15 STEM and AI kits that teach how machine learning works

Train image, sound or pose models in a browser with no code. The fastest way to show a kid that a model learns from examples and inherits their mistakes.

1. Pick two or three things to tell apart, like toy cars and blocks.
2. Record examples of each with the webcam.
3. Train and test, then try to fool it.
4. Add more varied examples and see it improve.

Kids wear or hold a micro:bit, record movements like waving or clapping, train a model, and then code what happens when it spots each one.

Watch for: Needs a micro:bit V2 and a laptop browser.

Kids train text, image or number models, then use them inside Scratch to build games and chatbots they fully understand. Seeing a model trained on 20 examples fail is the best lesson in AI limits.

Short, free activities such as training an AI to tell ocean trash from fish, plus lessons on bias and how AI decides. They are made for classrooms and work just as well at the kitchen table.

Free classroom-ready lessons from MIT on what AI is, how it learns and how to use it responsibly, sorted by grade. The activities are easy to adapt for a family afternoon at home.

Doodle and a neural network tries to guess what you drew before time runs out. Afterward, browse other people's drawings to see the data the model learned from, and why it gets confused.

Kids build a six-legged robot and teach it to respond to their own gestures and sounds. It is machine learning you can hold, and it shows that training data comes from people.

Its AI and data unit has students use ready-made classifiers and then train their own with bricks, a motor and a sensor. Work stays on the device, with no student logins or saved images.

Watch for: Built and priced for classrooms.

A modular robot with machine vision, speech recognition and self-driving projects, coded in Scratch and Python. Kids build a robot that sees and hears, then learn what is happening inside.

An all-in-one vision module for Makeblock robots that teaches recognition of objects, colors and lines. Attach it to an mBot2 and kids can build a car that follows signs or spots objects.

Watch for: Best with an mBot2 (about $150).

A camera sensor with more than 20 built-in models and support for training custom YOLO models. It works with micro:bit and Arduino, so a teen can add real vision to robots they already have.

Runs vision neural networks on the camera itself, so a teen can build a bird counter or a pet detector without sending video to the cloud.

An add-on board for the Raspberry Pi 5 that runs small language and vision-language models locally, with no network connection. It lets teens see how a chatbot works from the inside, privately.

Watch for: Needs a Raspberry Pi 5. Local models still make things up; teach checking.

Add a recognition camera to Petoi's robot dog and program it to react to faces or objects. It teaches how a robot connects what it sees to how it moves, the heart of robotics.

A small computer with a GPU for running real AI models at home, from vision to small language models. It suits the serious teen who has outgrown the Raspberry Pi and wants to go deeper.

Watch for: Sold through partners; setup takes Linux comfort.