Droid AI

Technology fundamentals

How do humanoid robots work

A complete guide to the hardware and software systems inside a modern humanoid robot. From actuation and sensing to the AI stack that turns movement data into intelligent behaviour.

Published 20 September 2026 by Droid AI

A humanoid robot is a machine designed to operate in environments built for humans, using a body plan that broadly mirrors our own. Two legs, two arms, a torso, a head. Understanding how these machines actually work requires looking at several distinct layers: the physical structure, the actuation systems, the sensor array, the control software, and the AI layer that ties everything together.

The physical structure and mechanical design

A humanoid robot's skeleton is typically built from lightweight aluminium alloys or carbon fibre composites. The design challenge is significant: human proportions are not optimally engineered for stability. A tall, narrow base with a heavy upper body creates constant balance problems that must be solved in software, in real time, at every moment the robot is moving.

The number of degrees of freedom, meaning the number of independent joints, varies by design. A high degree-of-freedom robot can move more naturally and handle a wider range of tasks. Most commercial humanoid robots have between 30 and 50 degrees of freedom across the whole body. The hands alone account for a significant proportion of this, since human-like dexterous manipulation requires independent finger movement.

Joint design is critical. Each joint must be precise, strong, and power-efficient, while also being safe around humans. Many current designs use series elastic actuators, which incorporate a spring element to absorb shocks and improve force sensing. Others use high-torque motors with strain wave gearboxes, a compact and backlash-free design well suited to precise position control.

Actuation: how the robot moves

The vast majority of commercial humanoid robots use electric servo motors for actuation. Each motor drives a joint, with the motor's rotation translated into the angular movement of that joint through a gearbox. The motor can be commanded to move to a specific position, hold a specific torque, or move at a specific velocity. Modern servo systems sense all three simultaneously and report back to the control system many times per second.

Battery technology is a significant constraint. Humanoid robots are power-intensive machines. A robot performing physical tasks may draw several kilowatts during peak exertion. Current lithium-ion battery packs give most humanoid robots two to four hours of operation under moderate workload. Swappable battery modules and charging docks are common solutions for sustained commercial deployment.

Sensing: how the robot perceives the world

A humanoid robot's perception system typically combines several types of sensor. Cameras, usually stereo cameras for depth perception, provide visual information about the environment. Depth sensors, often using structured light or time-of-flight technology, give precise three-dimensional maps of nearby objects and surfaces.

Inertial measurement units placed in the torso and limbs track orientation, acceleration, and rotation. This proprioceptive sensing, awareness of the body's own state, is essential for balance. Joint position and torque sensors tell the control system exactly where each limb is and what forces are acting on it. Force and tactile sensors in the hands allow the robot to feel when it has grasped an object and how much force it is applying.

More advanced platforms add lidar for environment mapping, microphones for speech input, and additional cameras in the hands for close manipulation tasks. The sensor data from all these sources must be fused together, in real time, to give the control system an accurate picture of both the robot's own state and the environment around it.

Control systems: from sensors to movement

The control system is the software layer that translates sensor data into motor commands. At the lowest level, individual joint controllers run at very high frequencies, making continuous small adjustments to keep each joint on its commanded trajectory. Above this, whole-body controllers coordinate across all joints to produce coherent movement, maintain balance, and manage contact with the environment.

Walking on two legs is a particularly demanding control problem. Each step requires a carefully managed transfer of weight from one foot to the other, while the swinging leg is precisely placed for the next support phase. Model predictive control, which anticipates future states and plans accordingly, is widely used for gait generation. Reinforcement learning has also produced excellent walking controllers, capable of handling uneven terrain that would defeat hand-coded approaches.

The AI layer: intelligence above control

Above the low-level control systems sits the AI layer that has transformed what humanoid robots can do in recent years. This layer combines several technologies.

Large language models handle natural language understanding. A robot instructed to "pick up the red box on the table and put it on the shelf" must parse that instruction, identify the relevant objects, plan the required actions, and execute them. LLMs provide the language understanding component. The translation from instruction to executable action plan requires integration with the perception and control systems.

Computer vision systems, trained on large datasets, classify objects, estimate their poses, and identify surfaces and obstacles in real time. Neural networks trained on human movement data allow robots to learn motion skills from demonstration rather than requiring explicit programming for every task. Reinforcement learning allows robots to improve their skills through practice, discovering solutions to manipulation and locomotion challenges that human programmers would struggle to specify directly.

The integration of these AI components into a coherent system that can operate safely around humans, in unstructured environments, in real time, is the central engineering challenge of the field. It is a challenge that is being actively solved, but it is not yet fully solved. The gap between a controlled demonstration and reliable deployment in an arbitrary real-world setting remains significant.

Common questions

Frequently asked questions

How do humanoid robots keep their balance?

Humanoid robots maintain balance using inertial measurement units (IMUs) that track orientation and acceleration in real time. Control algorithms, often based on model predictive control or reinforcement learning, calculate the adjustments needed to keep the centre of mass over the support polygon at each step. Modern systems can recover from significant pushes and uneven terrain.

What type of motors do humanoid robots use?

Most current humanoid robots use electric servo motors with integrated position, velocity, and torque sensing at each joint. Some designs use hydraulic actuation for strength-critical applications. Electric servo systems are preferred for their precision, programmability, and quieter operation. Battery capacity and joint power density remain active engineering challenges.

How long can a humanoid robot operate on a single charge?

Operating duration varies significantly by design and workload. Most current humanoid robots are designed for continuous operation of two to four hours on a full charge under moderate workload. This is an active area of development. Swappable battery systems and docking stations are common solutions for sustained deployment in commercial settings.

Related guides