How Does a Humanoid Robot Work?

A humanoid robot is a machine built on the model of the human body: it has a torso, two legs, two arms and usually a head with cameras serving as eyes. This form is no accident. The world is designed by humans, for humans, so a robot with human proportions finds it easier to climb stairs, open doors or use tools designed for human hands.

Three layers: body, perception, decision

Every humanoid robot consists of three layers. The first is the body: electric motors (actuators) in the joints, gearboxes, batteries and the mechanical structure. The second is perception: cameras, force sensors, gyroscopes (IMUs) and sometimes LiDAR, which build a picture of the surroundings in real time. The third is the decision layer: software that uses sensor data to decide how to move each joint to keep balance and complete the task.

How does a robot know how to walk?

Modern humanoid robots learn to walk mainly through reinforcement learning in simulation: millions of virtual trials and errors before the algorithm is loaded onto a real machine. Thanks to this, the robot can keep its balance even on uneven terrain, though it still falls short of a human in completely unforeseen situations.

What are VLA models?

A newer generation of robot software (such as the Helix system from Figure, introduced in February 2025) is built on so-called vision-language-action models (VLA), which combine image recognition, understanding of natural-language commands and motion control in a single system. They are what makes possible the scenes increasingly common in press footage, where a robot picks up a specific object from a table on a spoken command.

Balance: why a robot on two legs does not fall over

Staying upright is the hardest part of walking. The robot measures the tilt of the torso with an inertial sensor (IMU), reads the position and torque in the joints and corrects the legs several hundred times a second so that the center of mass stays over the support base. Classical control rested on physical models (for example a simplified inverted pendulum), and newer control moves that work to a neural network trained in simulation. In practice the two approaches mix: the network decides the motion, and a safety layer watches the torque and speed limits. That is why manufacturers are so eager to show a robot recovering its balance after a push, because it is the simplest proof that the control works.

Hands and grasping: where the biggest gap remains

Walking is largely solved today, grasping is not. A human hand has dozens of degrees of freedom and a sense of touch, while robot hands range from a single two-finger gripper to designs with a dozen joints and pressure sensors. The more complex the hand, the more expensive, heavier and failure-prone it is, so cheaper and educational versions often have simpler grippers. How many degrees of freedom individual models have and what that number really means is explained in the article “Degrees of Freedom (DOF) in Humanoid Robots”.

Actuators and battery: physical limits

Every joint is a motor with a gearbox and sensors, and their quality decides the force, smoothness and noise of motion; we describe the details in the text on “Actuators in Humanoid Robots”. The second limit is energy: the robot carries its battery along with its own mass, so runtime is usually from about an hour to a few hours (for example about 2 hours for the Unitree G1 according to the maker, and 90 minutes for the announced Digit 5), and charging or swapping the battery interrupts work. That is one reason why the first factory deployments involve short, repeatable tasks and not entire shifts; more in the article “Humanoid Robots in Factories 2026”.

Where the robot gets its skills

Besides learning in simulation, imitation learning is used: a person controls the robot remotely (teleoperation), and the system records images and motions from which it then learns its own policy. The amount and quality of such data is one of the main bottlenecks today. That is why companies boast of hours of recordings and open datasets, and a buyer should ask what tasks the robot performs in practice, not only at a demo. Details on models that combine vision, language and motion are in the text “VLA Models: How AI Controls Humanoid Robots in 2026”.

A demo versus a deployment

Trade-show footage shows the result of many attempts, often with a prepared environment and an operator within reach. To judge maturity a few simple questions are worth asking: does the robot work without teleoperation, for how long without a break, what share of tasks ends in success, who is responsible for service, and is it available in Poland through a distributor. We collect the answers on the model pages, and a comparison of prices and running costs is in the article “Humanoid Robot TCO: How to Calculate Total Cost”; an overview of the robots themselves is in the category “Humanoid robots”.

Why this matters to a buyer

If you are looking for a robot for research or educational use, check which software stack (AI stack) the manufacturer offers and whether it is open to your own algorithms. This often matters more than the mechanical specifications alone. Full technical specifications of individual models can be found on their profile pages in our catalog.