A selection of what I have designed, built, and tested, across student teams, design competitions, and research.
A life-sized, waist-up humanoid robot built by a 16-person student team to serve as an AI-powered teaching assistant, commissioned by a physics professor at Western University.
Full-body shot of the robot winking, mid-assembly.
Full assembly CAD, waist-up.
Baker Bot is a waist-up humanoid robot commissioned by Dr. Mark Baker, a first- and second-year physics professor at Western University, to act as an AI-powered teaching assistant that answers questions and reacts during lectures. A team of 16 students split the build across six subsystems: software, electrical, arms, head, neck, and torso. The finished robot turns and tilts its head, moves its eyelids and jaw, and gestures with both arms while responding to spoken questions. The full build came in under $2,000.
I led the arm subsystem, a two-person team, and served as mechanical manufacturing lead for the whole robot, doing the bulk of the mechanical fabrication and final assembly.
The full arm assembly in 3D. Drag to rotate, scroll to zoom.
Each arm has three functional stages: a manually indexed shoulder that sets the working height, a powered elbow that provides the main range of motion, and a swappable hand on the end of a carbon fiber forearm. The design assumes the robot sits at a desk in front of students, so the arm's job is to gesture expressively within reach of the tabletop, not to lift heavy loads.
The shoulder pivots on a single large bolt fixed to the torso. An aluminum plate on the arm carries five holes, and a spring-loaded indexing plunger in the body drops into the chosen hole to lock the shoulder at one of five angles. Because the robot works seated at a desk, this lets the operator set each arm's height to match the table before a session, freeing the powered elbow to handle the expressive motion. Two cut-and-drilled aluminum sheets per arm carry the load into the pivot.
Indexing the shoulder by hand keeps it simple, strong, and power-free, saving the actuation budget for the elbow, where it matters most.
Shoulder pivot internals in CAD.
The printed shoulder housing and its indexing plunger.
The elbow is the only powered joint in the arm and carries roughly 110 degrees of range. It is driven by a 12V DC worm gear motor. The worm drive is non-backdrivable by design, so the elbow holds any angle at zero holding current: the arm draws no power and generates no heat while posed, which matters for a robot that spends most of a lecture holding a gesture. A potentiometer built into the joint reports elbow angle in real time, so the controller always knows where the forearm is. The whole joint is built from seven 3D-printed parts that fit together around the motor, potentiometer, and limit switch to form one compact, self-contained module.
Elbow CAD: worm gear motor, limit switch, and potentiometer.
Both assembled elbow modules.
Sizing the Elbow
The forearm and hand hang off the elbow as a cantilever, so the joint only has to hold their weight at full extension.
The selected worm gear motor is rated near 3 N·m, giving roughly a 3x safety factor on the static hold, with the non-backdrivable worm carrying the load at rest instead of the motor.
A same-size aluminum forearm would weigh close to double. Keeping the forearm light with carbon fiber keeps the elbow moment low, which is what lets a small, low-power motor hold position for a full lecture. Tip deflection of the carbon tube under the hand's weight is only about 2 to 3 mm, so the arm stays visually rigid while gesturing.
The forearm is a carbon fiber tube, chosen for stiffness at low weight. The hand mounts to the end through a quick-swap interface: a keyed turn-to-lock collar with a magnet that holds the hand seated. This lets the robot change hands for context: a right-hand-rule hand for physics demonstrations, an open pointing hand for gesturing and explaining, and a closed hand that can hold an object. Because the interface is shared, new hands can be designed and swapped in later, so the robot can be customized for special occasions. The hands themselves are printed and fixed rather than actuated, a deliberate scope cut to stay on budget without losing the expressive value of a posed hand.
The quick-swap hand mount.
A printed swappable hand, open palm.
I handled all of the arm's electrical work and integrated it into the robot's shared control system. At startup, a homing routine sweeps each arm until limit switches trigger, establishing the elbow's zero reference and the height of the table in front of the robot. During use, the in-joint potentiometer provides live elbow position. The arm's 12V worm gear motors run through a DC motor driver commanded by the system's ESP32-WROOM-32 controller, with the elbow potentiometer and limit switches wired back to that same board. Power comes from a 12.8V LiFePO4 battery, stepped down through buck converters on the electrical subsystem's power distribution board.
Gestures tie into the AI pipeline. An onboard Jetson Orin Nano handles speech-to-text and text-to-speech and passes each question to an external AI model, which returns a response plus an emotional tag. The robot speaks the answer while the arms play the gesture mapped to that emotion. Head and neck servos are driven separately through a PCA9685 PWM controller.
Every structural part on the robot was 3D printed, so I set a manufacturing standard that made parts come off the printer ready to assemble, with no cleanup and no failed fits. Details I built into every part:
I printed and assembled the complete arm subsystem along with parts of the head and neck, prototyping in PLA to iterate cheaply and reprinting final parts in PETG for strength and temperature tolerance, roughly 50 unique printed parts across the robot. The non-printed structure I fabricated by hand: cutting and drilling the aluminum upper-arm tubes and shoulder plates, and cutting the carbon fiber forearms to length.
An elbow part, showing the print details built into every part on the arm.
Leading a subsystem taught me to break a build into realistic long-term and short-term goals: near-term targets that keep a teammate learning and shipping at a steady pace, set early enough to leave room for team review and revision before integration. I taught the shoulder and hand designs through the CAD itself, so the knowledge stuck and the work held up. Balancing my own design and fabrication load against mentoring and whole-robot assembly was the core challenge, and it is the part of the project I would point to for any team-based engineering role.
Being both arm subsystem lead and mechanical manufacturing lead for the whole robot meant moving between CAD, electrical wiring, firmware, and a 3D printer farm in the same week, and being accountable for how well those pieces fit together at final assembly, not just for my own arm. A 16-person student build with a sub-$2,000 budget also has no room for a part that has to be redesigned after it is printed, which is why the manufacturing standards, the sizing math on the elbow, and the shoulder and hand mentoring all point the same direction: get it right on paper, with a tested reason behind the choice, before committing material or a teammate's time to it. That mix of technical range, budget discipline, and ownership past the edge of my own subsystem is what I'd bring to any team building real hardware on a deadline.
A two-U-joint drive shaft for Western Baja SAE's four-wheel-drive off-road car, with custom yokes I designed so we can machine them ourselves. Splitting the shaft lets both the engine and the driver's seat sit lower, with the engine dropping about 6 inches and bolting straight to the frame.
The team's current car during the 4-hour endurance race.
The team at Baja SAE Ohio 2026.
Western Baja SAE designs and races a single-seat off-road car in the Baja SAE series, which runs endurance races, hill climbs, and suspension courses over rough terrain. The car is four-wheel drive, so power has to get from the gearbox at the back of the car to a front differential at the front. In 2026 I competed with the team at Baja SAE New York in June and Baja SAE Ohio in September.
With a one-piece shaft, the gearbox output and the front differential have to line up in a straight line, which forces the engine to sit high in the frame. My split drive shaft removes that constraint. The shaft leaves the gearbox off to one side of the car, angles across through a U-joint at each end, and arrives at the front differential on the car's centreline. That frees the engine to drop about 6 inches and mount directly to the lower frame, which makes the mounting stronger, lowers the centre of gravity, and opens room to lower the driver's seat too.
I packaged the shaft in the frame, set the joint geometry, and designed the custom yokes. Then I ran the stress analysis and weight optimization that settled the final design: an all-steel shaft, now being finalized and built.
Seen from underneath, the problem the shaft solves is clear. The engine and gearbox sit at the back of the car, and the gearbox output is off to one side. The front differential sits on the centreline at the front. A straight shaft between the two would need the engine raised to line them up.
Instead, a short section leaves the gearbox, a U-joint turns the shaft onto an angled path, and a long section runs forward along the floor of the frame. A second U-joint at the front turns it back into line with the differential. The shaft stays low, below the driver, and clear of the frame tubes along its full length.
Lowering the engine by about 6 inches is the main payoff. The engine is one of the heaviest parts of the car, so dropping it lowers the whole car's centre of gravity, which helps the car stay stable on side slopes and through fast corners. Mounting the engine straight to the lower frame members also gives it a stiffer, more direct load path than a raised mount.
Bottom view of the shaft path, gearbox to front differential.
A single U-joint running at an angle does not turn at a constant speed. Even if the input spins perfectly steadily, the output speeds up and slows down twice every revolution, and the bigger the angle, the bigger the swing. On a drive shaft that becomes vibration, and extra load on the joints, the gearbox, and the differential.
The fix is to use two joints at equal angles. The first joint makes the middle shaft speed up and slow down; the second joint, set at the same angle and phased to mirror the first, adds exactly the opposite variation. The two cancel out, so the front differential receives the same smooth rotation that left the gearbox. That is why the geometry is built around one joint at each end of the long section, both running at the same angle.
Why Equal Angles Matter
The same ~4% swing that cancels at the differential still exists inside the middle section of the shaft, so the shaft and yokes are sized with margin for it.
The first design used two metals: a 4130 chromoly steel stub at the gearbox, which carries the splines that engage the gearbox output, and a 6061 aluminum tube for the long run forward, to save weight. Each yoke is welded to its shaft, and steel can't be welded to aluminum, so the yoke on the long section had to be aluminum as well.
The yokes are built around a purchased cross from All Balls Racing, and choosing a small cross keeps the whole joint light. A small cross means small yokes, though, and that's where aluminum ran into trouble. I loaded both yoke designs in SolidWorks Simulation with the full 300 N·m (2,655 lb·in) from the gearbox. The cross pushes on the two ears of the yoke as a pair of equal and opposite forces, so each force is the torque divided by the distance across the cross: 300 N·m over about 34 mm, or 8,824 N.
At the size the small cross allows, the 6061-T6 aluminum yoke went past its 275 MPa yield even unwelded, and not just where the cross bears on it: the sides of the ears yielded too. Welding only makes that worse, because the heat softens 6061 next to the weld well below its normal strength, so aluminum was ruled out. The same yoke in 4130 steel held the load. With steel yokes at both ends, the long section went to steel too, so the shaft is now steel from end to end, with every yoke welded straight on.
6061-T6 aluminum yoke, unwelded, at 8,824 N: red is past its 275 MPa yield.
4130 steel yoke under the same load, against a 460 MPa yield.
Reading the Plots
In both yokes the highest stresses sit on the inside faces of the bearing bores, exactly where the cross pushes on them. The model applies that force over a small contact patch, and FEA reports stress at a spot like that which keeps climbing as the mesh gets finer, instead of settling on a real value. Those readings come from how the load is modelled, not from the part, so I ignored them. What decides whether a yoke holds is the stress just away from the contact, through the ears and down into the base. In the aluminum yoke that stress still passes yield on the sides of the ears; in the steel yoke it stays below steel's 460 MPa yield.
Steel is almost three times as dense as aluminum, so every bit of material in a steel yoke costs weight. SolidWorks can plot the factor of safety at every point in the part, and those plots showed whole regions with extremely high factors of safety, material that was carrying almost no load. I cut those regions out in ways that are simple to machine, taking roughly 30% off the weight of each yoke while leaving material everywhere the load actually flows.
The joint also needs room to move. Shock loads and vibration make the joint flex and rattle as it spins, so I added extra clearance between the two yokes to make sure they can never collide, without giving back the thickness the optimization had trimmed. I worked out that clearance in the CAD assembly and checked it with 3D-printed models of the joint.
The steel U-joint in 3D: two lightened yokes and the cross. Drag to rotate, scroll to zoom.
Switching to steel settled the strength question but raised a weight one: how much heavier would an all-steel shaft be? I wanted to compare it with the mixed aluminum-and-steel shaft it replaced, with the current car's drive shaft, and with titanium, which is lighter than steel for the same strength but much more expensive and harder to machine.
For that comparison to be fair, every option had to be at its own lightest, so I built a spreadsheet that optimizes the tube for each material. A tube carrying torque can fail in two ways. It can yield, which calls for a thicker wall, and for a given torque a bigger diameter needs less wall, so yielding on its own keeps pushing toward bigger, thinner tubes. Or it can buckle, where a large tube with a thin wall wrinkles and collapses as it twists. Buckling pushes the other way, so between the two there is a real lightest tube.
For each material, the spreadsheet sweeps the outer diameter in 0.01 inch steps. At every diameter it finds the thinnest wall that passes both checks, each with a factor of safety of 2, and never goes below the thinnest wall that can be made and welded in that material. Then it picks the lightest. The yield check uses the shear stress from the 300 N·m design torque, and the buckling check uses the Timoshenko critical stress for a long, thin-walled tube, with a 0.7 knockdown factor because real tubes buckle below the ideal value.
What the Optimizer Found
These sizes are the optimization step. Next, each tube is rounded to the closest-fitting standard tube size before anything is ordered.
The full spreadsheet is below, rebuilt to run in the browser. Every input is editable: change the torque, the safety factors, or any material property, and the optimum, the chart, and the design comparison all update.
Editable! Click to open the optimizer, then change any blue cell and it will recalculate.
The all-steel design is being finalized and built now. It goes on the team's new car, which will be racing before the end of the school year.
Alongside the drive shaft, I've worked on other parts of the car, including sprag bearing wheel hubs and a redesign of the wheel rims to reduce weight.
A noble-metal-free high-entropy alloy catalyst engineered for overall alkaline seawater splitting, developed under Dr. Amir Mirzaei in Western's Department of Physics.
About 95% of hydrogen production today still relies on fossil fuels, and the catalysts that make green hydrogen viable through water electrolysis are usually made from platinum, iridium, or ruthenium, some of the rarest and most expensive metals on Earth. Scaling green hydrogen means replacing them with catalysts made from abundant, affordable elements, and doing it with seawater rather than freshwater, since freshwater is a limited resource while seawater is effectively infinite. My project, under Dr. Amir Mirzaei in the Department of Physics, was to design and optimize exactly that: a bifunctional catalyst for overall alkaline seawater splitting built entirely from earth-abundant metals. In parallel, the lab also had me working on high-entropy oxide semiconductors for photocatalytic hydrogen evolution, a related but separate thread not covered in the presentation below.
The catalyst is a high-entropy alloy (HEA), NiFeCoMoCr, electrodeposited onto nickel foam for high surface area. HEAs combine five or more elements into one alloy, which produces electronic and structural properties, the "cocktail effect", that none of the individual metals have on their own. I used AI and machine learning to process literature and experimental datasets and narrow down which metal combinations were most promising before committing lab time to them. Once the composition was set, I ran a systematic, iterative optimization of the fabrication process: testing nickel foam against titanium oxide nanotubes and carbon paper as substrates, sweeping electrodeposition pulse counts to tune coating thickness, and finally applying a hydrothermal ammonium fluoride treatment to reconstruct the surface and expose more active area. Each stage was evaluated by electrodepositing samples and testing them with linear sweep voltammetry, cyclic voltammetry, and Tafel analysis on a potentiostat, narrowing in on nickel foam, 6,000 electrodeposition pulses, and a mild 0.05M fluorination step as the best-performing combination.
I validated the optimized catalyst in simulated alkaline seawater (1M KOH with 0.5M NaCl added to approximate natural seawater chlorides) rather than just clean freshwater electrolyte, and it held up: strong resistance to chloride-driven corrosion and a stable potential over extended chronopotentiometry testing. Benchmarked against five recent high-entropy alloy catalysts from the literature and against noble-metal standards (RuO2 and Pt/C), the final catalyst matched the best oxygen-evolution performance and exceeded every catalyst in the comparison, noble metals included, on hydrogen evolution. I presented these results at Western's 2026 USRI Research Conference, and they're now being carried forward for publication, with an incoming student continuing the work in the lab.
With Dr. Amir Mirzaei and fellow members of his research group.
Hydrogen and oxygen evolving off the electrodes during a bench-scale electrolysis test.
A short presentation covering my work in more depth.
Click to view the full presentation.
A two-machine system that sweeps green erasers off a set, slightly uneven floor area, carries them back to base, and sorts them by colour with no human input and no remote control. Built in eight weeks for a second-year design competition, where it placed 2nd.
Full system CAD, scoop raised over the sorting station.
The same assembly in 3D. Drag to rotate, scroll to zoom.
The competition brief gave us 120 seconds, a set area of roughly 15 m², and a slightly uneven floor scattered with erasers. The system had to locate and collect them, tell green erasers apart from every other colour, and deposit them into the correct bin at a home base, entirely on its own, with scoring tied to both how many it collected and how accurately it sorted them.
The decision that shaped everything else was splitting those two jobs across two machines. A mobile scavenger robot drives the area and sweeps up erasers without trying to identify anything, and a stationary sorting station at the home base does all the colour classification. Collection is fast and indiscriminate; sorting is slow and careful. Neither job compromises the other.
The sorting station was mine end to end: mechanical design, electronics, and firmware. My teammates built the scavenger.
Why Sort at the Base Instead of Onboard
We scored three full system concepts through a Go/No-Go screen and then a weighted decision matrix. Onboard sorting kept failing the same way: a colour sensor riding on a moving robot sees changing ambient light, vibration from the drivetrain, and erasers tumbling past at unpredictable angles. Every one of those degrades a colour reading, and a misread costs points directly.
Moving the sensor to a fixed station fixes the sensing environment: consistent geometry, consistent lighting, one eraser presented at a time. The cost is that the robot has to come back to base, but it already had to, because that is where the bins are.
The station takes a scoop-load of mixed erasers dumped in at the top and processes them one at a time. Two inclined V-shaped platforms vibrate the pile into a single-file line, with a drop between them that breaks apart whatever is still stuck together, a colour sensor reads each eraser as it reaches the bottom, and a servo-driven chute swings to send it left or right.
The feed platform, and the chute with the colour sensor at its mouth.
The robot tipping its scoop onto the feed ramp.
The station separating, reading, and diverting erasers.
The feed is two V-shaped platforms in series, each with its own vibration motor, and a short drop from the first into the second. The arrangement is borrowed from industrial sorting lines: potato and produce graders use staged vibrating beds with transfers between them, because one bed on its own will not reliably present a single item. The V walls channel the pile into a line, and the drop between the stages knocks apart anything still stuck together on the way down.
Neither vibration motor is a vibration motor. Each one is an ordinary DC gearmotor with a hex nut glued to the encoder disc on the back of the motor shaft, upstream of the gearbox, where the shaft still turns at the motor's own speed rather than the reduced output speed. An offset mass spinning that fast is the entire working principle of a vibration motor, and putting it before the reduction is what makes it work at all: the geared output turns far too slowly to shake anything. Two nuts and some glue stood in for a part we did not have.
Each V sits on four springs, and the spring rate mattered as much as the motors did. Too soft and the platform wallowed without moving anything along; too stiff and it barely moved at all. Getting erasers to index down one at a time meant tuning spring rate and motor speed together rather than either on its own.
The two stages are also deliberately out of phase: the firmware runs motor one for 500 ms, then motor two for 500 ms, with no gap between. The reason is the drop. An eraser leaving the first V while the second V is also shaking lands on a moving surface and bounces, sometimes back up the incline, sometimes straight past the line. Offsetting the two means the platform receiving an eraser is always the one holding still.
At the bottom, a single servo drives the chute between two fixed positions: 0° for non-green and 70° for green. There is no in-between and nothing to calibrate at runtime. The chute rests in the non-green position, so a sensing failure or a power glitch defaults to the reject path rather than contaminating the green bin.
Nearly every hard problem on this machine sits upstream of the sensor. Classifying is the easy half: given one eraser, held still, under fixed light, calling it green or not green is close to solved. Getting exactly one eraser to arrive, alone, is what took the iterations.
Erasers are close to a worst case for a feeder. They are rectangular blocks with flat faces and square edges, so they neither roll nor flow. A heap of them locks together instead of pouring, and it will bridge across the full width of the ramp and sit there indefinitely. The faces are also large relative to the mass behind them, so two erasers in contact have a great deal of surface holding them together and very little weight trying to pull them apart.
Then there is the rubber itself, which is engineered to grip. Erasers stick to the platform, stick to each other, hold static, and shed dust that makes the surface grippier as a run goes on. That leaves the ramp angle with no good setting: shallow enough to meter erasers out individually and the pile simply stops moving, steep enough that it always moves and the whole clump slides down as one body. Gravity and geometry alone never separated them. The separation had to come from the vibration and from the drop between the two stages, which is why the feed ended up staged rather than as one long shaking ramp.
This is also where the residual error lives. On single erasers the classifier does not miss; in full runs the system sorts at 90–92%. Essentially all of that gap is the feed rather than the sensor: when a whole scoop-load lands at once, two erasers occasionally reach the aperture stuck together and the sensor reads a blend of the two. Every improvement we made to sorting accuracy after the first week was an improvement to feeding, not to the algorithm.
The sensor is a TCS34725 running over I²C at a 50 ms integration time and 4× gain, and each decision averages three consecutive reads to damp sensor noise. The important step happens before any threshold is applied: the raw red, green, and blue counts are normalized against their own sum, so what the algorithm actually compares is each channel's share of the total light rather than its absolute value.
That one change is what makes the classifier survive real conditions. A green eraser under a bright lab light and the same eraser in shadow produce very different raw counts but nearly identical normalized ones. Absolute thresholds would have needed recalibrating every time someone turned on a different bank of lights in the room.
The sensor board recessed into the chute wall.
On top of the normalized values, an eraser has to clear six independent gates simultaneously to be called green. Any single failure rejects it:
| Gate | Threshold | What it rejects |
|---|---|---|
| Clear channel | ≥ 120 | An empty chute with nothing under the sensor |
| Normalized green | ≥ 0.40 | Anything not predominantly green |
| Normalized red | ≤ 0.32 | Yellows and oranges |
| Normalized blue | ≤ 0.32 | Teals and cyans |
| Green dominance | ≥ 0.080 | Greys and washed-out neutrals |
| Hue angle | 108–145° | Yellow-greens and blue-greens at the edges |
The gates overlap on purpose. Green dominance, the normalized green share minus whichever of red or blue is larger, catches the case where a colour is technically green-leaning but so desaturated that it is really grey. The hue window catches the opposite case: a colour saturated enough to pass the channel tests but sitting at the yellow or blue edge of green. A single threshold on green alone would let both through.
Mounting mattered as much as the thresholds. The TCS34725 has to sit close to the eraser, because its reading falls off quickly with distance and starts picking up whatever else is in view, so the board is recessed into the chute wall with its face level with the gap the eraser passes through. Everything around it prints in black, which keeps stray colour from bouncing into the reading and gives the sensor a consistent dark background to read a coloured object against. That proximity is also why the feed springs ended up stiffer than the feed alone would have wanted: a softer spring fed slightly better but shook the sensor enough to smear the reading, and a still sensor was worth more than a marginally better feed.
Passing the gates once is not enough to move the chute. The reading has to stay green continuously for a full second before the servo commits, and it has to stay non-green for a full second before the chute swings back. An eraser tumbling past the sensor, a shadow crossing the aperture, or a momentary reflection all produce single-frame misreads, and none of them survive a thousand-millisecond hold.
This is the cheapest reliability improvement in the whole project: a few lines of timestamp comparison that eliminated essentially every false trigger we saw on the bench. Across more than a thousand bench samples, the classifier produced no misclassifications.
The scavenger is a 3D-printed rear-wheel-drive chassis with two DC gear motors, a rotary brush across the front, and a servo-actuated scoop. The brush sweeps erasers back into the scoop as the robot drives forward, and reverses direction when the robot backs up so that the erasers already collected stay in rather than being swept back out. It averaged eight to nine erasers per collection pass.
A full autonomous run, collection through sorting.
Fast physical prototypes were what got us to a working design. Model a version in SolidWorks, print it that night, test it on the bench the next morning, change the model the same day. A part in your hand tells you things a CAD model will not, especially with a material as awkward as eraser rubber, and once a rough version worked the same loop was what turned it into a good one. Almost nothing on the finished system is a first revision, and the parts that decided whether it worked went through several each: the scoop, the feed platforms, and the chute.
The scavenger's heavy first design, before the weight rework.
The sorter's first design: a tilted rotating funnel.
Neither machine resembles its first model. The rotating funnel above is a centrifugal bowl feeder in miniature, and it is a completely reasonable way to single-file parts, as long as those parts slide. Once there were printed erasers-and-plastic tests on the bench it was obvious that no amount of spin walks a sticky rubber block up a smooth wall in an orderly line. The vibrating platforms that replaced it were simpler and far more reliable, and they took one less servo to build. The first scavenger went the same way: it was a heavy robot, and keeping the robot light was one of the competition objectives, so what shipped was reworked around minimising weight.
That is the part worth keeping from this project. The binding constraint was not CAD time, it was how many times we could afford to be wrong before the deadline, so we optimised for that: keep every part cheap to print, keep the models easy to re-cut, and never let a single idea carry so much work that killing it hurts. A wrong design cost us an evening instead of a week, and the machine that placed 2nd is essentially the one that survived the most rewrites.
The station's firmware is a single non-blocking loop driving an explicit state machine, with nothing blocking execution while it waits. That matters here because the vibration motors and the colour sensor have to run concurrently: the motor cycle cannot stall colour reads, and a colour read cannot stall the feed. Each keeps its own timestamp and advances independently, which is also what makes the 500 ms offset between the two feed stages exact.
The scavenger runs the larger program, built the same way: 27 states covering three collection passes, each one a sequence of drive out with the brush running, pre-lift the scoop to trap the load, reverse to base, align, lift fully, dump, lower, and turn to the next heading. The scoop lifts in timed one-degree increments every 25 ms rather than being commanded straight to its target, so the load is not thrown out by a sudden servo slew. During the reverse leg the scoop holds a 50° mid-lift, high enough to retain erasers but low enough to clear the station on approach.
Letting the Mechanism Absorb the Error
Our early design assumed closed-loop heading control for navigation. What actually shipped is dead reckoning: timed drive segments, a per-motor PWM trim to correct the chassis pulling to one side, and a potentiometer to tune overall drive speed between runs.
Position accuracy comes from the ultrasonic sensor and the station geometry instead. The robot reverses until the ultrasonic reads 15 cm from the base, then keeps driving for a fixed five seconds while a pair of funnel-shaped guide rails on the station physically pull the chassis into alignment. Rather than measuring the error and correcting for it in software, the rails mechanically remove it. It is a less sophisticated answer than closed-loop control and a considerably more reliable one over a 120-second run.
Full firmware for both machines, CAD, engineering drawings, circuit schematics, and the complete design report are on GitHub.
As VP Combat Robotics, I lead Western's first combat robotics program. Two student teams each design and build a 3 lb beetleweight, fight each other, then take what they learn into competition against other universities in the BotBrawl Canada league.
Reference photo, not our robots: 30 lb robots from the University of Illinois's iRobotics Combots teams at Robobrawl 2025. Photo: Mason Daniels, via iRobotics.
Bot Brawl is WE First's combat robotics program. It runs two teams of about 15 members, each building its own 3 lb (1.36 kg) robot. Each team has a build lead who runs the robot overall, plus mechanical, electrical, and software leads who each own a subsystem and guide the members working on it.
A big part of the program is teaching: members learn robotics by designing, building, and fighting a real competition robot. Splitting into two smaller teams instead of one large one gives more members hands-on work, and gives each robot a real opponent to test against.
Each robot fits drive motors, a weapon, a LiPo battery, speed controllers, and a radio receiver into 3 lb. Both teams are in the design stage, and the two types we're considering are:
When both robots are finished, they fight each other. After each fight, the teams tear their robots down, record what broke, and redesign those parts, so both robots have been through real fights before outside competition.
Reference footage, not our robots: the 3 lb final at NHRL's 2025 World Championship, where Pinevictus, a beater-bar vertical spinner, takes on ColdSnap, an undercutter. Video: NHRL's BrettZone on YouTube.
At a Glance
Tentative, and will change as the builds progress.
Recruit members, form the two teams, and pick the build and subsystem leads.
Each team picks its robot type, then concept CAD, weight budgets, and design reviews before parts are ordered.
Fabricate and assemble both robots, then drive and weapon testing.
The two robots fight, with a teardown and fix after every fight.
Rebuild with what the internal fights taught us, then compete against other universities in the BotBrawl Canada league.
Top photo by Mason Daniels, from the University of Illinois iRobotics Combat Robotics page; the robots in it belong to the iRobotics Combots teams. Fight footage by NHRL's BrettZone, embedded from YouTube. The robots in it belong to their builders: Pinevictus by Alex Wang and ColdSnap by Myles Sim (details via the NHRL wiki). The BotBrawl logo and arena photo on my homepage card belong to BotBrawl Canada and are used only to reference the league we plan to compete in; the program is not affiliated with BotBrawl Canada.
An enclosure for the Prusa XL that anyone can build for about $162 in parts: every bracket printed on the printer it encloses, every panel a plain rectangle of hardware-store acrylic, and an active filtration unit running on a $5 filter. Placed 1st in Western's First-Year Engineering Design Competition.
The full assembly: printed parts in red and black, acrylic panels clear.
The competition table: printed part kit, fasteners, and build manual.
Printing without an enclosure means an open bed exposed to room drafts and a workspace breathing whatever the hotend gives off. The enclosures that fix this are expensive, rarely modular, and often insulate poorly, which is why the makerspaces that need them most tend not to have them. Our brief was Western's FAST Lab: a high-traffic student shop where an enclosure has to survive constant handling and constant modification.
The design constraints were tight and mostly financial: under $150 for the base enclosure, under $50 for whatever value-add we chose to bolt onto it, standardized or 3D printed parts only, recyclable materials, clear sightlines to the print, and an open-source design someone else could actually reproduce. We took the fume problem as the value-add and built an active particle filtration unit into the enclosure itself.
I led the team of five and worked across the whole project rather than owning one subsystem: the CAD for the printed parts, the filtration unit and its control circuit, and the dimensioned build manual we shipped as the deliverable.
Designed to Be Copied, Not Just Built
"Open source" was a constraint we took literally, and it turned out to be the constraint that decided most of the design. Every part had to clear the same bar: could a stranger with one 3D printer, a drill, and no workshop reproduce it? That ruled out laser cutting, welding, custom PCBs, injection-moulded clips, and any proprietary hardware.
What survived is a parts list you can fill anywhere. The acrylic is nine plain rectangles in six sizes, cut from two sheets by whoever sells them to you, with M3 clearance holes drilled to the dimensions in the manual. The brackets are 30 printed parts across nine designs, all of which fit a standard bed and print in PLA: 941 g of filament, about 26 hours, one spool. The fasteners are M3 screws and 10×3 mm magnets from a bulk pack. Nothing needs a soldering iron.
The other half of it is documentation. CAD files alone are not a reproducible design, so the deliverable was a 23-page build manual: panel drawings with every hole dimensioned, part-by-part assembly renders, a hole-by-hole breadboard wiring guide written for someone who has never wired anything, the full bill of materials with prices, and operating instructions. It is the actual product, and the enclosure is what falls out of following it.
The hardest requirement to satisfy cheaply was attachment. The enclosure has to hang off a printer we are not allowed to modify, with no drilling, no adhesive, and nothing that voids a warranty on a machine a lab depends on.
The Prusa XL's frame is extruded aluminum, so the answer was to use the T-slots already in it. Part 001 is a two-piece printed clamp: 001A drops into the channel and rotates 90° to lock, acting as a printed T-nut, and 001B slides on from the outside. A single M3 screw pulls them together and traps the acrylic panel between them. Eight of these carry the structural panels, and the whole enclosure comes off the printer as fast as it went on.
Panels meet each other the same way. An L connector (002) screwed in from the inside holds two sheets at a true 90° without needing a jig, a corner bracket, or a clean edge. The part defines the angle, so the acrylic does not have to.
The T-slot integrator clamped into the printer's extrusion.
The L connector holding two panels at 90°.
Every panel is clear for a reason. One of the objectives was monitoring a print without opening anything, because opening the doors dumps the heat the enclosure exists to hold and releases the fumes the filter is busy capturing. A user who can watch the print through the wall does not have a reason to break the seal.
The front doors run on printed hinges (007) mounted on the inside face and close onto neodymium magnets rather than any kind of latch. The magnets are glued into grooves in two printed holders: a corner holder (009) on the frame and a drop-down holder (008) on the door. The closing force comes from twelve $0.25 magnets and two printed parts instead of hardware nobody can source. Printed handles (004) bolt through both faces of each door.
The practical benefit is that the doors cannot bind. A magnetic closure has no alignment to lose and nothing to wear out of adjustment, and a door that pops open under a pull anywhere along its edge is the right failure mode for a lab where people are reaching in with one hand full.
The two magnet holders meeting as a door closes.
The value-add had a $50 ceiling, which is roughly what a single replacement cartridge costs on a commercial fume filter. So the unit is built around a filter you can buy at IKEA: an UPPÅTVIND air-purifier refill, $4.99, stocked in most of the world.
A printed housing (005 body, 006 lid) bolts to a side panel and holds the filter in a slot in front of an 80 mm silent fan, pulling enclosure air through the media instead of letting it drift into the room. Because the filter is square and only the quarter in front of the fan loads up, the housing lets you pull it out and rotate it 90° when it is spent. Four positions, four filter lives, one $5 part, which is the difference between a consumable people replace and one they quietly stop replacing.
The filtration housing and lid in a side panel.
A filter nobody can see is a filter nobody changes, so an Arduino Uno tracks run time and blinks an LED when the media is used up. A touch sensor resets the timer whenever the filter is rotated or swapped. The fan itself runs through a relay off the printer's power rather than its own switch, so filtration starts when the print starts and stops when the printer does. There is no step in the workflow a user can forget.
The circuit is deliberately unimpressive: a relay, a resistor, an LED, a touch sensor, a step-down converter, and a breadboard. The manual assumes no soldering setup and no experience, and walks the builder through it one breadboard hole at a time. No part in it costs more than $10.
The cost breakdown is the part of this project I would point at first. Of the $111.64 base enclosure, $98.98 is the two sheets of acrylic. Every bracket, hinge, handle, magnet holder, screw and magnet in the structure, the entire designed portion, comes to $12.66. The design is essentially free; what a builder pays for is glazing, at whatever their local price for a sheet happens to be.
| Metric | Result | Requirement |
|---|---|---|
| Base enclosure cost | $111.64 | ≤ $150 |
| Active filtration add-on | $49.58 | ≤ $50 |
| Acrylic share of base cost | 89% | — |
| Printed parts | 30 parts, 9 designs | — |
| Filament and print time | 941 g · ~26 h | — |
| Total build cost (CAD) | $162.40 | — |
The Three Concepts We Threw Out
We scored four enclosure architectures before committing. Modular snap-fit panels were expensive, restricted us to one material, and would not survive a shop. Magnetic hexagonal panels were expensive and relied on magnets strong enough to be a problem, with no obvious way to scale to a different printer. A tarp-style shield was genuinely cheap, and looked it. In a space where the enclosure is also a demonstration of what the lab can make, that mattered.
The winner was the least inventive one: flat panels fitted to the shape the Prusa XL already is. Taking the printer's own geometry as the frame meant no structural design to get wrong, no material we had to defend, and mounting points already built into the machine. Most of what makes this design cheap is that we let the printer do the structural work.
The full 23-page manual: panel drawings, assembly renders, wiring, bill of materials, and operating instructions.
Scroll to read it here, or click to open it full size.
By the end of the term we had printed the complete part kit and built and run the filtration unit and its control circuit; the enclosure shell itself is documented as a cut-and-drill package against the panel drawings. Both budgets held: $111.64 against $150, and $49.58 against $50, with 42 cents to spare.
The project placed 1st in Western's First-Year Engineering Design Competition. What I took from it is that the reproducibility constraint was not a limitation on the design; it was the thing that made it good. Being unable to reach for a laser cutter or a custom part is what pushed us onto the printer's own T-slots and onto a five-dollar filter anyone can find.
The team at the competition with the filtration unit and manual.
A small autonomous robot that drives a set course, finds an IR beacon, grabs a ping pong ball with its claw, and brings it back to where it started.
The finished robot: ESP32, ultrasonic sensor, and the claw.
The course is laid out in taped-together 8.5" × 11" sheets of paper, and the robot has to stay on the paper the whole time. Fully on its own, it has to:
The hard part was getting the robot to move the same way twice. The paper surface was uneven and the wheels had poor grip, so the same drive command could produce noticeably different distances and turn angles from one run to the next, and small errors early in the course added up by the end.
The IR signal was also noisy. Stray readings could pull the robot off target, so steering purely by IR was unreliable.
The approach that worked combined all three: a pre-planned route for the driving legs and the turn, the ultrasonic sensor to judge the distance to the beacon and stop in the right place, and IR only for small final adjustments to line up with the ball. With that combination, the robot grabbed the ball and returned to its starting position successfully and repeatably.
The course: out, a 90° turn to the beacon, and back.
A few smaller projects and experiences. Hover over each for a short description.