How humanoid robots work

Four systems behind one movement

This guide focuses on two-legged humanoids. During a reach, cameras help locate the object while joint and inertial measurements describe the robot's motion. Control software turns the desired movement into actuator commands. Forces at the feet still have to support the body as the arms move, so reaching and balance affect each other. The four guides below explain those parts in detail. [1] [2] [3]

Locomotion. Foot placement and joint effort change the forces that support and redirect the body. Read How a humanoid takes one step.

Manipulation. A reach needs a calibrated object pose, feasible contact points and enough friction to hold the load. Read How a robot plans and checks a grasp.

Perception. Images and range measurements become a scene estimate with positions, timestamps and uncertainty. Read From sensor measurements to a scene representation.

Decisions and control. Task selection, motion planning and joint feedback operate at different levels of the same physical action. Read From a bottle instruction to a checked result.

A body is only part of the system

A human-like body does not tell you who selects its next action. Researchers have used NASA’s Valkyrie with an operator choosing contact points while software checks whether the planned posture is physically feasible. A person can therefore guide the task while the robot handles parts of the motion calculation. Autonomy needs a task description and an account of human involvement. [4]

The teleoperation guide traces the operator’s input, motion mapping and local feedback loop. It explains which decisions can remain with a person.

Sensors inside and outside

NASA's 2015 Valkyrie hardware description identifies stereo and laser sensing in the head, cameras on the torso and two inertial measurement units in the pelvis. Those systems observe different things. Cameras and range sensing describe the surroundings; inertial measurements track body motion. Joint measurements complete the picture of the robot’s own pose. [1] [3]

Motors, joints and contact forces

Valkyrie is electrically actuated. Its upper arm uses series elastic rotary actuators, and each complete arm has seven joints. NASA also documents linear actuators in the wrists and ankles. A visible joint and its driving mechanism are therefore not always a simple one-motor-to-one-axis pair. [1]

Unitree's published G1 seven-joint-arm example sends target position, target velocity, feedback gains and feedforward torque in each motor command. These fields let feedback and an additional torque request contribute to the joint effort. The example documents a command interface for that arm configuration. It does not establish the joints or development access of every G1 edition. [3]

Why a reach can move the feet

When a biped lifts one foot, the remaining contacts must support its weight and control its motion. Friction limits the sideways force each foot can apply. Reaching or carrying an object changes the forces and moments the controller must handle, so arm movement cannot always be planned independently of the torso and feet. Foot placement is part of whole-body control. [2]

A documented warehouse task

GXO’s June 2024 agreement with Agility Robotics describes Digit taking totes from mobile robots and placing them on conveyors at a SPANX facility. Agility Arc coordinates the workflow. That is a specific commercial operation with defined objects and handoff points. It provides a more useful reference than a claim that the same machine can perform every job in a warehouse. [5]

Where comparisons break down

Compare the task as closely as the hardware. Record the object weight, grip, floor surface, walking distance, intervention count and completed cycles. A successful lift says little about recovery from a dropped tote. A short walk leaves charging and long-shift wear unanswered. These are separate questions that need separate measurements.

The profiles below document three Unitree variants. They are a starting point for reading published specifications, not a ranking of the whole market. NASA’s Valkyrie is included here as a documented engineering reference, and Digit as an example of a named customer task.

The EngineAI T800 analysis separates body joints, finger joints, computers and sensors across the published configurations.

For a platform with several generations, read the Atlas hardware and operating record. Product ratings and earlier laboratory results retain their original hardware context.