What Is a Humanoid Robot, and Why the Human Shape?
A humanoid robot is a machine built on the plan of a human body: a torso, a head that carries sensors, two arms that end in hands or grippers, and usually two legs. The shape is a design decision with real costs, and it only makes sense when you know what it buys. This guide covers what sits inside a humanoid, why companies keep choosing the human form, and where a wheeled base or a single-purpose machine is the better answer.
A working definition
Definitions vary between companies and researchers, so a practical one helps. Call a robot a humanoid when it has a human-like upper body with two manipulating arms, and it is meant to work in spaces and with objects designed for people. Most humanoids walk on two legs. Some builders mount a humanoid upper body on a wheeled base and still call it a humanoid, which is fair as long as you know the tradeoff they made.
A humanoid is different from a robot arm bolted to a table, which repeats one motion in a fixed cell, and from a warehouse cart robot, which moves shelves or totes on flat floors. The humanoid's pitch is generality: one body that can walk to a task, pick something up, and do a different task after lunch.
The four subsystems inside every humanoid
Strip the shell off any humanoid and you find the same four systems. Each one limits what the whole machine can do.
Actuators: the muscles
Every joint needs something to move it. Most current humanoids use electric motors paired with gearboxes or other transmissions, one actuator per degree of freedom (each independent way a joint can rotate or slide). Hydraulics were the older route for strong, fast motion. Boston Dynamics ran hydraulic Atlas robots for years, then announced the retirement of its hydraulic Atlas in April 2024 and moved to a fully electric design. Electric actuators are quieter, cleaner and easier to control precisely, which matters for a machine that may work near people.
Actuator count adds up quickly. A standard Unitree G1 has 23 degrees of freedom, and its EDU version reaches 43 with optional extra waist and wrist joints and a pair of three-fingered hands, according to Unitree's product page. Every one of those joints is a motor, a sensor and a possible point of failure.
Sensors: how the robot knows where it is
Humanoids carry two broad kinds of sensing. Proprioception tells the robot about its own body: joint encoders report angles, an inertial measurement unit reports tilt and acceleration, and force or torque sensors report load. Without these, balance is impossible. Exteroception tells the robot about the world: cameras, depth cameras and lidar build a picture of the room, the objects in it and the people nearby.
Touch is the newer frontier. Figure AI says the fingertip sensors on its Figure 03, announced October 9, 2025, can register forces as small as the weight of a paperclip. Fine touch matters because grasping a soft or slippery object by vision alone fails often.
Compute: fast loops and slow thinking
A humanoid runs several control loops at once. The lowest layer keeps joints on target and the body upright, and it has to react within milliseconds. Above that sits motion planning (where to put the feet, how to reach), and above that, perception and task logic (what is on the table, which item to grab). Designers often split this across more than one computer so a slow perception step never starves the balance controller.
Power: the runtime problem
Walking, balancing and holding arms out all draw power continuously, and batteries add weight that the legs must carry. Unitree lists about 2 hours of runtime for the G1. Builders attack the problem in different ways: Boston Dynamics says its production Atlas can walk to a station and swap its own batteries, and Figure says Figure 03 charges wirelessly by stepping onto a stand at 2 kW. When you compare robots, runtime under a real workload tells you more than any peak strength figure.
Why build it in the human shape
If legs and dozens of joints are so expensive, why not build something simpler? Three reasons come up again and again.
The world is already built for people
Stairs, door handles, shelf heights, cart handles, tool grips, light switches and the width of an aisle all assume a person of roughly human size and reach. A robot that matches those dimensions can use existing spaces without a rebuild. A factory or warehouse that would need new conveyors, ramps or fixtures for a custom machine can, in principle, drop a humanoid into a workstation a person used yesterday.
People can show it what to do
A robot shaped like a person is easier for a person to demonstrate tasks on. An operator's arm motion maps to a robot arm with similar joints, and a human hand pose maps to a five-finger robot hand far more directly than to a suction cup. Recordings of people doing work also become more useful when the robot's body resembles theirs. Much of the current interest in humanoids rests on this bet that human-shaped data will be the cheapest route to general skills.
One body, many tasks
A special-purpose machine is built around one job. A humanoid is built around the assumption that the job list will change. For sites with many small, varied tasks that no single automation project could justify, a general machine is the appeal, even if it does each task slower than a specialist would.
What the human shape costs
The same choices that make humanoids flexible make them hard to build and run.
- Balance. A two-legged robot has to control its balance actively all the time. A wheeled cart does not tip over when its computer freezes. A walking robot can.
- Complexity. Dozens of actuators mean more parts to wear out, more calibration and more ways to fail.
- Strength at reach. Arms held out from the body lose lifting capacity quickly, and a humanoid must also keep the load within its balance limits.
- Speed. A general machine is rarely as fast as a machine built for one motion. A dedicated palletizer will outpace a humanoid stacking boxes.
- Cost. More joints, more sensors and more compute cost more to make, and the software to coordinate them costs more to build.
Humanoid, wheeled or single-purpose
Most real deployment decisions come down to three options.
| Option | Strengths | Weaknesses | Good fit |
|---|---|---|---|
| Legged humanoid | Stairs, steps, tight human spaces, many tasks | Balance risk, complexity, runtime | Varied tasks in spaces you cannot rebuild |
| Wheeled humanoid upper body | Stable, simpler, longer runtime | Needs flat floors and ramps | Flat facilities with varied manipulation |
| Single-purpose machine | Fast, reliable, cheaper per task | Does one job only | High-volume, repetitive work |
If your floor is flat and your task list is short, the table usually points away from legs.
A few real examples, as of September 2026
These show how different builders have made the tradeoffs above. Each claim comes from the company's own materials.
- Agility Robotics Digit. Agility describes its 2024 agreement with logistics company GXO as the first formal commercial deployment of humanoid robots under a robots-as-a-service model. Digit's job there was specific: in Agility's words, "moving totes from cobots and placing them onto conveyors."
- Boston Dynamics Atlas. In January 2026, Boston Dynamics unveiled the production version of the electric Atlas and said production would start immediately at its Boston headquarters. It lists 56 degrees of freedom, lifting up to 110 pounds and a reach of 7.5 feet, and said all 2026 deployments were already committed, with fleets going to Hyundai's Robotics Metaplant Application Center and Google DeepMind.
- Figure 03. Figure covered its third robot in soft textiles rather than hard machined parts, with covers that can be washed and removed without tools. It also charges wirelessly, as described above, and Figure says its first-generation factory line can build up to 12,000 humanoids a year.
- Unitree G1. A smaller machine, about 4 feet 4 inches tall and about 77 pounds, with a depth camera and 3D lidar. Unitree lists the standard model from $13,500, with a separate EDU version that adds joints and hand options.
Match the body to the task before anything else
Before asking which humanoid to buy, write down the tasks. List the surfaces the robot must cross (stairs, ramps, thresholds), the heaviest object and how far from the body it must be held, how long a shift runs, and how often the task list changes. If there are no stairs and the tasks rarely change, a wheeled robot or fixed automation will likely do the job for less. If the space was built for people, cannot be rebuilt, and the work shifts week to week, the human shape starts to earn its cost.