How AI systems learn

Yesterday I briefly introduced the five levels of AI autonomy, capturing the range of actions and responsibilities robots and vehicles can take on—from fully human-controlled machines to systems capable of making independent decisions in dynamic environments.

So today, I want to dive into how AI-powered robots or vehicles are trained to learn how to do things—and just as importantly, what to do in any given situation.

Before we jump in, remember how I mentioned that Physical AI relies on perception and computation? To be more precise, there are three fundamental elements that make it all work: perception, decision-making, and action (also called execution or actuation). These form the core loop of any intelligent system—sense, decide, and act.

So, how do we actually teach machines to follow that loop? There are several approaches, but let’s start with two of the most widely used:

Supervised learning teaches machines by example—feeding them large amounts of labeled data so they can learn patterns and make predictions.

Reinforcement learning, on the other hand, is more like trial and error. The AI interacts with an environment, receives rewards or penalties, and gradually figures out the best way to act through experience.

Each training method comes with its own strengths and limitations, and choosing the right one depends largely on what you want the AI to learn.

Tomorrow, I’ll break down how these methods are applied in real-world AI systems—and what challenges we still need to overcome.