Lesson 3 — Build a Rock–Paper–Scissors Image Model

By /
⚠️ Save before you close either website. Your eLearning progress does not save your editable MRT AI Studio or Teachable Machine project. Download the Studio project, save the Teachable Machine project file, and keep the model URL separately.

Lesson 3 — Build a Rock–Paper–Scissors Image Model

What you will make: Train an Image Model to recognise Rock, Paper and Scissors, upload it, connect it to MRT AI Studio and test the complete code.

Use an Image Project in this lesson. Pose Projects begin in Lesson 4.

1. Open the Extension Library

In MRT AI Studio, select the extension button in the lower-left corner.

Step 1 — Open the MRT AI Studio extension library

2. Add Image Project (TM)

Select the blue Image Project (TM) extension. Do not select Pose Project for this practice activity.

Step 2 — Add the Image Project TM extension

3. Prepare the Model Download Block

Drag the Download training model block to the code canvas. You will paste your model URL into this block after training.

Step 3 — Add the model download block

4. Start a New Image Project

Open Teachable Machine from the Studio menu and select Image Project. This opens Google Teachable Machine in a new tab.

Step 4 — Open a new Image Project

5. Create the Three Classes

  1. Rename Class 1 to Rock.
  2. Rename Class 2 to Paper.
  3. Select Add a class and name it Scissors.

Use these exact English class names. MRT AI Studio uses the class names to decide which event should run.

Step 5 — Empty classes in Teachable Machine

Collect useful training images

  • Use Webcam and collect approximately 80–120 varied images for each class.
  • Move the hand left, right, higher, lower, closer and farther away.
  • Include small natural changes in angle while keeping the whole hand visible.
  • Do not record Paper or Scissors inside the Rock class.
  • Keep the three class sizes reasonably similar; they do not need to match exactly.

6. Check the Training Images

The screen below is an example target. Your image counts and camera background may be different.

Step 6 — Rock Paper Scissors training samples

Before training: delete samples with the wrong sign, a missing hand, severe blur or most of the hand outside the frame. A model learns from the examples you keep.

7. Train the Image Model

Select Train Model, keep the Teachable Machine page open and wait until training is complete. If you add or remove images later, select Train Model again.

Step 7 — Trained Rock Paper Scissors model

8. Test Before Exporting

Use Preview to show Rock, Paper and Scissors several times. Test different positions and distances. Aim for at least 8 correct results out of 10 for each class; one high confidence result does not prove the whole model is reliable.

Step 8 — Test the model and select Export Model

9. Upload the Trained Model

Select Export Model, keep TensorFlow.js and Upload (shareable link) selected, then select Upload my model.

Step 9 — Upload the Teachable Machine model

10. Copy and Save the Model URL

Wait for the upload to finish, then copy the shareable URL. Keep the complete URL, including the final slash.

Step 10 — Copy the shareable model URL

Save two different Teachable Machine items:

  1. Editable project: use the ☰ menu and download the project file so you can reopen and retrain it later.
  2. Model URL: save the URL shown above so MRT AI Studio can load the trained model.

11. Paste the URL into MRT AI Studio

Return to MRT AI Studio and paste your Image Model URL into the Download training model block.

Step 11 — Paste the model URL into MRT AI Studio

12. Add the Camera Extension

Add the Camera extension. The camera must be on before Image Classification can check live images.

Step 12 — Add the Camera extension

13. Build and Test the Complete Code

Build the blocks as shown. The green-flag stack downloads the model, turns on the camera and repeatedly classifies the image. Rock, Paper and Scissors each use a separate event stack.

Complete Rock Paper Scissors Image Classification code with camera and forever loop

Final test: select the green flag and show all three signs. If recognition is unreliable, return to Teachable Machine, add varied examples of the failure, retrain, update the cloud model and test again.
Optional improvement: Background / No Hand

Add a fourth class containing the empty scene, a neutral hand and transitions between signs. This can reduce false actions when no clear gesture is shown. Because the existing example uses three classes, complete the basic activity first.

Next: Train a Pose Model

You have learned the complete model workflow with a small Image Project. In Lesson 4, you will apply the same data-quality process to body poses for your game or robot.

Scroll to Top