures.

MY CONTRIBUTION

Microfluidics

Microfluidics for molecular gastronomy [2017-2018]

Microfluidics involves the manipulation of fluids in devices with microscale channel dimensions, enabling precise control at small volumes. This field has gained popularity due to its wide range of applications, particularly in biological research. Microfluidic devices are biocompatible, and their small volume requirements are advantageous since many biological samples are expensive and difficult to obtain in large quantities.

One common microfluidic device is a droplet generator. In its simplest form, two immiscible fluids, such as oil and water, are fed into a device and meet at a junction to create micro-droplets. The left image shows a schematic of this setup, while the right image provides a close-up of the junction where oil and water meet to form water droplets (video from Elveflow).

During my master's studies at the University of Toronto, I was fortunate to conduct research in the field of microfluidics. Additionally, this research extended to bioprinting, intertwining 3D printing and biology for promising applications in healthcare. This section focuses on the use of bioprinting and microfluidics, showcased through side projects in molecular gastronomy.

Molecular Gastronomy with Microfluidics

Molecular gastronomy is where food meets science to flip traditional culinary practices. With the same ingredients, the perception of taste can be altered through molecular gastronomy. Here, we focus on using microfluidics to generate micro-scale edibles. The preparation of 3 different dishes will be shown. (Please be prepared to exercise some level of imagination... Thanks.)

For each dish, a different microfluidic device is used. All 3 devices are flow-focusing devices to create alginate gels, dependent on the reaction between sodium alginate and calcium chloride (CaCl₂). The two solutions create an ionic reaction, producing an edible, tasteless calcium alginate gel. All devices contain a junction where an alginate-filled channel is sandwiched in between two other channels. By controlling the flowrate ratios, the dimensions of the gels can be tuned.

1) Mould Design

A 3-layer device: top and bottom layers are hot embossed; middle layer has laser-cut through holes connecting channels. Black = through holes, green = alginate channels, light blue = oil channels. Fluid enters at the centre and exits at 16 perimeter outlets.

2) Mould Making

Two moulds are required to create the top and bottom layers. (a) A PMMA master is micro-milled. (b) An epoxy resin is poured onto the master. (c) After setting, the stiff elastomeric epoxy is peeled from the master to create the mould.

Falooda!

Hot embossed thermoplastic device

3) Pattern Transfer

The hot embossing process is shown in the schematic, where a mould is pressed against a PMMA slab. The right image shows the resulting embossed PMMA layer, a replica of the PMMA master.

4) Assembly

After two PMMA layers are embossed and the middle PMMA layer is lasercut, the three layers are bonded together via solvent bonding. Pipette tips are epoxied around the two inlet holes, used to connect to external tubing.

5) Fiber Formation

Alginate and oil are loaded into 2 syringes mounted on a syringe pump. Tubing is connected to the device, which is placed in a CaCl₂ bath for crosslinking alginate.

Fibers are pulled out of the bath for examination. Fibers of different diameters can be extruded depending on the inlet flowrates.

Falooda is a sugary vermicelli-like fiber. Here are our fibers on some ice cream with an attempt to make an Indian dessert, kulfi falooda. For sanitary reasons, the fibers were extruded differently for consumption.

Bucatini!

Extrusion and Controlled Deposition using PDMS device

Bucatini is a tubular pasta made from an extrusion process. With a PDMS microfluidic device, we can produce candied micro-bucatini (alginate-based tubes with sub-millimeter dimensions). Refer here for details on the microfluidic device. Apart from shrinking food to alter food texture, a dish can be elevated through presentation. Through the controlled deposition of our bucatini extrusion, we can entice with presentation, psychologically enhancing the perception of taste.

Our bucatini maker consists of three syringe pumps (1) that are connected to a microfluidic device (2). The bucatini extrudes from a glass tube (3) and onto a plate fixed to a moving XY stage (4). All components are controlled through the custom GUI (5).

Here, we show the ability to deposit our candied bucatini in an alluring spiral pattern. You will have to trust me that these are indeed tubes.

Caviar!

Laser-cut thermoplastic device

Spherification is a technique used in molecular gastronomy to create spherical gels. Using a flow-focusing microfluidic device, alginate-in-oil emulsions can be formed to make alginate droplets and later crosslinked to form gels.

This 3 layer laser-cut PMMA device has 5 junctions. The flow focusing junction is a standard design, but is hidden in this design where the microfluidic channels are in the shape of petals. The flower is a stylized Bauhinia blakeana, featured on the flag and emblem of Hong Kong.

On the left, one of the junctions is shown making mono-disperse 200 micron diameter alginate droplets. The other phase is oil, flowing 20 times faster than the alginate phase to form droplets. The alginate droplets exit the device and enter a CaCl₂ bath, where the alginate immediately gels to form "caviar".

Microfluidic device after bonding

Crackers and caviar

MY CONTRIBUTION

Reconfigurable Display

An electronic display with the ability to change size and aspect ratio [2016]

Moore's law drives technological advancement to make electronics smaller. However, the cellphone, the most prevalent electronic device, defies this trend. Why is this? Once touchscreens became practical, it created a trade-off between convenience and practicality. It is more convenient to carry around a smaller phone, but large electronic displays offer more utility. Arguably, the main difference between a cellphone, a tablet, and a laptop is the size of the electronic display. Electronics with reconfigurable displays would maximize the convenience and utility benefits of these devices.

The common approach in the flexible display research area is to use semiconductor fabrication with the key difference being swapping the silicon substrate for flexible, stretchable substrates (see figure below). The key issue with this approach is once the substrate is stretched, the pixel density decreases, reducing the image quality. Also, this approach is dependent on the substrate's mechanical properties, which diminish when stretched. This project proposes a different approach to this problem.

The Initial Design

The first design proposed a solution where 16 individual square "islands", each representing an LED, would move uniformly. Each island was constrained to move linearly on a track and connected to other islands via guide arms. By moving one island, all islands would move simultaneously.

Two major problems of this design forced a redesign. First, this design required extra space to function, defeating the purpose of having a compact design. Second, this design shared the same flaw as the flexible substrate approach, which did not conserve pixel density when stretched.

Re-Designing the Prototype

The islands needed to be first visualized in a compact configuration. 16 islands were placed on 4 height levels in collapsed state.

Upon expansion, the bottom islands needed to rise to the top level. By linking islands on different levels through rigid beams, only 1 degree of freedom between 2 coupled islands is allowed.

The inspiration for this design came from the Hoberman sphere.

From left to right, the figure shows the transition from collapsed to expanded state. Depending on the direction of the applied force, the prototype can expand to different aspect ratios. This figure shows expansion in two directions for uniform scaling.

Bringing the Design into Reality

(a) After finishing the design, all the components were 3D-printed. (b) After assembly, the prototype was built, and demonstrated the ability to expand and collapse. (c) LEDs were wired so each level of LEDs could be independently controlled. (d) One LED was placed on every island.

Bringing the Design into Reality

(a) A 2x2 version of the prototype could be expanded and collapsed to either a 2x3 or a 3x3 state. (b) With the current manufacturing method of 3D printing, the prototype could be scaled down to under a centimetre in height. The device height to width ratio remained the same with scaling. (c) Pixel density was conserved in different device states.

This design could shrink further with MEMS fabrication. However, the limitation of this design was the large height-to-width ratio. A future iteration of this design should focus on flattening the cube-shaped prototype. Nevertheless, this design showed a different, counterintuitive approach to flexible displays.

*Figures from Qaiser et al., Adv. Mater. Technol. (2018).

MY CONTRIBUTION

Balance Rehabilitation Game

An independent, novel balance rehabilitation trainer [2017]

Individuals suffering from spinal cord injuries (SCI) often have balance instability caused by a reduced neural control of lower-limbed muscles. Rehabilitation tools have been demonstrated to be effective in repairing the disconnect between brain and body. Two rehabilitation tools of interest are visual feedback training (VFT) and functional electrical stimulation (FES) therapy.

Wii Fit is a commercial example of visual feedback training. Training games can be used to strengthen muscles and improve balance.

MyndMove is a commercial example of functional electrical stimulation. Surface electrodes are placed on the body to activate targeted muscle groups to generate functional movements.

VFT + FES Therapy

In this project, we propose a balance rehabilitation tool that combines visual feedback training with functional electrical stimulation therapy. By combining an engaging training game with stimulation therapy, we hypothesize that our balance training tool will amplify and accelerate rehabilitation compared to single-modal approaches.

Just as the Wii Fit can serve as a personal trainer, a fully developed version of our proposed system can have the same at-home functionality for individuals requiring attention for balance rehabilitation. We can use this system as a platform for SCI patients to rehab daily in the comfort of their home while giving therapists access to training data for monitoring.

Four Task-Oriented Games

One training session included 4 games played 3 times each. The games were designed to emulate a collection of muscle strengthening exercises for balance rehabilitation. Each game displayed a transient coloured dot on the screen, displaying the individual's centre of position on the force plate. Game performance was measured by assigning game scores.

TARGET: This game trained an individual to stand as still as possible in the optimal balance position. The objective was to keep the red dot as close to the centre as possible.

HUNTING: A randomly spawning blue target appeared and the goal was to move the red dot inside the target and holding the position for 3 seconds. Once successful, a new target would appear in a different location. This game trained muscle reflexes by inducing constant shifting of body weight.

ELLIPSE: A target followed an elliptical orbit at the edge of an individual's range of motion. Once in the target, the target would increment along its path. This forced the participant to keep their centre of balance away from neutral position for the whole game.

COLOUR MATCHING: In this game, the dot changed colour. The individual needed to move the dot to its matching coloured target. Once successful, the dot changed to a different colour. This game required mental processing and drastic shifts in body weight.

Two Balance Assessments

In addition to the games, the training session included two balance assessments to track progress of balance improvement over multiple training sessions. These assessments were standard tests, giving performance indicators that show strong correlations to an individual's balance ability.

Static Balance Assessment

An individual was asked to stand on the force plate with their eyes closed as still as possible for 100 seconds. During this time, a centre of position (COP) time series would record and track the involuntary sway of the individual. For these studies, three balance indicators were taken: root means squared (RMS) in the (1) anterior-posterior (front to back) and (2) medial-lateral (side to side) direction, and (3) 95% confidence ellipse of the data.

Dynamic Balance Assessment

An individual was prompted by our GUI to perform the "star reaching task", where the individual leaned as far as possible without losing balance in the prompted green direction. After reaching in 8 directions, 8 data points were gathered, plotted, and joined to form an octagon. The area of the octagon gave the balance indicator for dynamic balance called dynamic balance area (or base of support).

System Debugging with Able-Bodied Individuals

Seven able-bodied individuals ran through the training session to test the effectiveness of the system. Three out of four games (y-axis) showed bivariate correlation between balance ability (x-axis), indicating the games were appropriately designed. For more details on this project, refer here.

Future Direction

One problem was for two games, ellipse and colour matching, everyone had high scores. This cluster made it difficult to distinguish an individual's balance ability with game scores. Immediately after the trial run, this problem was addressed by changing the scoring system to favour individuals who could keep their balance position closer to the centre of the game targets.

The next step would be to conduct another trial run on able-bodied participants, this time incorporating functional electrical stimulation. Once completed and further system optimizations are performed, the system would be ready for use in clinical trials on individuals with incomplete spinal cord injuries to determine the effectiveness of this rehabilitation tool.

MY CONTRIBUTION

Vehicular Rear-View Camera System

A camera-based module to replace the wing mirror [2016-2017]

Vehicles come in all sorts of shapes, sizes, and utilities - from sports cars to common sedans, hybrids to vintage cars, or trucks to Smart cars. While the rest of the vehicle has morphed to fit modern times, the wing mirror design has stagnated and looks comparable between most vehicles (see vehicles below).

The wing mirror design has been a major design flaw for over a century and requires a drastic design change. First, the wing mirror is a safety hazard. It creates blindspot zones, making it standard procedure for drivers to keep their head on a 180° swivel for blindspot checks before changing lanes. This flaw has existed for over a century and has never been fully addressed. At fast speeds, fractions of a second can be the difference between life or death. The safest option is for drivers to always keep their eyes on the road in front of them. Second, the wing mirror is the only part of the vehicle that unexpectedly extrudes out from the body of the car. Although relatively small, this obtrusion can cause up to 7% of total vehicle drag during driving.

Automakers are pushing to eliminate the law requiring vehicles to have anachronistic wing mirrors. We (hopefully) preview the future of wing mirror-less vehicles.

Compact Rear-View Camera Module

Inner Components

Due to the scope of the project, the system was designed around an off-the-shelf Raspberry Pi camera board, serving as the geometric constraint for our module. To not further increase the footprint of our module, a custom PCB was fabricated and placed over the camera board. The PCB includes high power infrared (IR) LEDs for night time viewing, heater connections for de-icing, a temperature sensor, and an ambient light sensor.

Outer Housing

The outer housing was designed to minimize drag created by this system. This camera module would be placed on the outside of the vehicle near the wing mirror. The housing prototype is 3D printed and provides a water-tight seal from the outdoor environment. A dual bandpass filter is placed in front of the camera lens to allow transmission of visible light and a narrow range on the same wavelength as our IR LEDs.

Aerodynamic Benefit

Here is a size comparison between a typical vehicular wing mirror and our camera system. After running a simulation for a standard highway driving condition, our camera system was 99.8% more efficient than the wing mirror.

Safety Enhancement: Blindspot Zone Elimination

On the left, the camera (large frame) and side mirror (small frame) views show our car being passed by another car. A schematic is shown on the top-right. At first, the car is in both the camera and side mirror view. When the car leaves the side mirror view, it enters the normal blindspot of our vehicle. Yet, in our camera view, the passing car is still in frame and only leaves the frame when the car has fully passed our vehicle. Thus, our camera eliminates the blindspot of cars in the adjacent lane.

Safety Enhancement: Low Lighting Conditions

Here is a comparison of our camera operating at night through the same highway section. The left is our camera with the infrared LED turned on, and the right is the camera with the light off. The right view is similar to the view of the side mirror. It is clear from these videos that the IR LED improves visibility at night. When the car enters the section without streetlights (or any other external light), the IR LED allows the camera to pick up the lane markings on the road. In comparison, the side mirror would have shown a pitch black view, demonstrating a driver's current dependency of external light. This camera module allows the driver to see potential vehicles with their headlights off (or moose, bears, zebras, kangaroos, or camels) that sneak up on us.

Demo Day

This was an undergraduate capstone project to demonstrate to the University of Waterloo community the potential advantages of a camera-based rear view system. For demo day, we showcased an interactive experience of our system on our supervisor's Porsche, demonstrating the safety benefits of our system. With the monitor placed in front of the driver, we demonstrated a driving experience without having to tilt your head to check blindspots.

Due to the limitations of time and money, the project was narrowed down to a manageable scope, demonstrating a few key concepts. Ideally, we would have put an emphasis on seamlessly integrating the monitors into the vehicle's dashboard to improve driving experience. With more time, another point of emphasis would have been coupling the camera system with haptic feedback in the steering wheel so drivers can feel their surroundings. There are numerous other ways in which technology can be used to improve our driving experience, so why not use it?

MY CONTRIBUTION

Alarm Clock Bed

Integrating an effective alarm clock into a bed [2015]

Have you ever slept through your alarm and missed an important meeting or exam? Are you the type of person that hits the snooze button 25 times every morning? Do you typically lie in bed in a depressing state of Weltschmerz, requiring all the will and strength in your mind and body to drag yourself out of bed? We've all been there. Mornings are the worst and traditional alarms clocks just don't cut it.

The Blue Clock

Introducing the solution to your morning problems. A bed that literally kicks you out of bed. The tilting alarm bed functions as a normal bed until the alarm is triggered. Once activated, a latch mechanism retracts causing one side of the bed to tilt over along with you. The bed will no longer lock in horizontal position for 30 minutes after activation, forcing you to not only wake up, but not climb back into bed. Stay woke.

Proof-of-Concept Functional Prototype

Once designed, an initial prototype is built out of Popsicle sticks. This prototype allows us to visualize the vision and assist with elucidating the direction for future design iterations. For instance, pinch points become more obvious and the observed drop height and drop speed help with understanding the mechanics of rolling out of bed. Shown is lil' Matt Damon rolling out of bed in style.

Putting it Together

After the design was finalized, a trip to Home Depot was planned. After weeks of arduous labour, the bed was built. Below are a few components that make up the bed, specifically the mechanisms involved with triggering the alarm.

All disassembled parts and sub-assemblies of the bed.

Latch assembly for actuation of the bed. Retract to trigger, extend to lock.

A limit switch fixed to the back post is used to detect the position of the top frame.

The actuation of the latch is triggered by a POC app via Bluetooth.

Full Scale Functional Prototype

Here is the bed in action. Instead of the traditional approach of waking up through sound, this bed uses the sense of (an excessively aggressive) touch. This alarm clock comes with a 100% guarantee of waking someone up. Although promising in its effectiveness, further persistence of getting this product to market would be met with too many lawsuits. BTW, if this gif had sound, you would have heard a gunshot. RIP Blue Clock (2015-2015).

MY CONTRIBUTION

Woodworking Arts and Crafts

yin and yang [2022-Present]

Engineering and arts are contrasting disciplines. Engineering is objective and logical, whereas arts are subjective and creative. Lately, I've taken an interest in woodworking and feel like it's a good medium for projects that fuse art with engineering. Through these woodworking projects, my aim is to take what I've learned in engineering for storytelling and personal expression. To begin, here were some traditional woodworking projects that I started with to learn the fundamentals of woodworking.

@thekumikoproject

After picking up woodworking fundamentals, I wanted to find a niche. Kumiko is an incredibly time-consuming, meticulous Japanese technique of assembling wooden lattices into geometric patterns. The traditional jig-and-chisel approach doesn't scale for large panels, so after doing some research (and finding out there's limited English resources online), I went to Japan to purchase specialized blades that cuts the most time-consuming step down significantly. I hate social media but here's my Instagram account. Some of my 3D printed designs are here and here, which are 3D prints that have helped me make Kumiko.

Wooden Keyboard

I wanted to work on projects that combine my engineering experiences with my new hobby of woodworking. This project was inspired the Hacoa Kiboard. I wanted to figure out how to replicate it with the addition of hiding it in a typical-looking woodworking project box. It's hard to see, but the keycaps are made from a single plank of wood, giving it the subtle detail of a continuous grain keycap set. Documentation found here.

"Unique" animal carvings

My first foray with wood was a bird carving course just before the pandemic. That bird on the left was the outcome of that class, and I vividly remember the instructor looking at it and was like, "Oh, that looks...unique?". Can't say I disagree. While stuck at home during the pandemic, I tried a second bird. This was an attempt at a three legged crow (yatagarasu), which was probably just as "unique". I kept at it and after a few blisters carving a chess set, I felt like I was past the "unique" phase.

I'm not great at enunciating my words. If I say "I am panicking", it sounds like "I am panda king". So here's a panda king. Or a panicking. I don't know anymore. Whatever it is, I was quite pleased with it and retired my wood carving career on a high and enrolled in a real woodworking course.

Vases (or Bowls)

Super Bowl

This is going to require some mental gymnastics to comprehend, please feel free to skip. My intention here was to create a super bowl for a friend's birthday gift. One might ask, what makes this glass vase a super bowl? I was trying to find a glass bowl and couldn't, and found the next closest thing which was this glass vase. Now assuming you can convince yourself that a vase is a bowl, what makes this bowl super? If you go to the subreddit called r/superbowl., you will not find any super bowls nor football-related content, but instead, superb owls. This happens to be my friend's favourite animal and there's a few owls etched onto the vase. Hence, superbowl. It's the thought that counts.

Quality Street Vase

This next vase is a bit more straight forward. I have a friend whose favourite chocolate is Quality Street, which comes in a very distinct purple octagonal box. I tried to recreate this box using purpleheart wood and added some details with the help of a laser cutter and epoxy.