Supervised learning

Learning from labeled examples: inputs paired with the correct answer. The model adjusts itself until its answers match yours — then generalizes to inputs it never saw.

The loop

1. Data — pairs of (input, label). On this platform: (camera frame, your steering button).
2. Training — the model predicts, is scored against your label (the loss), and nudges its weights to shrink the error. Repeat over the whole dataset, many times (epochs).
3. Inference — frozen weights, new inputs, instant predictions.

What actually matters: the data

• Coverage — the model only knows situations like the ones you labeled. All straight-road photos → helpless in curves.
• Consistency — if identical frames get different labels, you are teaching noise.
• Balance — 90% "straight" labels teach the model that "straight" is always a safe guess.

This is why the supervised racing course gets better when you label more curves — not when you train longer.

Limits

Supervised learning imitates its teacher — it cannot exceed you, and it needs you to label everything. When showing the right answer is impossible but scoring an attempt is easy, you want reinforcement learning.