How Do Humanoid Robots Work?

The sense–think–act loop

Almost every robot, including humanoids, runs the same basic cycle thousands of times an hour: take in information about the world, decide what to do, and move. What makes modern humanoids capable is how good each of those three stages has become — especially the "think" stage, thanks to recent advances in AI.

1. Sense: how a humanoid perceives the world

A humanoid robot carries several kinds of sensor, each answering a different question:

• Cameras answer "what is around me?" — recognising objects, people and obstacles.
• Depth sensors answer "how far away is it?" — building a 3D map of the space.
• An inertial measurement unit (IMU) answers "which way is up, and am I tipping?" — essential for balance.
• Force and torque sensors answer "how hard am I pushing or gripping?" — so the robot can handle objects without crushing or dropping them.
• Joint encoders answer "exactly where is each limb right now?"

2. Think: the AI that turns perception into action

The "brain" is software running on processors inside the robot. The most important recent shift is the move to vision-language-action (VLA) models — AI that takes in camera images and a plain-language instruction and outputs the movements needed to carry it out. This is the engine behind what the industry calls physical AI: intelligence that acts in the real world rather than just on a screen.

In practice the "think" stage blends several techniques: classical control for balance and safety, planning to sequence steps of a task, and learned AI models for perception and dexterous manipulation.

3. Act: actuators, joints and hands

The decisions become movement through actuators — almost always electric motors at each joint. A humanoid has dozens of these "degrees of freedom": several per leg for walking and balance, several per arm, and many in the hands for grasping. Precise control of each motor, updated continuously, is what lets a humanoid stand on one foot, recover from a nudge, or place a part gently into a fixture.

How a humanoid learns a new task

Re-tasking a humanoid is mostly a software exercise, which is the whole point of a general-purpose machine. Common methods include:

1. Teleoperation and imitation — a human operator demonstrates the task (sometimes by wearing motion-capture gear) and the robot learns from those demonstrations.
2. Simulation training — the robot practises in a virtual "digital twin" of the workspace millions of times before touching real hardware.
3. On-site fine-tuning — the behaviour is adjusted to the specifics of a real line, with safety limits in place.

Power and runtime

Humanoids run on rechargeable batteries, typically delivering a few hours of active work before needing a charge or a battery swap. Planning around charging — like scheduling it across shift breaks — is part of deploying them well.