What is a humanoid robot?

Look around the room you are sitting in. The door handle is at the height of your hand. The steps outside are cut for the length of your leg. The tap, the switch, the shelf and the spanner were all made for a body with two arms, two hands and eyes on top.

A humanoid robot is a machine built to that same plan. It has two legs, a trunk, two arms, and a head with cameras in it. The reason is plain. A machine shaped like a person can use the world that people already built, so nobody has to rebuild the building.

That is the promise. The catch is that a human body is very hard to copy.

What counts as a humanoid robot

The word describes the body, and it says nothing about how clever the machine is. A HUMANOID IS A SHAPE, NOT A SKILL. Some humanoids can barely stand. Some can carry a crate across a warehouse.

Most have a head, a trunk, two arms and two legs. Many useful ones stop at the waist and sit on a wheeled base. Wheels are cheaper, steadier and far kinder to a battery than legs. Engineers still call those humanoids, because the working half is the human-shaped half.

You may also meet the word android. That usually means a humanoid with a face and skin, made to look like a person rather than to work like one. However, most machines built for real jobs make no attempt to look human.

Why anyone builds a human shape

Ask the awkward question first. Wheels beat legs on flat ground, and a fixed arm is cheaper than a whole body, so why bother?

The answer is the world we already have. Stairs, ladders, narrow doorways, high shelves, hand tools and vehicle pedals were all sized for people. Changing a whole factory to suit a machine costs money. Changing a machine to suit the factory does not.

There is a second reason. One human-shaped machine could, in theory, do many jobs in one week. It could unload a lorry on Monday and stack shelves on Tuesday. A machine that does one job well has to be sold for that one job. That is the bet, and it is still a bet.

How it stays upright, step by step

Standing still looks easy and it is not. Watch a person stand on a bus and you will see small corrections without pause. A humanoid does the same, only faster.

1. Sensors in the trunk report tilt and turn many times a second.
2. Sensors in the feet or ankles report how weight is pressing on the ground.
3. Encoders in every joint report the exact angle of every limb.
4. The controller works out where the weight of the whole body falls.
5. If that point drifts towards the edge of the feet, the robot is starting to fall.
6. Motors move the hips, ankles and arms to pull the weight back over the feet.

That loop runs hundreds of times a second, without a break, even when the machine is doing nothing else. Walking is the same trick with the brakes off. The robot lets itself fall forward and puts a foot out in time to catch the fall. That is why a small push, or a wet patch, can put a humanoid on the floor.

How a humanoid learns to move

Some movements can be written down as rules. Reaching for a fixed point is one. Balancing on gravel or turning an odd shape in the fingers is another matter, because no engineer can write every case down.

So the training moved into simulation. A model of the robot practises inside a computer, thousands of times faster than real life. It earns points for staying upright and moving well. That method is reinforcement learning, and it now sits behind much of what you see in walking demos.

The trained behaviour is then copied onto the real machine. This is the hard step, because the real world is never quite the simulation. Floors are slippery, motors run warm, and parts wear. Engineers call the trouble the reality gap. The usual cure is to train across thousands of slightly different pretend worlds, so the skill does not depend on any one of them.

For hands and tools there is a second source of skill. A person wears trackers, or holds controllers, and drives the robot through the job. The robot records what was done and learns from many such runs. This is also why some demo videos are less impressive than they look, since a human may be driving.

What they are good at today

Moving through human spaces is the first real strength. A humanoid can open a normal door, climb normal stairs and stand at a normal bench. Other robots, however, need ramps, rails or a floor cleared for them.

Carrying and placing is the second. Lifting a box from a pallet and setting it on a belt is now within reach for several machines, at a steady, unhurried pace.

Inspection is the third, and it is the quietest success. A machine that walks a fixed route through a plant, reading dials and listening for leaks, does useful work. This is because it needs almost no hand skill.

What they are still bad at

Hands come first. A human hand moves in more than twenty ways and feels pressure everywhere. Robot hands have far fewer joints and much less feeling, so soft, floppy and slippery things remain hard.

Power comes second. This is because legs, balance and arms all drain a battery. Working time is therefore counted in hours at best, and heavy work makes it shorter. Most machines need a charge or a battery swap during a shift.

Cost comes third. A humanoid carries dozens of motors, many sensors and a great deal of software. It therefore costs far more than an arm that does one job well. The job it replaces has to be worth that.

Safety comes last and matters most. A tall machine that can fall is a hazard near people. That is why humanoids in real workplaces are still kept in marked areas. A cobot, by contrast, is built from the start to share a table with you.

How to judge a humanoid demo video

Videos are the main way these machines reach the public, so learn to read them. Ask six questions.

• Is the clip sped up? Real speed is often much slower.
• Is it one unbroken shot, or a cut every few seconds?
• Is a person off camera driving the machine?
• Is the floor flat and clean, and the lighting perfect?
• Does it do the task once, or ten times in a row?
• Is there a cable or a safety strap holding it up?

Then ask the question that decides everything. How often does it fail across a full shift, and who picks it up when it falls? A machine that works nine times in ten is a fine demo and a poor employee. Look for that number the next time a new humanoid is shown, and notice how rarely anyone gives it.

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.