Humanoid Robot Parts: The Complete Anatomy Guide
A humanoid robot has 9 core subsystems: frame, actuators, sensors, hands, compute, software, battery, cables, and safety systems. See what each does and what buyers need to know.
By PointL4 Editorial
Published
A humanoid robot consists of nine core subsystems, each with a distinct engineering role. Understanding what each subsystem does and what specifications matter helps buyers distinguish between platforms, compare costs, and ask the right questions before purchase. A useful analogy: the frame is the skeleton, actuators are muscles, sensors are sensory organs, the onboard computer is the brain, and the software stack is the nervous system.
Quick answer
- A humanoid robot's main parts are: structural frame, actuators and joints, sensor suite, robotic hands, onboard computer, software stack, battery system, wiring and connectors, and safety systems.
- Actuators (the joint motors) are the single largest hardware cost: a typical humanoid has 31 actuator modules at $200–$800 each, putting actuators at $6,000–$25,000 of a robot's bill of materials.
- The Unitree G1 ships with 3D LiDAR, depth camera, RGB camera, IMU, and force sensors; the G1 EDU adds dexterous hands with 15–20 additional actuators per hand.
The Structural Frame
The structural frame is the skeleton: it supports all other components and defines the robot's shape and rigidity. Materials differ by application: aluminum 7075 is used for load-bearing structural parts where strength-to-weight ratio is critical; titanium is used at high-stress joints where fatigue failure is a concern; carbon fiber composites cover limbs and external panels where low weight and vibration damping matter; engineering plastics form non-structural housings and covers.
The frame must be designed to route cable harnesses, mount actuator modules in precise alignment, accommodate battery packs as structural elements, and provide crash-protection for sensitive electronics. In a well-designed humanoid, the frame is not just structure; it is also thermal management, cable routing, and the mechanical interface for every other subsystem.
Actuators and Joints
Actuators are the muscles of a humanoid robot, converting electrical signals into joint motion. A full-size humanoid requires between 23 and 43 actuator modules for its primary joints; a complete dexterous hand adds 10–20 more. Each actuator module contains: a brushless DC motor, a harmonic drive gearbox (for torque multiplication and backlash elimination), a torque sensor (for force feedback), an encoder (for position feedback), and a controller board.
The harmonic drive is the critical precision component. It uses a flexible spline (flexspline) that deforms elastically to engage with a circular spline, eliminating the backlash that conventional gear trains produce. This makes harmonic drives the standard in precision robotics, including the Unitree G1, Figure 03, and Apptronik Apollo. The cost of harmonic drives at low production volumes ($200–$400 per unit) is a primary driver of humanoid robot pricing.
A humanoid robot's actuator system consists of 23–43 actuator modules in the main body, with each module integrating a brushless DC motor, harmonic drive gearbox, torque sensor, encoder, and controller board. The Unitree G1, the world's best-selling humanoid with over 20,000 units shipped, uses 23 actuators in its base configuration and up to 43 in its EDU Ultimate configuration with dexterous hands. Harmonic drive gearboxes are the standard precision transmission in commercial humanoid robots; they eliminate backlash by using a flexible spline that deforms to engage gears, producing smooth, repeatable motion. At current production volumes, actuator modules cost $200–$800 each, placing actuator hardware alone at $6,000–$25,000 of a typical humanoid robot's bill of materials.
Sensor Suite
The sensor suite is the robot's perception system. A production humanoid integrates multiple sensor types because no single sensor provides all the information required for safe, capable operation.
3D LiDAR creates dense point-cloud maps of the environment, enabling obstacle detection and navigation in unstructured spaces. The Unitree G1 ships with a 3D LiDAR sensor in its head.
Depth cameras provide dense 3D spatial information at shorter range, suitable for manipulation tasks where the robot needs precise distance information for grasp planning.
RGB cameras provide color image data for object recognition and visual AI tasks.
Inertial Measurement Unit (IMU) measures linear acceleration and angular velocity, giving the robot its sense of balance and orientation. The IMU is the equivalent of the inner ear: it tells the control system whether the robot is falling, tilting, or accelerating.
Force and torque sensors at joints and in the hands give the robot proprioception: the sense of how much force it is applying and where. Without force sensing, a robot cannot grasp soft objects without crushing them or detect when it has made contact with a surface.
Robotic Hands and End-Effectors
Robotic hands are the most mechanically complex subsystem per unit mass. A single dexterous robotic hand contains 15–20 finger joints, 10–20 miniaturized actuators, 20+ force sensors, and 15+ encoders. The challenge is fitting this into a form factor and weight envelope that does not unbalance the robot's arm or limit speed.
BrainCo produces dexterous hands used in Unitree's G1 Pro E configuration. Dex3-1 hands appear in the G1 Pro A configuration. These are interchangeable components; the Unitree G1 EDU line is explicitly designed to support different hand configurations so researchers can select the end-effector appropriate to their task.
Simpler grippers (2-finger or 3-finger) are cheaper, lighter, and more reliable for structured tasks but cannot perform the manipulation variety that a 5-finger dexterous hand enables.
Onboard Computer and AI Systems
The onboard computer processes sensor data, runs AI models, and sends commands to the actuator network. In the Unitree G1 EDU, this is an NVIDIA Jetson Orin module with 100 TOPS (Tera Operations Per Second) of AI compute. Enterprise platforms like Figure 03 and Agility Robotics Digit use more powerful custom compute to run end-to-end neural network control policies.
The onboard computer communicates with actuators over a real-time fieldbus, typically EtherCAT or CAN, which guarantees low-latency, deterministic control updates. Latency in actuator control directly affects stability during dynamic locomotion.
Software Stack
The software stack operates at three levels. The lowest level handles real-time motor control with microsecond response times, where torque commands are issued to each joint. The middle level handles motion planning: pre-programmed behavior libraries, inverse kinematics solvers, and safety override systems. The top level handles AI: computer vision, natural language understanding, task planning, and the foundation model policies that allow the robot to generalize across novel situations.
ROS 2 (Robot Operating System 2) is the standard middleware in research platforms. The Unitree G1 EDU runs ROS 2, making it compatible with the broad ecosystem of ROS-based research tools, simulation environments, and community-developed software packages.
Battery and Power System
The battery functions as a structural, thermal, electrical, and safety system simultaneously. Lithium-ion battery packs in humanoid robots must balance energy density (for runtime), weight (for center-of-gravity management), and thermal management (to prevent thermal runaway in a system that generates heat from dozens of motor controllers).
The Unitree G1 provides approximately 2 hours of operation per charge. The 1X NEO from 1X Technologies provides 4 hours. Battery replacement intervals and spare battery costs ($800 for the G1) are real operational considerations for lab deployments.
A humanoid robot's sensor suite combines multiple sensor types because no single sensor provides complete environmental awareness. The Unitree G1 ships with a 3D LiDAR sensor for obstacle mapping, a depth camera for close-range spatial measurement, an RGB camera for visual AI, an IMU for balance and orientation sensing, and force sensors in the limbs for proprioception. The Unitree G1 EDU adds dexterous hands with 15–20 force sensors per hand for fine manipulation feedback. Enterprise platforms like Figure 03 and Agility Digit run additional sensors for end-to-end neural network control. The software stack that processes this sensor data runs at three levels: real-time motor control (microsecond latency), motion planning (inverse kinematics and behavior libraries), and AI (foundation model policies for task generalization).
Wiring, Cables, and Connectors
A full-size humanoid contains hundreds of wires, multiple communication buses, and separate power networks for motors, compute, and sensors. Flexible harnesses allow limb articulation without cable fatigue. Shielded connectors prevent electromagnetic interference between motor drive signals and sensor data lines. Poor cable management is a common source of field failures in custom robotics builds; production platforms like the Unitree G1 and Digit use engineered harness routing to minimize this risk.
Browse actuator components, sensors, and robotics parts at the PointL4 parts and upgrades marketplace. See the full physical AI hardware catalog for complete platforms.
Frequently Asked Questions
What are the main parts of a humanoid robot? The nine main subsystems are: structural frame, actuators and joints, sensor suite, robotic hands, onboard computer, software stack, battery and power system, wiring and connectors, and safety systems. Actuators are the largest cost component; the sensor suite and compute are the primary AI capability drivers.
What material is a humanoid robot frame made of? Most production humanoids use a combination: aluminum 7075 for load-bearing structural elements, titanium at high-stress joints, carbon fiber for limbs and covers, and engineering plastics for non-structural housings. The Unitree G1 uses aluminum and polymer composite; heavier-payload platforms use more titanium at critical joints.
How many sensors does a humanoid robot have? It varies significantly by platform. The Unitree G1 ships with 3D LiDAR, depth camera, RGB camera, IMU, and joint force sensors. The G1 EDU with dexterous hands adds 15+ force sensors per hand and 15+ encoders per hand. Enterprise platforms run additional sensor modalities for autonomous operation.
What software does a humanoid robot run? Most research-grade humanoids run ROS 2 as the middleware layer. Above this, platforms run motion planning libraries, inverse kinematics solvers, and AI models for perception and task planning. The Unitree G1 EDU is ROS 2 compatible. Enterprise platforms like Figure 03 run proprietary end-to-end neural network control policies trained in simulation.