Last year (2 months ago), I created a game using p5.js and ml5.js that uses Machine Learning to detect the ASL alphabet. I made this project for my course "IoT with Machine Learning" at my University and it taught me a lot about the process and how IoT is shaping the world!
One of the things that I utilize in this project is using MQTT (Message Queuing Telemetry Transport). It is different from Web Sockets because of how it is use for IoT specifically since it is designed for constrained devices such as Arduino, which I used for this project.
But I want to use MQTT for just the multiplayer functionality. I had some practice with Web Sockets, but thought to myself..."How far can I go with MQTT"?
Overview of the Project
This project is a desktop application. It is a simple game where your goal is to spell out as many words as you can in ASL. You can play either Single player or Multiplayer. It utilizes Machine Learning to perform Pose Estimation on the user's hand points using ml5.js. You can learn more about it on their documentation.
Technologies I Used
To make the desktop application, I used Tauri which is framework to create your own desktop/mobile app. It is very nice since you can do Web Development stuff inside of Tauri such as React, Vue, or just Vanilla HTML, CSS, and JavaScript. For this project, I stick to Vanilla HTML, CSS, and JavaScript. I also used p5.js and ml5.js for my libraries as I mentioned earlier. I also used SupaBase to store high scores of the words they accomplished.
ASL Machine Learning Model
This rough flow diagram shows the following of the game functionality. When you do ASL, it detects your hand points and the gesture you are doing. It then feeds the data points into the Machine Learning Model and it predicts which alphabet you are gesturing in real-time. Once the game is done, it uploads your data to the database of your score and coins you have received.
The Arduino part is an extension that shows you data, such as your "Letters per second" and your username, on an LCD which is shown here:
MQTT Multiplayer Functionality
To accomplished this, I use Mosquitto as my broker. This diagram shows the flow of how the Multiplayer game works:
In short, you can create a room and it will provide you a code to share with other users to join the room. In the lobby, when everyone clicks "Ready" the game starts. Demo is shown here below:
Challenges I Faced
There are are few big things I faced when creating this project.
1. MQTT setup
I had the idea of how it works, but it is hard to implement in practice since I am not really using an IoT device to communicate to the broker (Unless you count the camera as an IoT device). Since it was vibe coded, I had to take the time to understand the flow and how data is being pass/received.
2. Creating my own Model
I started with Image Classification. But the problem is that it captures the WHOLE image in addition to your hands. Meaning if I use the model in a different setting, it would not predict it accurately. I then learn about Hand pose in ml5.js documentation and I start building my own model. I had to create a separate vibe coded project that is design to create your own Hand pose Model.
3. Training problems
For context, I trained letters from A-Z ("J" and "Z" are excluded). For each letter, it captures around 2000 Post Estimation data. After training is complete, it forgets the last last two letters, which are "X" and "Y". The solution I came up with is that during data collecting, if I reach either "X" or "Y", collect more than 2000 Post Estimation data. In this case, "X" will get 4000 and "Y" gets 6000 (Just to be safe). If anyone knows why this is the case, please let me know!
Summary
It was a fun project! I learned a lot about not only how to connect the Arduino to the web, but also how to use MQTT to the limit (to my knowledge). If you are interested more about this project, check out the GitHub Repository here!
Any questions, or comments? I would love to hear from you!
Top comments (4)
Wow, Francis, that's incredible work! You've really made this project come together, and I'm loving the fact that you were able to get MQTT working for multiplayer functionality. I can imagine how tough it must have been to figure out the Hand pose model and train it, but you persisted and got great results.
Thanks Aryan! It did took me a bit since I was conformable doing Image Classification. What helped was watching an Overview of the Hand pose from The Coding Train. I think @richardpascoe would love this channel since he does web development related stuff in terms of Creative Coding.
But overall, I was happy to make the multiplayer game work since that was one of my goals I wasn't able to accomplish perfectly till now. Thanks for your reply!
Wow - just wow - congratulations on accomplishing that goal you set for yourself! Looking forward to learning so much more from you!! This is great!
Thanks for the heads-up, Francis! Will check it out, and re-read this post, when I have a little more time too! Sounds rather interesting though!