Machine learning for robots
BEGINNER · 6 MIN READ
Instead of writing every rule, you show the robot examples and let it find the pattern.
Normal code is rules you write: if the sensor reads below 10 cm, stop. Machine learning flips this. You give the computer many examples, and it works out the rules itself.
Think of teaching a child to spot a cat. You do not list the rules for whiskers and ears. You show many cats and not-cats until they get it. A robot learns the same way, just with numbers.
Imagine a robot sorting ripe fruit (orange) from unripe (blue). Each dot is one example it was shown. Click the picture to add examples. The coloured background shows what the robot would decide for any new fruit.
Try this: add 2 or 3 mislabelled examples inside the orange area and watch the blue patch appear. That is what bad training data does. Hold Shift and click to ask the robot about a new fruit. With no examples it can only guess.
To sort ripe tomatoes from green ones, you show a camera thousands of labelled photos. It learns what ripe looks like, even for tomatoes it never saw.
You can solve problems that are too messy to describe with rules, like telling ripe fruit from unripe in different light.
You would need to write rules for every lighting, angle and shade. That is impossible, and the robot breaks on the first case you missed.
Take 20 photos of two objects with your phone and notice how few examples already separate them.
Common questions
How much data does a robot need to learn?
It depends. Simple tasks may need a few hundred examples. Hard ones can need millions. More varied examples matter more than a huge pile of similar ones.
Can a learned model be wrong?
Yes. It can be confidently wrong on things unlike its examples, so safe robots keep simple safety rules on top.
Keep learning
Get your personal invite link to share R2BOT with friends.
Teachers and bloggers: paste the embed code into your page to show this demo for free.
Plain English. Free. Unsubscribe in one click.