B4 · Domains of AI: Vision, Language, Speech and Robotics

Artificial Intelligence in Robots

AI gives a robot the ability to cope with situations that were not planned in advance. It helps the robot perceive its surroundings, work out where it is, plan a path, handle objects and move safely near people, and to improve these skills from data.

6 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

Definition

A factory robot can weld a car perfectly but cannot pick up a sock from the floor. Why is the sock harder?

AI gives a robot the ability to cope with situations that were not planned in advance. It helps the robot perceive its surroundings, work out where it is, plan a path, handle objects and move safely near people, and to improve these skills from data.

In simple words

A robot without AI is like a dancer who has memorised one routine: perfect on the same stage every night, and lost if a chair is moved. A robot with AI is more like a waiter in a crowded restaurant, who looks around, picks a path between tables and changes it when someone stands up.

Why it matters

The real world is untidy and keeps changing. AI is what lets robots leave fenced-off factory areas and work in warehouses, hospitals, fields and homes. It also explains why robots still struggle with things a small child finds easy.

Key terms

Perception
Turning raw sensor readings into knowledge of what is around the robot.
Localisation
The robot working out where it is.
Mapping
Building a map of the surroundings from sensor data.
Path planning
Choosing a safe route from where the robot is to where it must go.
Manipulation
Picking up, holding and moving objects.
Simulation
A computer model of the world in which a robot can practise safely.

Explained in detail

Mapping and locating at the same time

Put a robot vacuum in a house it has never seen. To clean efficiently it needs a map of the rooms, and to build a map it needs to know where it is. But to know where it is, it needs a map. Roboticists solved this puzzle with methods that do both together: the robot moves a little, notes what its sensors see, estimates how far it has travelled, and gradually builds a map while tracking its own position on it. Each time it passes a place it recognises, it corrects the small errors that have built up. The same idea guides warehouse robots, delivery drones and self-driving cars.

Learning to pick things up

Grasping an object is astonishingly hard. The robot must judge the object's shape, decide where to place its fingers, close them with enough force to hold and not enough to crush, and adjust if the object slips. Writing rules for every object is impossible. Instead, robots learn. In some laboratories, arms have practised hundreds of thousands of grasps on bins of mixed objects, recording which attempts succeeded. From this experience a model learns to predict a good grasp for an object it has never seen.

Practising in a virtual world

Learning by trial and error is slow and risky on a real machine. A robot that falls a thousand times while learning to walk would be destroyed. So much of the practice happens in simulation, a computer model of the robot and its surroundings with realistic physics. Thousands of simulated robots can practise at once, far faster than real time, and a crash costs nothing. The difficulty is that a simulation is never exactly like reality: real floors are slippery, real motors are imperfect. Skills learned in simulation often need further adjustment before they work on the real machine.

Learning from people

Another route is to learn from demonstrations. A person guides the robot's arm through a task, or controls it remotely while it records everything, and a model learns to imitate the movements. Recent research goes further and connects robots to large models trained on language and images, so that a robot can be told in plain words, "put the red cup on the top shelf", and work out the steps. This is progressing quickly in laboratories and is still far from reliable in ordinary homes.

Where AI robots already work

SettingWhat the robot doesWhy AI is needed
WarehousesCarries shelves and sorts parcelsRoutes change constantly; other robots and people are moving
FarmsSprays weeds, picks fruit, monitors crops from the airEvery plant and fruit is different
MarsDrives across rocky ground between instructions from EarthRadio signals take many minutes, so it must avoid hazards itself
Disaster sitesSearches rubble and unsafe buildingsNo map exists, and conditions are unknown

Responsibility for a moving machine

A robot adds physical risk to every limitation of AI. If a vision model mistakes a person for a box, the consequence is no longer a wrong label but a collision. Robots that share space with people are therefore built with several layers of protection: limits on speed and force, sensors that stop movement on contact, emergency stop buttons, and extensive testing. Someone must also be accountable when a robot causes harm: the maker, the owner or the operator. The more a robot decides for itself, the more important it becomes to settle that question before it is switched on.

How it works

SkillQuestion the robot must answerHow AI helps
PerceptionWhat is around me?Vision models recognise shelves, boxes and people
LocalisationWhere am I?Combines sensor readings with a map to estimate position
PlanningHow do I get there?Searches for a short, safe path and re-plans when blocked
ManipulationHow do I pick this up?Learns grips for objects of different shapes
SafetyIs a person too close?Detects people and slows down or stops

Case study: a warehouse robot

In a large online-shopping warehouse, hundreds of low, flat robots carry shelves of goods to human packers. Each robot knows the warehouse map and reads markers on the floor to track its position. A central system assigns jobs and plans routes so that robots do not collide or block each other. When one robot stops unexpectedly, the others are re-routed around it in seconds. People do what the robots cannot: pick items of every shape from the shelf, notice damaged goods and solve unusual problems.

Why easy things are hard for robots

Researchers noticed long ago that tasks people find hard, such as chess or long calculations, are fairly easy for computers, while tasks a one-year-old manages, such as walking over toys or picking up a soft cloth, are extremely hard for robots. This is known as Moravec's paradox. A sock can lie in endless shapes, it is soft, and gripping it needs a sense of touch. Welding a car is the same precise movement every time.

Try it yourself

A warehouse robot meets a surprise

A robot must carry a shelf from aisle 4 to a packing station.

Easy or hard for a robot?

Think about how much the task changes each time.

  • Walking through a crowded vegetable market

  • Carrying a shelf along a marked path

  • Tightening the same bolt 2,000 times a day

  • Picking ripe tomatoes without squashing them

  • Painting car doors on a production line

  • Folding a pile of mixed clothes

Common mistakes

  • Thinking robots are close to doing every household chore. Handling soft, varied objects is still very hard.
  • Thinking AI in a robot means it thinks like a person. It runs separate models for seeing, locating and planning.
  • Forgetting safety. Physical machines need stricter testing than software alone.

Check yourself

Question 1

1. What does AI mainly add to a robot?

Question 2

2. What is localisation?

Question 3

3. Why is picking up a sock harder for a robot than welding a car?

Summary

  • AI helps robots perceive, locate themselves, plan, handle objects and stay safe.
  • Repeated tasks in controlled places are easy; varied tasks in the messy real world are hard.
  • Robots are physical, so safety and human supervision matter even more.

Finished reading?

Sources: AI4K12: Five Big Ideas in AI · Elements of AI (University of Helsinki)