What is a humanoid robot? Definition, joints, real limits

A humanoid robot is a machine built to the general proportions of a human body: a torso, two arms with hands or grippers, a sensor head, and usually two legs. Humanoid means human-shaped, and the form is a design decision rather than a capability claim. Robots are given human shape so they can work in rooms, doorways, and stairs, with tools built for people. What follows covers how they are specified, how they stay upright, and where the hardware stops.

What counts as a humanoid robot

No certifying body owns the term, but the working definition in industry is consistent enough to use. A humanoid has an upright torso, two arms capable of manipulation, and a sensor head. Legs are usual but not universal. Several capable manipulation platforms sold today put a humanoid upper body on a wheeled base, because wheels are more efficient and cheaper to keep running than legs.

The Astribot S1 is a clear example. It stands roughly 1.7 m tall with two 7-degree-of-freedom arms, a 4-degree-of-freedom torso, and a 3-degree-of-freedom base that is wheeled and omnidirectional rather than legged. It is a humanoid in the sense that matters for manipulation, and not a biped. Most confusion about humanoid capability comes from treating legs, arms, and autonomy as one package; they are three problems at three maturity levels.

Degrees of freedom, and why the count matters

A degree of freedom is one independently controlled axis of motion, which here means one motorised joint. Six are the minimum to place an end effector at any position and orientation in reach. A seventh makes the arm redundant: it can hold the hand still while moving the elbow, which is how a robot reaches around an obstacle or avoids a singularity.

Whole-body counts add up quickly. The LimX Dynamics Oli EDU carries 31 active degrees of freedom excluding end effectors: 7 per arm, 6 per leg, 3 at the waist, 2 in the neck. It stands 165 cm tall, weighs up to 55 kg with the battery, walks at up to about 5 km/h, and lifts 3 kg per arm.

A higher count is not automatically better. Every axis adds a motor, a gearbox, a driver, wiring, mass, heat, a failure mode, and a term in the control problem.

Actuators: what actually moves the joints

Almost every humanoid shipping today is electrically actuated. A brushless motor drives a reduction gear, typically harmonic, planetary, or cycloidal, with a driver closing a current loop around it. Hydraulics gave earlier research machines great power density along with pumps, hoses, heat, and leaks, which is why electric drives won indoors.

The design tension is between torque and transparency. High gear ratios multiply torque and let a small motor hold a heavy limb, but they make the joint hard to backdrive, so it cannot easily feel or yield to an outside force. Low ratios, often called quasi-direct-drive, stay backdrivable and impact-tolerant at the cost of continuous torque. Legs favour the second; arms mix both.

Published figures show the range. Oli's highest-torque joint motor is rated to 150 Nm, which is what a knee needs to catch 55 kg of falling robot. The Astribot S1 sits at the other extreme, reaching a peak end-effector speed of 10 m/s while holding plus or minus 0.1 mm repeatability.

Balance and control: staying upright is an active process

A standing humanoid is not stable. It is continuously falling and correcting, which is why a powered-down biped topples. Balance means keeping ground reaction forces consistent with the intended motion, usually formulated around the centre of mass, the support polygon under the feet, and a criterion such as the zero moment point or the capture point that says where the robot must step to avoid falling.

Modern stacks run in layers. A planner sets footstep locations and body trajectory. A model predictive controller solves, many times a second, for forces that keep the predicted motion feasible. A whole-body controller turns those forces into joint torques within joint and contact limits. Increasingly that layer is a policy learned in simulation.

All of it depends on fast, clean state estimation. An inertial measurement unit gives orientation and angular rate, joint encoders give configuration, and foot contact sensing reports what the ground is doing. The rates are high: the Deep Robotics Lite3 Basic quadruped runs motion control at up to 1 kHz with a 9-axis IMU, and Oli pairs a 6-axis IMU with an RK3588 motion processor.

Legged robots that are not humanoids

Bipeds without a torso

Plenty of two-legged robots are not humanoids. They exist to study locomotion, so everything above the hips is left off. The LimX Dynamics TRON1 EDU is 854 mm tall, weighs under 20 kg, carries up to 10 kg, runs more than two hours per charge, and swaps between point-foot, sole, and wheeled ends.

Both LimX platforms are covered in detail in our article on the Oli EDU humanoid and the TRON1 EDU biped.

Four legs solve a different problem

Quadrupeds are often mislabelled as humanoids, and the difference is structural. With four feet the support polygon is large and the robot is statically stable at rest, removing most of the balance problem a biped solves every millisecond. The Lite3 Basic has 12 degrees of freedom, three per leg, weighs 12 kg, carries 5 kg, reaches 2.5 m/s, and climbs slopes to 40 degrees.

That stability is why quadrupeds reached commercial inspection work years earlier. Our piece on trends in quadruped robotics covers where that market is heading.

Manipulation does not require a human body

The strongest argument against the humanoid form is that most manipulation tasks do not need it. A machine built for one job does it with fewer joints, less compute, and a fraction of the price.

The Roborock Saros Z70 Robot Vacuum with OmniGrip™ Mechanical Arm makes the point at consumer scale: a five-axis folding arm on a robot vacuum, lifting items up to 300 g out of the cleaning path. It is not a humanoid and does not need to be.

Five body plans cover most of what is currently sold.

The gap between a demo video and deployed capability

Demonstration footage is honest about what a machine did once and silent about the rest. Four omissions are routine, and knowing them makes the videos easier to read.

First, who was driving. Teleoperation is central to how these robots are built rather than a secret: a human in a tracking rig performs the task, the demonstrations are recorded, and a policy is trained to imitate them. Astribot documents that workflow, pairing an open Python SDK and ROS 2 Humble integration with imitation learning from VR teleoperation. A clip that looks autonomous may be live teleoperation, a replayed trajectory, or a policy that only works on this table in this lighting.

Second, the number of attempts. Success rates on unfamiliar objects and cluttered scenes remain well below rates on trained ones, and one clean take says nothing about the ninety-ninth.

Third, endurance. Oli runs about two hours per charge on a 9,500 mAh pack; the S1 manages four to six hours of continuous operation and takes two to three hours to recharge. A full shift means battery swaps, a charging routine, and a plan for the hours on the dock.

Fourth, safety and cost. These machines carry serious mass and torque, are not certified for unguarded work beside untrained people, and are priced as capital research equipment. Current deployments sit in labs, pilots, and controlled industrial cells rather than shops and homes.

None of that makes the progress unreal. Walking on rough ground, two-handed dexterity, and learned recovery behaviours were research problems a decade ago and are product features now. The fair reading is that the hardware has arrived ahead of the autonomy, and the remaining gap is mostly software and reliability rather than motors and gearboxes.

Frequently asked questions

What is a humanoid robot in one sentence?

It is a robot built to roughly human proportions, with a torso, two manipulating arms, a sensor head, and usually two legs, so it can work in spaces and with tools designed for people. The form is chosen for compatibility with human environments, not because it is the most efficient shape for any one task.

What does humanoid mean outside robotics?

Humanoid means having the general form of a human: an upright body, two arms, two legs, and a head. Anthropology and fiction both use it that way. In robotics the word carries no performance guarantee, so humanoid and capable are separate claims that need checking separately.

How many degrees of freedom does a humanoid robot need?

Six per arm is the minimum for placing a hand anywhere in reach at any orientation, and seven adds the redundancy needed to work around obstacles. Full machines land much higher once legs, waist, and neck are counted. Oli EDU has 31 active degrees of freedom before any hand or gripper is added.

Can humanoid robots do useful work today?

Within limits, yes. In structured settings they handle trained pick-and-place, machine tending, inspection routes, and teleoperated tasks a person specifies. What they do not yet do reliably is manage unfamiliar objects, recover from unexpected states, or run an unsupervised shift. Buyers today are labs, universities, and integrators building toward a specific application.

Where to look next

If you are evaluating a platform rather than reading for background, start from the task. Decide whether you need locomotion, manipulation, or both, then check runtime, payload per arm, and the SDK first. The research-grade platforms we stock sit in the humanoid robots collection.