Machine Learning

Instead of telling your robot exactly what to look for, machine learning lets you show it thousands of examples and have it figure out the pattern itself. That single shift changes everything.

What is Machine Learning?

Machine learning (ML) is a branch of AI where a program improves its performance through experience rather than through rules you write by hand. You feed it labelled data — images of cats, sensor readings before a motor stalls, sonar distances paired with “obstacle / clear” decisions — and a training algorithm adjusts the model’s internal parameters until its predictions get good enough to be useful.

The three main flavours you’ll meet in robotics are:

• Supervised learning — you supply labelled examples (input → correct output). Most image classifiers and object detectors work this way.
• Unsupervised learning — the algorithm finds clusters or structure in unlabelled data. Useful for anomaly detection and mapping.
• Reinforcement learning — the robot tries actions, gets rewards or penalties, and learns a policy. This is how many walking and balancing controllers are trained.

The workflow

Every ML project, big or small, follows the same loop. You gather and label data, train a model on it, check how well it does, and only then deploy it to the robot — looping back to collect more data whenever it gets things wrong:

flowchart LR
    DATA["Collect data<br/>(photos, readings)"] --> LABEL["Label it<br/>(cat / not cat)"]
    LABEL --> TRAIN["Train model"]
    TRAIN --> EVAL["Evaluate<br/>(is it good?)"]
    EVAL -->|"not yet"| DATA
    EVAL -->|"good enough"| DEPLOY["Deploy to<br/>the robot"]
    DEPLOY --> PREDICT["Live predictions"]

The big mental shift from traditional coding is in the middle: you never write the rules yourself. The training step learns them from your labelled examples, which is exactly why messy, varied real-world data makes a stronger model.

Why robot builders care

Robots encounter a messy, unpredictable world. Writing if colour == "red": stop works in a lab; it falls apart the moment the lighting changes. An ML model trained on varied examples handles that variation naturally.

Concrete uses you’ll run into quickly:

• Object detection — spot a ball, a person, or an obstacle in a camera frame.
• Gesture or voice recognition — trigger actions from a wave or a word.

Even a small model trained on a few hundred images can be surprisingly capable once you understand the workflow: collect data, label it, train, evaluate, deploy.

Get started

The best first project is image classification because the feedback loop is fast and visual. Here is a concrete path:

1. Build a dataset — collect 50–100 photos each of two or three objects you care about. Your phone camera is fine.
2. Train on the Pi AI Camera — Kevin’s course Building Object Detection Models with Raspberry Pi AI Camera walks you through PyTorch training, optimisation for edge hardware, and deploying a live detector on real Raspberry Pi hardware.
3. Try it on a robot — the Elf Detector project shows how Kevin trained a model with Viam’s point-and-click training tool and then dropped it straight onto a robot. It is a great template for any “spot this thing and react” idea.
4. Go deeper with Computer Vision — once your model is running, pair it with OpenCV (see the Computer Vision with CVZone course) to add tracking, distance estimation, and more.

Python is the lingua franca here. scikit-learn covers classical ML (decision trees, k-nearest neighbours, SVMs) and is a great place to learn the fundamentals. PyTorch and TensorFlow handle deep neural networks. Start with scikit-learn on simple sensor data, then move to a vision model when you are ready — the concepts transfer cleanly.