At first glance a humanoid robot looks like a single machine, but it is really three systems working together: a body that can move, senses that perceive the surroundings, and software that decides what to do. This article walks through those three layers, and keeps a clear line between what works today, what is in development and what is still in the future.
Why a human shape?
Factories, warehouses, shops and homes are built for people. Stairs, door handles, workbenches and cupboards are sized for the human body. A robot with arms, hands and legs can, in principle, work in those environments without everything around it being redesigned. That is the main reason companies are investing in this form.
The International Federation of Robotics (IFR) takes a sober view. In its position paper of August 2025, the industry body writes that humanoids are not expected to replace existing robot types, but to complement them. A fixed robot arm on a production line will remain faster and cheaper for many tasks.
Layer 1: the body
Joints and actuators
Every joint of a humanoid robot is driven by an actuator: an electric motor with a transmission, sensors and control electronics. A conventional industrial robot arm usually has six or seven joints. According to research by Fraunhofer IPA and P3, a humanoid often has more than forty. For illustration: Unitree lists 23 degrees of freedom for the base version of its G1 and up to 43 for the research version. Boston Dynamics lists 56 degrees of freedom for the new electric Atlas.
Broadly, there are two types of actuator:
- Rotary actuators for shoulders, elbows, hips and torso. They pair a motor with a reducer, such as a harmonic drive (strain wave gear) or a planetary gearbox. The reducer turns the motor's high speed into usable force.
- Linear actuators for joints that need a lot of force, such as knees and ankles. These often use planetary roller screws, which convert rotation into a straight push or pull.
Actuators are the most expensive part of the robot. Schaeffler puts their share at roughly half of the total cost, Euronews reports. Analysts quote different percentages, but agree that this is the largest cost item.
Hands
Hands are the hardest part technically. A human hand has many small joints and a fine sense of touch. Manufacturers therefore compromise: simple grippers for industrial tasks, and multi-fingered hands with tactile sensors where finer handling is needed. When it introduced Figure 03, Figure AI described fingertip sensors that detect forces from 3 grams and a camera in each palm.
Energy
Batteries determine how long a robot can work. Unitree states about two hours for the G1; 1X states an 842 Wh battery and about four hours of use for NEO. To limit downtime, manufacturers are working on autonomous charging and battery swapping. Atlas drives itself to a station and swaps its own batteries, and Figure 03 charges wirelessly through coils in its feet.
Layer 2: the senses
A humanoid robot has to know where it is, what is happening around it and what position its own body is in. It combines several sensors for this:
- Cameras, often stereo or with depth sensing, to recognise objects and people.
- LiDAR, a laser scanner that measures distances in 3D. The G1, for example, has a depth camera and a 3D LiDAR.
- Inertial measurement units (IMUs) that measure acceleration and tilt. They are essential for balance.
- Encoders and force sensors in the joints, so the robot knows the position of each joint and how much force it is applying.
- Microphones and speakers for voice interaction.
All that data has to be processed within milliseconds. That is why humanoid robots carry powerful onboard computers, often with dedicated AI chips.
Layer 3: the software
Balance and movement
Walking on two legs is a form of controlled falling. The controller continuously calculates how the robot should shift its weight to stay upright. This used to be done mainly with mathematical models. Today manufacturers increasingly train movements in simulation using reinforcement learning: the robot practises millions of times in a virtual world, and the result is then transferred to the real machine.
Understanding and acting: vision-language-action models
The biggest change of recent years is in the "brain". So-called vision-language-action (VLA) models combine vision, language and movement in a single AI model. The robot sees a situation, receives an instruction in plain language and translates it into movement. Examples:
- NVIDIA Isaac GR00T N1, presented in March 2025 as an open foundation model for humanoids. It uses a "fast" system for motor control and a "slow" system for reasoning.
- Gemini Robotics from Google DeepMind, also from March 2025, built on Gemini 2.0 and used by Apptronik, among others.
- Figure AI's Helix, for which the Figure 03 sensor suite was designed.
These models are impressive in demonstrations, but not yet flawless in practice. Some manufacturers therefore use teleoperation: a human takes over remotely when the robot does not know a task. 1X openly states that for NEO a remote "Expert" can supervise and assist.
What works today, and what is in development?
Available today:
- Stable walking on flat and slightly uneven ground, and pre-trained movements such as dancing and running, as Chinese manufacturers showed publicly in 2025 and 2026 (IFR).
- Well-defined tasks in controlled environments, such as moving totes or handing over parts, often as pilots.
- Voice interaction through language models.
In development:
- Reliable grasping of unfamiliar, soft or fragile objects.
- Longer working periods without human intervention.
- Safe collaboration without fencing, including dedicated safety standards (see our insights on safety).
Future:
- Robots that learn new tasks in any household or business without preparation. The IFR does not expect mass adoption as universal household helpers in the near or medium term.
In closing
A humanoid robot is not a single invention but a stack of mechanics, sensors and AI, each still improving quickly. If you want to understand what a robot can really do, look at all three layers. More background is on our Technology page. Want to follow how KYNOR brings these developments to Europe? Sign up for Early Access.
Sources
- https://ifr.org/ifr-press-releases/news/humanoid-robots-vision-and-reality-paper-published-by-ifr
- https://ifr.org/ifr-press-releases/news/china-makes-ai-powered-robots-core-of-national-strategy
- https://euronews.com/2026/10/01/humanoid-robots-china-builds-them-but-germany-could-supply-the-joints-and-hands
- https://www.unitree.com/g1
- https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/
- https://www.figure.ai/news/introducing-figure-03
- https://www.1x.tech/neo
- https://research.nvidia.com/publication/2025-03_nvidia-isaac-gr00t-n1-open-foundation-model-humanoid-robots
- https://deepmind.google/discover/blog/gemini-robotics-brings-ai-into-the-physical-world/
- https://humanoid.guide/humanoid-actuator-supplier-landscape/
This article is based on public sources and reflects the situation on the date of publication. It is not advice on any specific purchase.