Humanoid robot anatomy: actuators, hands, sensors, compute
What is inside a humanoid robot: actuators and transmissions, hands, sensors, onboard compute and batteries — and how each limits capability and safety.
- A humanoid robot is built from five subsystems: actuators, hands, sensors, onboard compute and a battery, on a structure with dozens of joints.
- Actuators dominate cost and capability. The choice of motor and transmission trades off strength, speed, efficiency and how well the joint tolerates impacts.
- Hands are the least mature part. Dexterity, durability and cost are still in tension, and tactile sensing is far behind human touch.
- A humanoid balances actively, so cutting power makes it fall. That breaks the usual assumption in machine safety that removing power is the safe state.
Strip the shell from a humanoid robot and you find the same five subsystems in every design: actuators to move, hands to manipulate, sensors to perceive, compute to decide, and a battery to run it all. The differences between humanoids come from the trade-offs made in each — and those trade-offs set hard limits that no robot model can train its way around.
Actuators: the muscles
An actuator is a motor plus a transmission plus sensing, packaged as a joint. Humanoids have dozens, and they account for a large share of the robot’s cost, weight and failure modes.
Nearly all current humanoids are electric. The design question is the transmission, which converts a motor’s fast, weak rotation into slow, strong motion.
| Type | Strengths | Weaknesses |
|---|---|---|
| Harmonic (strain-wave) drive | Compact, high reduction, almost no backlash | Costly, sensitive to shock loads |
| Planetary gearbox | Robust, efficient, cheaper | Some backlash, bulkier at high reduction |
| Quasi-direct drive (large motor, low reduction) | Tolerates impacts, senses force through motor current | Lower torque for its weight, runs hot under sustained load |
| Linear actuator (motor driving a screw) | Very high force, suits knees and ankles | Complex, expensive, wear in the screw |
Two properties matter beyond raw torque. Backdrivability — how easily the joint can be pushed from outside — determines how well the robot absorbs impacts and how gently it can make contact. Torque density — torque per kilogram — determines whether the robot can carry a useful payload on top of its own weight.
Hands: the hardest part
A human hand has more than twenty degrees of freedom and thousands of touch receptors. Robot hands are nowhere near that, and it shows: many humanoid demonstrations involve tasks chosen to suit a simple gripper.
The main design approaches are:
- Simple grippers. Two or three fingers, robust and cheap, sufficient for a large share of industrial handling.
- Tendon-driven hands. Motors sit in the forearm and pull cables to move the fingers, as in human anatomy. This allows many joints in a slim hand, at the cost of complexity and cable wear.
- Direct-driven hands. Small motors inside the palm and fingers. Simpler to maintain but limited in strength and joint count.
Tactile sensing is the frontier. Without a sense of touch, a robot judges grip by vision and motor current alone — like working with numb fingers. Tactile sensors exist, but covering a hand in durable, affordable ones remains difficult.
Sensors: perception and proprioception
A humanoid senses the world and itself.
- Cameras in the head, and often the wrists, are the primary input for learned policies.
- Depth sensing — stereo cameras, structured light or lidar — supports navigation and obstacle avoidance.
- An inertial measurement unit in the torso reports orientation and acceleration, which is essential for balance.
- Joint encoders report the position of every joint.
- Force and torque sensing, either dedicated sensors at the wrists and ankles or estimates from motor current, tells the robot how hard it is pushing on the world.
Compute: the brain is on board
A robot cannot wait for a data center to respond before deciding how to keep its balance. Control loops run on board, in real time. Most humanoids carry two tiers of computing: a real-time controller handling balance and joint control at high rates, and one or more embedded GPU computers running perception and the learned policy.
That creates a ceiling. The size of the model a robot can run is bounded by the power and heat an embedded computer can sustain inside a body — one reason architectures that pair a slow, large model with a fast, small one have become popular.
Battery: the runtime budget
Humanoids carry a battery pack, usually in the torso. Walking, and simply standing upright under load, consume power continuously. Runtime on a charge is commonly a few hours and depends heavily on the task, which is why battery swapping and opportunistic charging feature in most deployment plans.
What the hardware means for safety
Three hardware facts shape any humanoid risk assessment.
It falls when power is cut. Conventional machinery is made safe by removing power. A humanoid balances actively, so an emergency stop that cuts power drops tens of kilograms of robot wherever it stands. A safe state has to be engineered — a controlled descent to a crouch or seat — and it has to work even when the main controller has failed.
It is heavy and strong. A robot built to lift useful loads has enough mass and actuator force to injure someone. Power and force limiting, as used with small collaborative arms, is much harder to apply.
It is mobile. Unlike a fenced arm, it brings its hazards to wherever people are.
These are the reasons early humanoid deployments keep people at a distance, and why a dedicated safety standard for actively balancing robots is being developed.
Frequently asked questions
How many degrees of freedom does a humanoid robot have?
It varies widely by design. The body — legs, waist, arms and neck — typically accounts for somewhere between twenty and thirty-odd joints. Dexterous hands can add a similar number again, so totals range from the low twenties to well over fifty.
What kind of actuators do humanoid robots use?
Almost all current humanoids are electric. They combine brushless motors with a transmission: harmonic or planetary gearboxes for rotary joints, and screw-based linear actuators where high force is needed, such as in the legs. Earlier humanoids used hydraulics, which are powerful but heavy, inefficient and prone to leaks.
Why are robot hands so difficult to build?
A hand needs many joints in a very small volume, strong enough to hold tools, delicate enough not to crush what it holds, and robust enough to survive constant contact with the world. Packing actuators, transmission and touch sensors into that space at a reasonable cost is an unsolved engineering problem.
Why is a humanoid robot harder to make safe than a robot arm?
A fixed arm becomes safe when its power is cut and its brakes engage. A humanoid stays upright only by continuously controlling its balance, so cutting power turns it into a heavy falling object. Its safe state has to be designed deliberately, for example a controlled crouch or sit.