What is a humanoid robot?

A humanoid robot has a torso, two arms, two legs and usually hands. It is a worse machine than a wheeled robot with one arm on almost every measure: slower, less efficient, more expensive, easier to knock over and far more complex to control.

The case for it has nothing to do with the robot. Our buildings are full of stairs, doorways, tool handles, shelves and pallets sized for a human body. Changing the machine is a purchase; changing the building is a project. A machine that fits what already exists can be deployed without an architect and moved to a different task in a day.

What is inside

Actuators. Twenty to forty motor and gearbox units, one per joint. They determine how much it can lift, how fast it moves and how long the battery lasts, and they are a large share of the cost.

Sensors. Cameras for vision, an inertial unit for balance, force sensing at the joints, sometimes depth sensors and touch sensors in the fingers.

Compute. A small fast computer on board running the control policy, sometimes with a larger model consulted over the network for planning.

Battery. The constraint nobody films. Two to five hours of light work is typical, after which it charges.

How it learns to do anything

Not by being programmed. A policy, which is a neural network, is trained on three kinds of experience: humans driving the robot remotely, video of people doing tasks, and physics simulations running thousands of copies in parallel. The policy takes camera images and joint positions and outputs the next motion, dozens of times a second.

This is the change that made the current wave possible, and it is covered in our read on robot foundation models.

What they actually do in 2026

In pilots, in small numbers, with human supervision available: moving containers between a conveyor and a cart, loading and unloading machines, inspecting parts, sorting items. All tasks a fixed robot could also do, in buildings where a fixed robot would have needed a cell and a fence.

What is still unsolved

The hands. Picking up something soft, transparent or unfamiliar, without crushing it, at better than 99 percent reliability, is the hardest remaining problem and the one that separates a demonstration from a deployment. Walking was solved years ago; grasping was not.

How to read a demo video

Look for cuts, a tether, and whether the robot ever works for more than thirty uninterrupted seconds. Ask whether it was teleoperated. And ask what the success rate is over a hundred attempts on objects the company did not choose.