Results for
I built Beat the PID for the One-Button Game Mini Hack: Beat the PID. Each tap switches an actuator on or off, and you try to hold a trace inside a moving band across six plants, from a thermostat to an inverted pendulum. A PID ghost plays the same level with the same on/off limit.
The yellow line is you, and the dashed gray line is the PID ghost. The band scrolls in from the right so you can plan ahead.
These are the things that surprised me, with measurements where I have them. I built it with help from an AI coding assistant (Claude Code), and we tested each change in MATLAB.
1. The character limit counts reformatted code
My first local count, with comment lines removed, said 9,862 characters. minihackContest.validate said 12,253. The validator's number matched this exactly:
code = char(tree2str(mtree(fileread('game.m'))));
numel(code)
mtree and tree2str are undocumented, so treat this as a useful local estimate. It matched on my game at several sizes and on another contest entry.
Reformatting changes what costs characters:
% What you write
ax.Color=[.035 .045 .07];
% What gets counted
ax.Color = [ .035, .045, .07 ];
- Comments are free, so comment generously.
- Squeezing out spaces saves nothing, because they get added back.
- Every value in an array costs extra for its comma and spacing.
- Each nesting level costs 4 characters per line. Moving a deep block into its own function can pay for itself.
- Grouping assignments helps:
[s.t,s.acc,s.in,s.ing,s.ei,s.dwg,s.n] = deal(0);
Those changes, along with cutting a few visual extras, took my game from 12,253 to under 10,000 and kept all six levels.2. min and max ignore NaN
I used NaN for trace samples not recorded yet, then clamped values to the scope:
max(-.02,min(1.02,NaN)) % returns 1.02, not NaN
Because min and max ignore NaN, every empty sample became a real value at the top of the scope and drew a stray line. If you need clamping, restore the gaps afterward with y(isnan(v)) = nan.3. Use a fixed physics step and expm
dt varies between frames, so I run the physics at a fixed 120 Hz and spend the elapsed time in whole steps:
s.acc = s.acc+dt;
while s.acc >= s.h && s.t < s.T && ~s.crash
s.acc = s.acc-s.h;
s = stepPlant(s);
end
Each plant is linear, x' = Ax + B[u;1], so I can discretize it exactly once at startup:
n = size(A,1);
M = expm([A B;zeros(2,n+2)]*h);
Ad = M(1:n,1:n);
Bd = M(1:n,n+1:n+2);
Each physics step is then just x = Ad*x + Bd*[u;1]. The contest player limits dt to 0.05 s, so a game running under 20 fps plays in slow motion, but the physics stays stable and each step stays the same.4. A one-button PID is a relay
The ghost's PID output doesn't set a power level. Its sign picks on or off, and a minimum switching interval stops it from toggling faster than a person could:
e = r-y;
ei = max(-.3,min(.3,ei+e*h));
v = e+Ki*ei-Kd*(y-yPrev)/h;
if timeSinceSwitch >= minDwell && (v > 0) ~= u
u = v > 0;
timeSinceSwitch = 0;
end
The integral term matters more than I expected on a relay. On the inverted pendulum level, the ghost averaged 52% of the time in the band with Ki = 0.2, and 7% with Ki = 0. These are averages from simulated runs at one set of gains, and the result is very sensitive to tuning.
5. Measure work and frame rate separately
MATLAB Online was by far the biggest performance surprise. On my connection, the same game behaved very differently:
Median game and drawing time per frame. I measured MATLAB Online on a slightly earlier build, before the final color and text changes.
The MATLAB Online time was almost the same on all six levels, even though each draws different things, so I suspect a fixed delay and not my drawing load. Other games lagged for me there too. Your results may differ, so test your own game there.
Two habits helped:
- Time the work, not just the frame. On my Windows PC, a paced loop took about 46 ms per frame even when the actual work was about 10 ms, because pause overshot. Frame rate alone would have sent me after the wrong problem.
- Send only what changed. Create graphics once, set fixed X data once, and update only Y data. My render also skips frames where nothing changed, which keeps idle menus quiet and still gives the same result if called twice:
st = [s.mode s.lv s.t s.u];
if isequal(s.ax.UserData,st)
return
end
s.ax.UserData = st;
A piecewise-constant band needs vertices only at its ends and steps, which cut it from up to 166 numbers per update to a handful. Scrolling the scope at 15 Hz while moving the player marker every frame also reduced updates without making the controls feel slower.Play it on the desktop
If the game crawls in MATLAB Online, you can run it in desktop MATLAB:
- Download the starter project, unzip it, and open minihackContest.prj. Opening the project installs the contest tools.
- Copy my entry's code into the project's game.m.
- Run minihackContest.play.
Tap Space or click to play. I'd love to know which plant beats you, and whether anyone can beat the PID on all six.
Hello MATLAB EXPO community! Now that everybody has had a chance to digest all of the exciting sessions you saw last week, I wanted to create a forum for further discussion of @Emmanouil Tzorakoleftherakis and I's session, Building Embedded Controllers with Agentic AI and Model-Based Design. Do you have questions, comments, or feedback following the session? Let us know! Some points to consider for discussion:
- What was something interesting you learned from the session?
- What is something you'd like to learn more about after having watched the session?
- How comfortable are you with the prospect of AI agents designing controllers? Did the session change that feeling at all?
We'd love to see your thoughts below!
By SunHuilin
What happens when a tiny pixel cat tries on eight different personalities? Sometimes you get instant cuddles. Sometimes your sofa gets a new ruler.
Purrfect Fit is my entry for the MATLAB One-button Game Mini Hack. Different pairs of eyes scroll through a cat's glasses. Tap or press SPACE to stop them. Fit the eyes inside the frames, and the kitty reveals its personality with a short message and matching pixel effects. Miss the target, and the kitty offers a little encouragement. Tap again to try another round.
Play the game: Purrfect Fit
Here are four techniques behind the game. The snippets below are simplified illustrations of the implementation, rather than a complete runnable game.
1. Treat an RGB array as a tiny canvas
The artwork starts as a 150-by-200-by-3 RGB array, displayed across an 800-by-600 game area. One artwork pixel occupies four game-coordinate units in each direction. Rectangles, filled polygons, and small bitmap patterns build the cat, glasses, lettering, and effects without loading external image files.
The key is to create the image object once, then update its CData:
% During initialization:
h = image(ax,[2 798],[2 598],background);
% During rendering:
frame = background;
% Draw the moving elements into frame here.
h.CData = frame;
The static background is prepared during initialization. Each render starts with that background and adds the animated elements. This avoids accumulating drawing operations from previous frames and keeps the number of graphics objects fixed.
I also draw the title with small bitmap letters. That gives the main heading a consistent pixel style without depending on an installed pixel font. The longer instructions and feedback use MATLAB text objects for readability.
2. Separate scrolling, stopping, and feedback
The game has three states: scrolling, success, and miss. The same button stops the eyes during scrolling and starts another round after the result.
Scrolling uses elapsed time rather than a fixed distance per update:
phase = mod(phase + speed*dt,192);
There are eight eye designs, each occupying a 24-row tile, so the strip repeats after 192 artwork pixels. Using dt makes the intended scrolling speed independent of the number of updates per second, although a slow display can still make fast motion harder to judge.
One small detail matters: when a press arrives, I check the current phase before advancing the strip. This keeps the timing decision tied to the state that was just rendered.
The fit test uses the visible vertical bounds of each eye design. Both bounds must lie inside the target opening; checking only the center could accept eyes that still stick out of the frame. On success, the strip snaps to the centered position for the final portrait.
A brief restart guard prevents an immediate second press from dismissing the result before the player can read it.
3. Animate the tail with a small coordinate shift
The cat's tail uses filled polygons. Instead of storing multiple animation frames, I shift their horizontal coordinates with a sine wave:
d = 3*sin(t*3);
x = x + round(d*(118-y)/21);
The vertical weighting keeps the base relatively steady while the tip moves more. Rounding preserves the pixel-art appearance.
Layering is just as important as motion. I draw the moving tail first, restore the static foreground artwork through a mask, and then draw the eyes and glasses. This keeps the tail behind the body and prevents it from painting over the face.
4. Store small patterns compactly, then reuse them
The eight eye styles share a palette and a common rendering path. Their pixel data is stored as compact runs and decoded during initialization, so rendering can reuse the prepared arrays.
The personality effects use a different representation: seven integers describe a 7-by-7 binary pattern. Each bit says whether one pixel is filled:
if bitget(pattern(row),8-col)
% Draw this pixel using the personality's color.
end
The same drawing loop produces paws, lightning bolts, sleepy Zs, sparkles, spirals, question marks, hearts, and musical notes. Changing the pattern and color gives each personality its own visual feedback without duplicating the effect renderer.
The messages work similarly: each eye style has three possible success lines, while misses draw from a shared pool of encouraging phrases. For example:
- Fierce: "Tiny paws. Big boss energy."
- Sleepy: "Wake me when snacks arrive."
- Loving: "Can I keep you, human?"
- Miss: "Oops! Let's blame the whiskers."
Random choices happen when a round starts or a result is selected. Rendering reads those choices, so the message does not change every frame.
What helped most under the character limit
The biggest savings came from reusing drawing routines, palettes, and animation logic, and from compact representations of the artwork. Shorter variable names helped too, but shared structure saved more space than shortening names alone.
The current build passed the official validator at 8,933 counted code characters. During compression, I compared rendered frames and seeded update sequences with the previous version to check that the smaller code preserved the existing appearance and gameplay.
My favorite part is that the eyes change more than the cat's expression: they change the little conversation you have with it.
Which personality would you add next: a suspicious detective kitty, a dramatic movie-star kitty, or something even stranger?
Development note: I used OpenAI Codex to assist with coding, iteration, and code review.
Currently, MathWorks provides free access to MATLAB Online for anyone with a personal account. Each account receives 20 hours of usage per month and includes access to the following 10 toolboxes:
- Control System Toolbox（控制系统工具箱）
- Curve Fitting Toolbox（曲线拟合工具箱）
- Deep Learning Toolbox（深度学习工具箱）
- Image Processing Toolbox（图像处理工具箱）
- Optimization Toolbox（优化工具箱）
- Signal Processing Toolbox（信号处理工具箱）
- Statistics and Machine Learning Toolbox（统计与机器学习工具箱）
- Symbolic Math Toolbox（符号数学工具箱）
- Text Analytics Toolbox（文本分析工具箱）
I would like to know whether it is possible to customise the selection of these 10 toolboxes. For example, I mainly work in computer vision, but Computer Vision Toolbox is not included in the current offering, which is quite disappointing because I cannot use it.
At the same time, several of the included toolboxes are not useful for my work, such as Control System Toolbox and Signal Processing Toolbox. Therefore, it would be great if users could choose the toolboxes that best fit their needs rather than being limited to a fixed set of toolboxes.
Thanks to Dr. Riyadh Kenaya and Arnie Berlin for the session and for answering my questions live: UE5 handles radial and tangential distortion, rolling shutter would be added in Simulink, and the synthetic data is meant for design of experiments on extrinsic limits rather than exact calibration.
A few follow-ups:
- The demo results showed a mean calibration error of 3.85 px, with camera 1 at 7.33 px. On synthetic images where the true intrinsics are known, what drives that error: the distortion model fit, board detection near the image edges, or something else?
- Can the ChArUco board poses be set from a script instead of the Control Center sliders? That would let a calibration campaign run as a Monte Carlo sweep over board placement and lighting.
- Does the camera model support fisheye or rational distortion models for wide lenses (around 100 degrees horizontal)?
- With a 32-channel lidar, only a few rings hit each board. How much does camera-to-lidar extrinsic accuracy degrade as the boards move farther away?
I would also be glad to try the example if the code is available.
Thanks to Siddharth Jawahar for the Misconceptions in Control Design workshop. I worked through the M3 to M5 models and have three follow-ups.
- Transport delay in PID Tuner. In M5, the tuned design showed 60 degrees of phase margin at 5.27 rad/s, while the gain margin sat at 25 rad/s for both the original and the tuned gains. That suggests the transport delay sets the phase crossover. Does PID Tuner linearize the Transport Delay block exactly, use a Pade approximation, or tune on the delay-free plant? If it drops the delay, the margins on screen would be optimistic.
- Missing the settling spec. The tuned design settled in 1.02 s against the 0.8 s requirement, while rise time and overshoot passed. Which would you trade first: a little more overshoot, derivative action, or reducing the delay itself?
- Windup in field logs. The M4 hand-built PI asked for about 26 V against a 12 V driver limit and overshot by 187.7 percent after the disturbance cleared. On real hardware, is the requested voltage pinned at the limit while the error keeps growing the signature you look for? Do you recommend logging requested and applied voltage as separate signals by default?
Context: I work on simulation validation and am building a fault-injection testbed for a small robot that injects latency, jitter, and packet loss on the control link. Variable delay is the case I care about most, so any pointers on modeling jitter (for example with the NetworkCondition block) would help.
Thanks!
Hello everyone, I hope you all had a fantastic MATLAB EXPO! I wanted to create a forum for people to ask follow-up questions or start discussions related to the workshop Hands-On Agentic AI for Engineered Systems with MATLAB and Simulink. We ran this workshop nine (!) times over the two days of MATLAB EXPO, and heard some incredibly insightful comments and questions in each session, so I wanted to make a central bulletin board for people to share their thoughts and hear insights. A few ideas:
- What did you most enjoy about the workshop?
- Are there any topics/exercises that weren't clear, or maybe an area you want to explore more?
- Do you have an idea for how you'd like to apply agentic AI in your work?
MATLAB EXPO may be over, but that doesn't mean the conversation has to end!
I have to say, fitting this into 10k characters was quite challenging. This is an original version coded many, many years ago that I wanted to reimplement to test my limits. I actually lost the code of the original project because the PC died...and since at the moment I feel nostalgic for that achievement I just wanted to go back my steps.
First of all: colours and textures. To stay in the limits I had to give up on them.
🆕 (the contest remake) https://it.mathworks.com/matlabcentral/communitycontests/uploaded_files/20973/preview.gif
Moreover, the implementation of the collision system definitely had to be simplified -guys be kind in the comments please- and reused for both the walls and the moving parts of the flipper. Of course it does not handle edge-cases and sometimes the ball travels where it should not. By the way, I think the real advantage of this 'reuse' approach is that the elasticity of each collision can be independently parameterized. And also, it's a simple Flipper game with one button, it's just for fun, no real physics, just creativity. Since I'm naturally competitive, I deliberately decided to make this one relaxing: no score, no pressure, and infinite lives. 💚
- of course the ball moves because at some point I say (x,y)=(x,y)+v*t and v=v+a*t,
- and of course I had to pack the more verbose geometric formulas (semi-obscured tilted ellipses 💀) with some clumsy function-handle
- ⬆ but this is totally off-topic.
About the flippers, I loved the idea to save some little extra space by rendering all the moving segments with the quiver function, since the junction point (x,y) in this game is fixed and calculating the relative (u,v) component comes for free.
I really had to abandon the wormholes (a kind of segment that if the ball enters it - exits from another portal) but at least I could recreate a weird kind of solar system in which any planet is used as a bouncer. (This year I got married 💍, so I plot many many saturn-rings 🪐- I love you Selena)
In the original game there were some conditions in which the gravity could flip direction (bottom-up) and this feature was smartly comunicated in-game with some asteroids that changed orientation (upside down) based on the situation. Kinda delightful, but the first rule of the game is clarity for the player: if it's too complicated to put inside and the code is already saturated, better let it go.
But what I really appreciate of such a strong language as matlab happens to be, are the small details in the classic syntax. Do you recall this simple code? ⭐ plot( ___ , 'p' ) ⭐ that generates star-symbols at no cost! The temptation to use it on space-themed works is SO STRONG!!! Couldn't resist. It's too cool! And to make the idea of an infinite flow of stars there is a buffer of coordinates that keeps resetting those ones that reach nearly the center of the black hole, so there is no need to delete and regenerate new symbols. Some coding expert probably would label it as Procedural Generation... not my case: it's just common sense - MATLAB Online renders so slowly 🥲
I guess I said enough... So, enjoy the Black Hole in the space flipper, a simple state-machine mechanic that when the ball passes too close makes the hole star-hungry and when it reaches the critical level, all the universe vanishes and all the trajectories collapse due to the gravity, until the timer resets the state 🫠. I'm not sure but Matthew Mcconaughey probably was in that tiny Hole during the first rendering step, I swear I saw him...!
Don't forget to smile for the right reasons. Byeeee
clc, clear all
(non AI generated text, the use of emojis was a built-in feature of my personality)
For the MATLAB One-Button Game Mini Hack I made KOI — and I fell straight into the trap: I skimmed the rules, got excited, and built the prettiest game I could (caustic water shaders, a spine-chain fish, score-reactive effects…). Then I met the second line of the rules: 10,000 characters. The next several hours were spent tearfully ripping out features and compressing code until no human could read it anymore.
Read the requirements first. The constraints are not the fine print — they are the design brief. "Beautiful" and "10k characters" lead to two very different first lines of code.
A thank-you to AI agents, by the way: my final file has dense comments and aggressively packed one-liners that a human reader would weep at — but the agent working alongside me still knew exactly what was going on, measured every cut before it was made, and caught every regression. If you use agents for this, a few things made a real difference:
- Clear structure — small, single-purpose functions (init / update / render / helpers) instead of one blob
- Skills and rules — giving the agent your constraints (contest contract, banned functions, testing workflow) explicitly, once, up front
- Sorted context — snapshots of earlier states on disk meant we could always trade features back in; nothing was ever lost, only parked
If I did it again, my order of work would be:
- Brainstorm — everything on the table, nothing wrong yet (a pretty game and a minimal game are both valid answers here)
- Decide what you actually want — the experience, not the feature list
- Write down requirements — including the contest rules, this time as step one
- Architecture — how state flows, what gets created where
- Individual base functions, then merge the ideas in
- Test headless from minute one
- Friends play it — they find what you're blind to
- Only then: polish
I did steps 4–8 with a lot of joy and steps 1–3 not at all. That order is free, and it would have saved me a very entertaining, slightly painful afternoon.
GLHF — and read the rules. 🙂
PS:
Idea:
Final result after cutting dynamic rock spawns, lilies on the leaves, drifting light particles, guide lines, wake ripples, a second pickup ripple, 3 of 4 title glow copies, the game-over box styling:
PPS: but it still looks beautiful though haha
My entry for the one-button game mini hack, GOING UP, makes you the lift attendant. The lift never stops on its own: one tap stops it, the next tap sends it on. Every passenger wears the colour of the floor they want.
How to play
- Stop level with a floor and the gate opens. Riders for that floor get off (1 point each), and waiting people get on until the lift holds four.
- Stop more than 12 px off a floor and you get a CLUNK. Nobody moves, and everyone waiting loses 1.5 s of patience.
- Stop within 4 px while someone gets on or off: PERFECT, +1 point.
- Waiting passengers have a patience bar. Three walk-outs and your shift is over.
- The lift speeds up and people arrive faster as the shift goes on.
Tips for a high score
- Skip floors with nothing to do. Every stop costs time.
- The lift only turns around at the top and bottom. Pick up people whose floor is ahead of you; anyone else rides to the end and back.
- Don't mash. In my bot tests random tapping lost in about 21 s, and never tapping in about 29 s.
MATLAB tricks inside
- The lift has no direction flag. Its height is a triangle wave of the distance travelled, cy = 460 - abs(mod(ride, 840) - 420), so stopping just pauses ride.
- Queue order without loops or sorting: ahead = sum(w' & s.pf' == s.pf & s.side' == s.side & s.pt' < s.pt, 2) counts, for every passenger at once, who arrived earlier at the same spot.
- Timers store when something happened (s.stopAt = s.t) instead of counting up every frame. Elapsed time is just s.t - s.stopAt.
- All 20 passengers are one patch of 6-vertex bodies plus one scatter of heads. Empty slots are NaN, so nothing is created after init.
What's your best shift?
I'm interested in chatting with anyone who is using Bionic or another agent to write MATLAB code. I have puny 6GB PC GPU RAM but large 64GB McMini RAM. What model are you using, what joys and woes have you had?
What does the participation matrix represent - from the Modern Grid Stability with IBRs: Leveraging Simulation and Control Design session.
We got wonderful questions from the Design For Every Ear workshop session and we did not have chance to respond to them. We use this thread to share some more resources and references to continue the discussion.
- Concepts, tools and examples: https://www.mathworks.com/solutions/medical-devices/hearing-aids.html
- Example references used in the workshop:
- The Hearing Loss Simulation Plugin by Angeliki Mourgela (Queen Mary University of London and BBC R&D), MATLAB File Exchange 79541.
- MathWorks shipping example: Cochlear Implant Speech Processor.
- MathWorks shipping examples, including Speaker Recognition Using x-vectors.
- Embedded AI technology and workflow: https://www.mathworks.com/solutions/embedded-ai.html
- Set up the AI tutor in your own MATLAB: MATLAB MCP Core Server (https://github.com/matlab/matlab-mcp-core-server), MATLAB Agentic Toolkit (https://github.com/matlab/matlab-agentic-toolkit) and Simulink Agentic Toolkit (https://github.com/matlab/simulink-agentic-toolkit)
Thanks!
Exploring various cutting-edge technologies was an inspiring experience. Thank you, MathWorks, for this invaluable opportunity!
I have been using MATLAB on a daily basis for more than 30 years. For the past few months, I have been working with version 2026, and the problems I have encountered are staggering, especially when debugging functions.
The issues started to appear gradually with R2026a. Suddenly, the cursor would disappear, copy and paste would stop working, and MATLAB often had to be restarted because of these issues. In addition, the new figure UI is painfully slow, and entering debug mode has become frustrating. When stepping through code, MATLAB can take several seconds to respond to a debugging command such as Step.
Today, I finally got fed up and reinstalled R2024b. It works perfectly fine.
My entry for the one-button game mini hack, LIGHTKEEPER, puts you in a lighthouse on a reef. Ships sail in blind from the dark. The beam sweeps on its own, and one tap reverses its spin. Any ship the light touches turns for home.
How to play
- A tap reverses the beam. It has inertia, so a reversal takes about a third of a second.
- Boats are rescued the moment the light touches them: 1 point.
- Freighters need 0.6 s of light: 3 points. The arc around the ship shows progress.
- A lantern blinking red means that ship is close to the reef. Three wrecks and the lamp goes out.
Tips for a high score
- Tap fast to make the beam hover. Quick reversals cancel out, and a slow beam lights freighters three times faster.
- Swing toward the ship closest to the reef, not the one closest to the beam.
- Don't mash. In my bot tests, random tapping lost in about 19 s, and never tapping in about 39 s.
MATLAB tricks inside
- Every position is a complex number, so z.' + outline.*exp(1i*heading.') turns and places the whole fleet at once. Use .', because ' also conjugates.
- All ships are one patch with a face per ship. NaN vertices hide the ones not sailing. The whole game is 24 graphics objects.
- The beam is a patch with per-vertex alpha ('FaceAlpha','interp'), built once and turned with an hgtransform.
- Setting the axes limits every frame, even to the same values, added about 5 ms to each frame. The screen shake now touches the axes only while it's running.
How long can you keep the lamp lit?
Hi everyone! I made PULSE // PALS - Lumi & Afterglow for the One-Button Game Mini Hack.
Play the published game:
https://www.mathworks.com/matlabcentral/communitycontests/contests/10/entries/16481
Press SPACE or click when a diamond reaches the small ring beneath Lumi. Release between taps. There is no aiming or dragging. The game has 64 authored notes across eight visual phrases, with a tempo that gradually rises from 96 to 136 BPM.
The submitted version passed minihackContest.validate at 9,778 characters, with no banned functions reported. Here are the most useful lessons from building it. The short code examples below use descriptive names for explanation; they are not additional files required to play.
1. Make the moving target and the notes share one beat coordinate
The target drifts and the rail bends, but the input is still about timing. A note must meet the ring at its scheduled beat, wherever the ring has moved.
For a note at noteBeat and a current beat b, define d = noteBeat - b. This simplified example uses the same rail formula as the game:
rail = @(d,b) 22*(sin(.72*d+.22*b)-sin(.22*b)) ...
+ 26*sin(pi*b/12).*tanh(d/2);
b = 12.5;
noteBeat = 14;
targetX = 218 + 58*sin(pi*b/8);
targetY = 306 + 37*sin(pi*b/6);
spacing = 91*(1+.09*sin(pi*b/4));
d = noteBeat - b;
noteX = targetX + spacing*d;
noteY = targetY + rail(d,b);
The useful invariant is rail(0,b) = 0. At d = 0, the note and target coincide. The hit test uses beat distance, rather than mouse position or a pixel collision. That keeps the animation and timing rule connected while leaving room to change the path.
2. Animate a character's parts, then batch the drawing
Lumi is procedural geometry, not an imported sprite. The ears, hands, blinking eyes, star pin, and two scarf tails have separate motion. A shared transform adds tilt and squash-and-stretch.
All 22 colored character polygons are held in one patch object; the mouth uses one additional line. Independent animation does not require an independent graphics object for every body part.
Complex coordinates are a convenient way to express 2-D transforms. For local points z, rotation angle theta, and a position x0,y0:
zWorld = (x0 + 1i*y0) + z*exp(1i*theta);
set(h,'XData',real(zWorld),'YData',imag(zWorld));
The game creates its graphics objects during initialization and reuses them. It also avoids setting text when the string has not changed:
if ~strcmp(label.String,newText)
label.String = newText;
end
The main waveform uses 177 on-screen samples instead of storing a whole song-length curve in the scene. Background geometry has a separate update gate. These reduce unnecessary work; they are not a promise of a particular frame rate in MATLAB Online.
One testing lesson: check the effective face colors, not only one expected CData array shape. With flat shading, per-face colors and per-vertex colors need to be interpreted correctly; for the latter, the first vertex determines the face color.
3. Make the strongest effects something the player earns
A catch makes Lumi rebound and produces a ring around the target. Eight consecutive catches outside Afterglow fill the constellation and start an eight-beat double-scoring interval for subsequent notes. The palette warms, the scarf extends, and the surrounding shapes become stronger.
I wanted the visual changes to explain what the player had achieved. The background stays comparatively restrained between these moments so that the next diamond remains readable. The chart is fixed rather than randomized, so replaying also means learning its pauses and short runs of notes.
4. A slow frame is not proof that the player left
An earlier version paused whenever an update interval exceeded 0.8 seconds. During MATLAB Online playtesting, it repeatedly paused just after starting. The code was treating elapsed time as if it were a window-focus event.
I removed that automatic-pause condition. The submitted version bounds a simulation step to 0.1 seconds and marks slow sessions PRACTICE. This avoids a large catch-up jump, but intentionally allows the game to run slower than wall-clock time under sustained load. It does not fix rendering throughput or network latency.
A browser prototype helped with visual iteration, but browser smoothness is not a measurement of the MATLAB version. Testing in the official runner caught a problem that screenshots and reference simulations did not.
5. Pack data, and let the official validator decide the size
Some geometry, palette, and chart tables are stored as character arrays and decoded into numbers. For a small integer table, the principle is:
values = [0 2 4 6];
packed = char(13312 + values);
restored = double(packed) - 13312;
assert(isequal(values,restored));
Those unusual-looking characters are numerical data. They are not dynamically executed code. Keep values in a safe character range, preserve the source encoding, and verify that decoding reproduces the original table.
The distinction between character arrays and string scalars matters: double('ABC') returns character codes, while double("ABC") attempts a numeric conversion and returns NaN. My decoder receives character arrays.
Raw file size was not a reliable way to judge the contest budget. I used minihackContest.validate for the final decision and kept an expanded numeric copy for inspection. Compact data, shared drawing helpers, and fewer repeated statements were more useful than merely deleting whitespace.
Working with AI
AI assistance was used for implementation, iteration, and testing support. I playtested the game and used visible problems to steer changes: the first designs were too static, the character felt like a sticker, and the slow-frame pause rule failed in actual use. The important lesson was to keep reference-model checks, native validation, and real playtesting separate.
Useful MATLAB documentation
Graphics performance:
https://www.mathworks.com/help/matlab/creating_plots/improve-graphics-update-performance.html
Patch colors and properties:
https://www.mathworks.com/help/matlab/ref/matlab.graphics.primitive.patch-properties.html
Character codes and numeric conversion:
https://www.mathworks.com/help/matlab/matlab_prog/unicode-and-ascii-values.html
What visual cues help you anticipate a beat when there is no sound? I would be interested to hear what other Mini Hack participants have tried.
One-Button Game Mini Hack Entry: Pump-Send
Hi everyone! Pump-Send is a side-view mountain bike game on an endless desert pump track. You never pedal: all your speed comes from pumping the rollers, and one button does everything. On the ground, a tap pumps or pops; in the air, it throws you into a backflip.
How scoring works
- You get 1 point for every meter you ride. The run ends if you run out of momentum or crash.
- Every clean backflip is worth 50 points, so one flip equals 50 meters of riding.
- Your score and flip count are shown in the top-left corner.
Definitions
- Pump: A tap while rolling downhill. You push the bike into the slope and gain speed.
- Scrub: A tap at the wrong time (going uphill, or a second tap right after a pump). It costs you a little speed.
- Lip: The rising face just before the top of a hill. It glows yellow.
- Pop: A tap on a glowing lip. It's the only way to get airborne.
- Flip: A tap in the air spins the rider backward. Each full rotation you land counts as one flip. Tap only; don't hold the button.
- Stall: Going too slow for too long. This ends the run.
Physics you'll want to know
- Steeper downhill = bigger pump. A pump on a steep drop gives you a big shove, about a third of your starting speed. On a nearly flat stretch, it does almost nothing. Aim for one good pump on each downhill.
- Scrubs add up. A badly timed tap only takes back about a third of a good pump, but a few in a row can leave you crawling.
- Watch the hills scroll. If the scenery slows down and you're barely making it over the crests, pump now or you'll stall.
- Speed is hang time. The faster you hit a glowing lip, the higher and longer you fly, and the more time you have to finish a flip.
- Spin fades in the air. Each tap adds spin, and the spin slowly dies down. More taps spin you faster, but you have to be wheels-down at the right angle when you touch the ground.
- Land matching the slope. Don't start another flip if you won't finish it in time to land at the right angle.
Quick tips
- Build speed with a few pumps before you try your first flip.
- On a big jump, one tap in the air is usually a full flip. Count, then stop tapping.
- When in doubt, skip the flip: distance points still add up.
- The game gets harder as you go, with tighter pump timing and landing requirements.
Behind the scenes
I built Pump-Send with a team of Claude Code agents: one designed the game, one wrote the tests (113 of them!), one built it, one reviewed the code, and one judged prototypes to help me choose among my four original ideas. The judging turned out to be less useful than I expected, since it was pretty obvious after playing which idea would pan out best. The most interesting lesson for me was about using bots to tune the physics.
Lesson one: the bot was too good
The first bot was simply too good at the game. It predicted where each jump would land and planned its flips accordingly, scoring hundreds of points. Then I took the wheel (err, handlebars) and crashed in seconds.
So I had the agents build a bot that plays like a person. It makes the same decisions, but each tap arrives 100 ± 60 ms late, it waits to see what one tap did before deciding again, and it taps at most 10 times per second. That bot survived a median of 8 seconds and crashed in 17 of 30 rides.
Here's what it uncovered and what I changed:
- A slightly late tap on a lip turned into a backflip spin on a tiny hop. Now you only leave the ground when you tap.
- Riders landed on the next uphill 55–68% of the time. Now a lip glows only if a pop at your current speed will land on a downhill.
- A flat dead zone at every crest turned late taps into scrubs. It's gone.
- Everything runs in gentle slow motion (0.8×), so every timing window lasts 25% longer.
- In most flips, the rider's head went through the ground. Now the rider spins around the hips.
After those fixes, the human-like bot survived a median of 53 seconds with 0 of 30 crashes. "The human-like bot survives at least 30 seconds" is now a permanent test for every new build.
Lesson two: bad players on purpose
Bots that play badly are just as useful. Every build is tested against `idle` (never taps), `spam` (taps every frame), and `masher` (8 taps per second). Each one has to die within 15 seconds or score under a fifth of what the scripted bot scores, so you can't win by ignoring or mashing the button. A `random` tapper rounds out the lineup.
Results for the current build, 20 rides of up to 60 seconds each (a different set of rides from the 30 above, so the human-like numbers differ a little):
+------------+----------------------------+-----------------+--------------+--------------+
| Bot | How it plays | Median survival | Median score | Median flips |
+------------+----------------------------+-----------------+--------------+--------------+
| idle | Never taps | 12.8 s | 40 | 0 |
| spam | Taps every frame | 2.2 s | 1 | 0 |
| masher | 8 taps per second | 2.3 s | 9 | 0 |
| random | Random taps | 6.2 s | 30 | 0 |
| human-like | Smart taps, 100 ms late | 46.3 s | 801 | 8.5 |
| scripted | Smart taps, perfect timing | 51.2 s | 1,391 | 14.5 |
+------------+----------------------------+-----------------+--------------+--------------+
Your turn
What's your high score? Did you beat the human-like bot? I'd love to hear if anyone lands three flips in one jump.
The MATLAB One-Button Game Mini Hack runs from November 2nd through November 29th, 2026! The challenge is to make a game that can be played with one button. I thought this would be a good opportunity to share some of the graphics and programming ideas behind Arcade, a collection of 15 games I developed in MATLAB using Claude Code and the MATLAB MCP Server.
Before the 2024 MATLAB Shorts Mini Hack, Tim Marston published his walkthrough of Little Nemo's airship. I learned a lot from Tim's article, and it inspired me to enter the contest. I hope that sharing some of what goes into these games might encourage a few more people to give it a try this time around as well.
You can download Arcade from File Exchange or GitHub, add the project folder to the MATLAB search path, and run:
Arcade();
Creating and updating graphics
Arcade sports neon-styled graphics. This is accomplished by layering ordinary MATLAB graphics against a dark background. The bright colors used come from the shared GameBase palette, which enables the different games to keep a consistent appearance and accounts for quite a bit of that look.
In Pong, the ball is made by drawing a colored dot, a semitransparent scatter marker and a smaller white dot at the same position. The ball and trail change from cyan through green, gold and red as the ball speeds up, while the center stays white. Against the dark background, these simple shapes give us the bright center and surrounding glow.
arcadeBallDemo();
worldSize = [854, 480];
position = worldSize / 2;
cyan = [0, 0.92, 1];
radiusParameter = max(3, round(min(worldSize) * 0.035));
baseCoreSize = 2 * radiusParameter;
baseDotSize = 5 * radiusParameter;
baseGlowArea = pi * (1.25 * radiusParameter)^2;
coloredDot = line(ax, NaN, NaN, ...
Marker=".", MarkerSize=baseDotSize, LineStyle="none", Color=[cyan, 0.15]);
glow = scatter(ax, position(1), position(2), baseGlowArea, cyan, "filled", ...
MarkerFaceAlpha=0.4);
core = line(ax, position(1), position(2), ...
Marker=".", MarkerSize=baseCoreSize, LineStyle="none", Color=[1, 1, 1, 1]);
ballLayers = {coloredDot, glow, core};
Here's how the ball looks as each layer is added:
Setting NextPlot="add" preserves the existing graphics as we add the layers. Specifying the axes limits by using XLim and YLim keeps the view in place as the ball moves.
Each drawing call returns a handle to its object. With those handles stored in ballLayers, moving the ball involves assigning its new position to the existing objects:
for layerIndex = 1:numel(ballLayers)
ballLayers{layerIndex}.XData = position(1);
ballLayers{layerIndex}.YData = position(2);
end
drawnow
MathWorks' graphics-performance guidance recommends retaining graphics and updating the properties that change. Pong also checks whether its scores have changed before assigning new score text. Its trail uses preallocated position buffers, wrapping around as new positions arrive, and two retained lines display that history as a broad dim stroke with a narrower bright stroke.
Handling inputs using callback functions
The ball example responds to mouse presses through a callback function. After creating the scene, it assigns a function handle to the figure's WindowButtonDownFcn callback property:
fig.WindowButtonDownFcn = @reverseVerticalDirection;
MATLAB executes the callback function when a mouse button is pressed in the window. In the example, it is nested inside the local animateBall function:
function reverseVerticalDirection(~, ~)
velocity(2) = -velocity(2);
end
The two arguments supplied by MATLAB are the source object and event data. The ~ placeholders indicate that this function doesn't use them. As a nested function, it can access velocity in the enclosing function. Each mouse press reverses the shared vertical velocity, which the next loop iteration uses to move the ball.
Arcade routes input through methods on the launcher. Its figure's KeyPressFcn callback property is assigned as follows, where obj is the launcher instance:
obj.Fig.KeyPressFcn = @(~, event) obj.onKeyPress(event);
The anonymous function forwards the event data to onKeyPress, where its Key property identifies the key. The launcher handles its own controls and passes appropriate input to the active game. Button release and scrolling have their own callback functions.
Mouse movement can control a paddle or aim a shot. To track the pointer as it moves within the figure, Arcade assigns a callback function to WindowButtonMotionFcn:
obj.Fig.WindowButtonMotionFcn = @(~, ~) obj.onMouseMove();
function onMouseMove(obj)
pointer = obj.Ax.CurrentPoint;
obj.MousePos = pointer(1, 1:2);
end
The axes' CurrentPoint gives the pointer position in game coordinates. MATLAB accounts for the axes' position and size, so the same code works when the playfield is resized or surrounded by margins.
On each game update, the launcher passes the stored position to the active game's onUpdate method. In Pong, obj now refers to the game instance. Once the incoming pos is available, its vertical coordinate determines the player's paddle position:
yLimits = obj.DisplayRange.Y;
if ~any(isnan(pos))
obj.PlayerPaddleY = max(yLimits(1) + obj.PaddleHalfH, ...
min(yLimits(2) - obj.PaddleHalfH, pos(2)));
end
The min and max calls keep the paddle's center at least half a paddle height from either edge. Moving the pointer into a margin therefore leaves the whole paddle inside the playfield. The graphics update then applies that position to the existing paddle patches.
The animation loop's drawnow displays pending graphics changes and processes pending callbacks. Keeping callback functions short allows control to return to the loop promptly.
Animating visual effects
When the orb is launched in Flick It!, a call to spawnHitEffect creates a ring and rays that expand and fade. The launcher advances these shared effects alongside the active game.
In arcadeEffectsDemo.m, each click creates a burst at the mouse position, allowing several bursts to expand and fade together. To try it out, download the function, add its folder to the MATLAB path, and run:
arcadeEffectsDemo();
The nested callback function combines mouse-press handling with the mouse-position tracking introduced earlier. Here's the part that creates and stores a new burst:
fig.WindowButtonDownFcn = @createManualImpact;
function createManualImpact(~, ~)
point = ax.CurrentPoint;
position = point(1, 1:2);
if any(~isfinite(position)) || any(position < 0 | position > worldSize)
return
end
[newRing, newRays] = createEffect(position);
activeEffects(end+1) = struct('Ring', newRing, 'Rays', newRays, ...
'Center', position, 'AgeSeconds', 0, 'LastTime', toc(animationClock));
end
The bounds check ignores clicks outside the playfield. createEffect creates the lines, and activeEffects stores their handles, center and timing state. The animation loop updates each burst independently, allowing further clicks while earlier bursts are still expanding.
We can drive the geometry, transparency and line width from a common progress value: 0 when the burst begins and 1 at the end of its lifetime. Inside the function, the ring is updated as follows:
progress = ageSeconds / lifetimeSeconds;
eased = 1 - (1-progress)^2;
opacity = max(0, 1-1.5*eased);
ringRadius = radiusParameter * (1+2.5*eased);
ring.XData = center(1) + ringRadius*ringCos;
ring.YData = center(2) + ringRadius*ringSin;
ring.Color = [cyan, 0.8 * opacity];
ring.LineWidth = max(0.5, 3*(1-eased));
The ease-out expression starts the expansion quickly and slows it down toward the end. The burst fades out before its nominal lifetime finishes. ringCos and ringSin hold the circle's fixed directions, calculated before the loop. Arcade controls the transparency of these lines by assigning an RGBA value to Color, with alpha as the fourth component, ranging from 0 (clear) to 1 (fully opaque).
The row vectors rayCos and raySin contain the eight ray directions. Each ray joins the inner and outer radii, which increase as the burst expands:
innerRadius = radiusParameter * (0.5+1.5*eased);
outerRadius = radiusParameter * (1+3*eased);
for rayIndex = 1:numel(rays)
rays(rayIndex).XData = center(1) + [innerRadius, outerRadius]*rayCos(rayIndex);
rays(rayIndex).YData = center(2) + [innerRadius, outerRadius]*raySin(rayIndex);
rays(rayIndex).Color = [cyan, 0.6 * opacity];
rays(rayIndex).LineWidth = max(0.5, 2*(1-eased));
end
The lines are created when the burst begins, updated throughout its lifetime and deleted when it expires.
Display scaling: Keeping the scene in proportion
Try making the ball example's window wider while keeping its height the same. Space appears beside the playfield, and the ball still travels between the same boundaries. The example's reference window is 854 by 480 pixels, matching worldSize numerically. Resizing scales the playfield on the screen while the game's coordinate system remains unchanged.
The figure and axes use Units="pixels". Taking the min of the width and height ratios relative to the reference window gives the largest scale factor that fits in both directions. Applying that same scale to both dimensions preserves the playfield's 16:9 aspect ratio.
For a 1200-by-480 window, the height limits the scale to 1. The fitted area remains 854 by 480, leaving (1200-854)/2 = 173 pixels on each side. In a 1280-by-720 window, the playfield grows to approximately 1.5 times its reference width and height. Here are four window sizes at a common page scale:
MATLAB already scales graphics defined in data coordinates as the axes resize. For example, a paddle drawn with patch grows with the axes. In the example, marker sizes, font sizes and line widths use points, while scatter sizes use points squared.
To make a circle twice as wide, we need four times its area. Doubling SizeData multiplies its width by only √2:
fig.SizeChangedFcn = @fitScene;
The nested fitScene callback function combines the axes calculation with these size updates:
function fitScene(~, ~)
if ~isgraphics(ax)
return
end
availableSize = max(fig.Position(3:4), [1, 1]);
displayScale = min(availableSize ./ worldSize);
fittedSize = worldSize * displayScale;
ax.Position = [(availableSize - fittedSize)/2 + 1, fittedSize];
% Always derive sizes from their reference values
core.MarkerSize = baseCoreSize * displayScale;
coloredDot.MarkerSize = baseDotSize * displayScale;
glow.SizeData = baseGlowArea * displayScale^2;
instruction.FontSize = baseFontSize * displayScale;
playfield.LineWidth = 0.5 * displayScale;
end
The first two entries of ax.Position set the left and bottom position, splitting the unused space equally between opposite sides. The last two give the fitted width and height. The axes limits remain unchanged, so movement and collision calculations keep using the same game coordinates.
At scale 1.5, the marker sizes are multiplied by 1.5 and the scatter area by 2.25. The updated sizes are calculated from their stored reference values. A 34-point reference marker becomes 51 points at scale 1.5, then 68 points at scale 2.
You can also use Units="normalized" to size and position the axes relative to the figure. Arcade's axes setup uses [0 0 1 1] to allocate the full available area:
obj.Ax = axes(obj.Fig, "Units", "normalized", "Position", [0 0 1 1]);
obj.Ax.XLim = obj.DisplayRange.X;
obj.Ax.YLim = obj.DisplayRange.Y;
Arcade's resize method obtains the available dimensions in pixels with getpixelposition and calculates the same display scale. It uses pbaspect to preserve the playfield's aspect ratio. MATLAB fits the playfield within the available space, leaving margins where needed:
axPx = getpixelposition(obj.Ax);
newPs = min(axPx(3) / 854, axPx(4) / 480);
gameAR = diff(obj.DisplayRange.X) / diff(obj.DisplayRange.Y);
pbaspect(obj.Ax, [gameAR 1 1]);
For Arcade, I also wanted markers, text and line widths to grow proportionally with the playfield. MATLAB also supports automatic resizing for text through FontUnits="normalized", which expresses font size as a fraction of the axes plot box height. This works well for text in a playfield with a fixed aspect ratio. Arcade's shared resize helper modifies current property values, using the ratio of the new scale to the previous scale; for example, moving from 51 to 68 points requires multiplying by 2/1.5. The function applies a common display scale to text, markers and line widths, in order to preserve their relative sizes as the window is resized.
FPS scaling: Let elapsed time determine the motion
Suppose that the ball moves 5 game units on each update. At 60 updates per second it travels 300 units in a second, but at 30 it travels only 150. Calculating the state, drawing and processing input all contribute to the time between updates, so we need to measure that interval to maintain a consistent speed.
Gianluca Carnielli discusses the relationship between time and animation frames in Animating Science: Creating Time-Sensitive Animations with MATLAB. In his example, he explains how to create a frame-based animation from time-based coordinates, by choosing the video’s duration and frame rate, calculating the simulation timestep between two consecutive frames, and resampling the coordinates using retime.
In a live game, we process player actions and calculate each new state using the time elapsed since the previous update. Arcade expresses many motion parameters per reference interval, using 60 frames per second (FPS) as its reference frequency. An elapsed time of 1/60 seconds represents 1 reference interval, while 1/30 seconds represents 2. In the example, we define the velocity and reference frequency before the loop and start the clock with tic once the initial scene is displayed:
velocity = [5, 3]; % Game units per reference interval
referenceFPS = 60; % Reference intervals per second
animationClock = tic;
previousTime = 0;
currentTime = toc(animationClock);
dtSeconds = min(currentTime - previousTime, 0.1);
previousTime = currentTime;
dtScale = dtSeconds * referenceFPS;
position = position + velocity * dtScale;
Here dtSeconds is in seconds, and dtScale counts reference intervals. A horizontal velocity of 5 now gives a movement of 10 game units over 1/30 second. Taking 30 steps of 10 units covers the same 300 units as 60 steps of 5 units. The samples are farther apart, so the motion can look less smooth:
The min(..., 0.1) caps a single represented interval at 0.1 seconds (10 FPS) to limit jumps after a stall. If an update arrives after 0.5 seconds, the code advances only 0.1 seconds, and the remaining 0.4 seconds are discarded.
The reference frequency is 60 FPS; the actual update rate depends on execution. We could also use game units per second and write position = position + velocityPerSecond * dtSeconds.
Arcade schedules its updates with a timer. The callback function assigned to its TimerFcn property calls the launcher's frame-update method:
obj.RenderTimer = timer(ExecutionMode="fixedSpacing", ...
Period=0.02, TimerFcn=@(~, ~) obj.onFrame());
start(obj.RenderTimer)
With fixedSpacing, the period begins after the preceding callback function finishes. The function's own work therefore contributes to the interval between updates. The launcher measures the full interval and passes the resulting time scale to the active game.
arcadeBallDemo(Inf, 0.04);
Its measured interval includes both the pause and the surrounding work. Both Arcade and the example use the full drawnow command. The limitrate option can reduce rendering work by limiting display updates to 20 per second; its suitability depends on the animation's required display rate.
Flick It! multiplies its velocity by friction^dtScale, where friction is the fraction of velocity retained over one reference interval. Over 2 intervals, the multiplier is the square of that factor, equivalent to applying it once in each interval. This makes the velocity decay follow the elapsed time.
For accelerating objects, the numerical approximation can still change with step size, and collisions need to account for movement between samples. Pong checks when the ball crosses the paddle's front face and interpolates its vertical position there, allowing a contact to be detected between two displayed positions.
Organizing games using classes
Arcade's registry stores a constructor function handle for each game. When a game is selected, the launcher calls the corresponding constructor to instantiate the game class. Here's an adaptation of the launch sequence, with Pong selected explicitly and obj referring to the launcher:
createGame = @games.Pong;
game = createGame();
game.FontScale = obj.FontScale;
game.init(obj.Ax, obj.DisplayRange)
game.beginGame()
obj.ActiveGame = game;
Each game subclasses engine.GameBase. The shared init method sets up the session and calls the game's onInit, which creates its state and graphics. Assigning FontScale beforehand lets newly created markers and text use the current display size. The launcher stops its render timer before creating the game and restarts it after the initialized game is assigned to ActiveGame. This prevents timer callbacks from running while the game’s state and graphics are being created.
GameBase declares four abstract methods that each game class must implement:
The launcher calls onUpdate on its active game, and MATLAB dispatches that call to the selected subclass. Pong implements paddle and ball movement; Flappy Bird advances its bird and pipes. All games inherit methods for scoring, combo handling and visual effects from GameBase. Window management, menus, timing and common services stay with the launcher.
As a handle class, GameBase lets the launcher keep a reference to the same game instance while its state changes. Position arrays, velocities, scores and graphics handles belong to that instance. Property access attributes allow the host to read the common score while leaving changes under the control of the game classes.
Flappy Bird creates its pipe patches in onInit, then reuses them throughout the game. Their handles and state are stored in private properties of the game instance. Its onUpdate method moves the active pipes across the playfield. Spawning a pipe selects an available pair, assigns its position and gap, and adds it to the active list. Once it leaves the playfield, the pair is hidden and its slot becomes available again:
In the default gravity-and-flap mode, the onKeyPress input handler sets a pending flap flag. The next update applies the upward impulse and clears the flag.
When the window closes, Arcade stops and deletes its timer, calls the active game's onCleanup method and clears its effects before deleting the figure. The smaller examples use MATLAB's onCleanup object to delete their own figure when the function exits. The requestClose callback function assigned to the figure’s CloseRequestFcn property sets a stop flag, letting the current update finish before cleanup removes the graphics.
Recent work on MATLAB graphics and performance
The MathWorks graphics team has also been improving the underlying rendering system. Maddie Moon's article about WebGL describes the R2025a transition, which brought the renderer already used by uifigure to regular MATLAB figures, along with work on stability, consistency and smoothing. Will Jackson also explains related changes to built-in panning, zooming and rotation in his article on the new interactions system.
R2025b focused on quality and stability improvements, building on the changes introduced in R2025a. It also improved performance when capturing successive frames in a loop with getframe, the function used by Arcade’s recording helper to capture and store frames for GIFs and videos.
R2026a included memory and stability improvements for graphics such as scatter and surface plots, described in Adam Danz's overview and the release notes. In addition, Jessa Karlberg describes the redesigned axes toolbar in R2026a, which allows seamless hovering without the animation slowdown issues caused by the toolbar’s continuous mouse tracking in earlier releases, eliminating the need to remove the toolbar as a workaround.
R2026b improved responsiveness for large data visualizations, with plot and plot3 being able to display lines and markers incrementally and shortening the wait time for the first visible results. MathWorks also reports automatic selection of a dedicated rendering GPU on supported Windows and Linux systems with multiple GPUs. These changes are documented in the release notes and the GPU-selection update.
info = rendererinfo;
info.GraphicsRenderer
info.RendererDevice
Mike Croucher also describes a new object-management system for R2026b, which can speed up object-oriented MATLAB code by reducing the cost of creating and deleting objects, as well as accessing their properties. The system is currently available as a beta and can be enabled through a package on File Exchange. Its documentation explains the changes to object lifecycle and compatibility.
Developing MATLAB code with AI
Arcade was developed with Claude Code through the MATLAB MCP Server, and the contest welcomes AI assistance as an option.
When working with an AI assistant, consider sharing your project's requirements and any existing code it can build on. Arcade's developer guide, for example, describes the shared game interface and explains how to add a new game. For your own game, you could begin with the main interaction and add features gradually, playing and improving it as you go. When something goes wrong, describe what happened and how to reproduce the problem, such as a ball passing through a wall at higher speeds.
The MATLAB Agentic Toolkit provides workflows for code review, analysis and testing. Its setup guide covers connecting MATLAB to AI applications including Claude Code, GitHub Copilot, OpenAI Codex, Gemini CLI and Amp, and installing MATLAB-specific skills. Through the MCP server, the AI assistant can run MATLAB code and tests, inspect the results, and use them to help diagnose problems and revise the code.
Several implementation choices in Arcade align with the MATLAB Coding Guidelines and performance rules: using a handle class for an object whose identity persists, restricting access to its state, retaining graphics, preallocating known-size arrays and cleaning up timers. The two downloadable examples presented above were checked with MATLAB Code Analyzer and tested for resizing, input, handle reuse and cleanup.
I hope that this article gives you something to experiment with, and perhaps a starting point for the One-Button Game Mini Hack. I look forward to seeing what you make!
What a great first day of MATLAB EXPO Online 2026! Thanks to everyone who joined us, participated in the sessions, and helped make Day 1 such a great experience.
We’re not done yet. Join us tomorrow, November 5, for another full day of great sessions, ideas, and inspiration.
What are you most looking forward to on Day 2? Let us know in the comments!
