Humanoid Robots · Series
Episode III · The Engineering

Under the Hood: What Happens Inside the Shell

From the physics of balance to motors and reducers, artificial intelligence, and the battery. Lifting the smooth exterior one layer at a time.

Published 2026·06·20 · 16 min read · by Lucky Please Editorial
Prologue

Beneath the Smooth Exterior

As we saw at the close of Part 2, the outward appearance of today's humanoids is, regardless of camp, growing steadily more alike; the broad outline of two legs, five fingers, and a camera on the head has effectively converged into one, and so the real competition takes place beneath that smooth skin, at the level of components and algorithms. In this part we open the shell and look, in turn, at the five essentials that actually make a humanoid move.

The order runs from the outside of the body inward. We start from the physical question of why "walking" is so hard, pass through the motors and reducers that actually produce that motion, arrive at the domain of artificial intelligence — what to see and how to judge — and finally come to a halt before the battery, which keeps all of it running for hours yet draws the least attention.

Chapter I

The Physics of Walking

Part 1 mentioned Honda's ZMP in passing, but since the concept is the starting point of humanoid walking it is worth stating once more, precisely. Whether for a person or a robot, the area where the soles touch the ground is called the "support area," and as long as the zero moment point, or ZMP — the single point to which the forces and inertia acting on the whole body are reduced — stays within that support area, the machine will not fall.

Balance = keep the ZMP under the foot CoM STABLE ZMP inside the foot CoM TIPS OVER ZMP past the edge
The key condition for walking. When the ZMP, which reduces the whole body's forces to a single point, lies inside the foot's support area, the machine is stable (left); once it passes the edge, the machine tips over (right). Diagram by Lucky Please.

The difficulty is that this condition must be held not while standing still but "while walking." The instant one foot lifts, the support area narrows to the area of the remaining foot and the body tends at every moment to topple forward, so the controller must recompute the angles and speeds of dozens of joints hundreds to a thousand times a second, continually drawing the ZMP back within the foot. The reason early humanoids walked so slowly and gingerly, and the reason recent machines can run and even jump, both come down to how quickly and precisely this computation is carried out.

Chapter II

What Stands In for Muscle: Actuators

When the command to keep balance comes down, what turns it into actual movement is the actuator seated in each joint, the "artificial muscle." As Part 1's hydraulic Atlas showed, the hydraulic approach that produces great force through fluid pressure was once the symbol of dynamism, but because of its weaknesses — heavy, expensive, and difficult to control precisely — almost every humanoid of the 2020s has turned to the electric motor.

At its core is a high-performance electric motor commonly called a BLDC. It produces more torque for the same weight, allows fine control, and is more favorable for managing heat and efficiency; yet the torque a motor produces directly is nowhere near enough to support a person's weight while walking. A motor is by nature good at turning "fast and weak," so a device to convert that fast rotation into "slow and strong" is indispensable — and that is the protagonist of the next chapter, the reducer.

Chapter III

The Costliest Handful: The Reducer

The reducer, as its name says, slows the motor's fast rotation and in exchange produces proportionally greater torque, and it is the core that effectively determines the performance and precision of a humanoid's joints. The most widely used among them is the precision reducer called the harmonic drive, whose distinctive structure — a thin, flexible toothed cup pressed by an elliptical input shaft so that it meshes with a rigid outer ring at only two points — produces motion that is small and light yet nearly free of the looseness between teeth known as backlash.

Harmonic drive reducer Circular spline rigid · fixed Flex spline flexible · OUTPUT Wave generator elliptical · INPUT Reduction 50:1 – 160:1 · near-zero backlash
A conceptual cross-section of a harmonic drive reducer. An elliptical input shaft (the wave generator) presses a flexible toothed cup (the flex spline) so it meshes with a rigid outer ring (the circular spline) at two points, yielding a large reduction ratio that is small, light, and free of backlash. Diagram by Lucky Please.

This component matters not for its technical excellence alone. The harmonic drive is hard to make, so a handful of firms effectively hold an oligopoly over the market, and since a single humanoid contains dozens of such precision reducers, one per joint, the moment robots are mass-produced, who supplies this "costliest handful" becomes an enormous business in itself. It is precisely this point — where a question of technology turns directly into a question of money — that sits at the heart of the stock value chain we will take up in the next part.

Chapter IV

The Head That Sees and Judges: Sensors and AI

However sturdy the legs and however precise the joints, without a "head" to judge what to see and how to move, a robot remains a machine that merely repeats fixed motions. The decisive difference between Part 1's ASIMO and today's humanoids lies exactly here: when cameras, an inertial measurement unit (IMU), and sensors that feel force and torque continually gather information about the surroundings and the body's posture, artificial intelligence interprets it and decides the next motion, and that command runs back down to the motors, a cycle that repeats hundreds of times a second.

① SENSE Camera · IMU · Force / Torque sensors ② THINK (AI) Perception · Balance · Planning ③ ACT Motors + Reducers drive the joints ④ BODY Joints · limbs move, posture changes closed loop · up to ~1 kHz
A humanoid's control loop. The AI interprets the information gathered by the sensors and decides the motion, motors and reducers move the body, and the result returns to the sensors — a closed loop that repeats hundreds of times a second. Diagram by Lucky Please.

The greatest recent change lies in the very way this "head" is built. In the past a person had to program every motion by hand, but the center of gravity has now shifted toward robots learning their motions on their own through vast quantities of video and simulation. In particular, as new artificial-intelligence models that learn vision, language and action as a single bundle have appeared, it is gradually becoming real for a robot to understand the single phrase "pick up the cup" and carry it out in an environment it has never seen before — and this is exactly why the companies we met in Part 2 are each staking everything on their own artificial intelligence.

Epilogue

These Parts Are About to Become Money

The control that keeps balance, the motor that produces force, the reducer that refines that force, the artificial intelligence that handles judgment, and the battery that keeps it all running. A humanoid, as a single machine, is in the end a precise ensemble of these five layers, and if any one of them is weak the whole comes to a halt.

Yet if you look quietly at this list of parts, it is also a vast industrial map. The company that makes the motors, the company that holds an oligopoly on reducers, the company that sells the AI chips, and the company that supplies the batteries each stand guard over a chokepoint in this flow. In the next part we will turn this anatomical chart of the technology over as it is and follow the stock value chain, to see where the money actually flows in an age when humanoids are mass-produced.

The Humanoid Robots Series
Next → Part 4: The Stock Value Chain — Where the Humanoid Turns into Money (coming soon)
  1. Part 1 · History — From Clockwork Automata to ASIMO and HUBO
  2. Part 2 · The Giants of Today — Atlas, Optimus, Figure, Unitree
  3. Part 3 · The Engineering — Locomotion, Actuators, Reducers, AI, Batteries
  4. Part 4 · The Stock Value Chain
  5. Part 5 · United States
  6. Part 6 · Japan
  7. Part 7 · Korea
  8. Part 8 · China

References · Sources

  1. Standard robotics literature on bipedal-locomotion control and the ZMP (zero moment point).
  2. Technical materials on electric and hydraulic actuators and on the torque density and control characteristics of BLDC motors.
  3. Materials on the structure of harmonic drive (strain-wave gear) reducers and on the precision-reducer market.
  4. Public research on robot learning (imitation and reinforcement learning) and on vision-language-action (VLA) models.
  5. Reporting and corporate disclosures on humanoid power consumption and the limits of battery runtime.

Diagrams

  1. Three diagrams (balance, harmonic drive, control loop) — created by Lucky Please (schematic, for conceptual explanation).