Black Scarab logoBLACK SCARAB

Deep Dive · Humanoid Robotics

Humanoid Robot Components and Suppliers: The Complete Hardware Anatomy Guide

A medical style dissection of the modern humanoid robot, from skeleton, joints, actuators, hands, batteries, and cable harnesses to vision, GPUs, tactile sensing, thermal management, safety, suppliers, failure modes, and service economics.

By Rodolfo Garcia Calderoni, CFA|Published September 7, 2026|Insights index
Discuss an OpportunityBack to Insights
Exploded anatomy of a humanoid robot showing its skeleton, joints, hands, battery, compute, sensors, wiring, and protective shell.

Open a humanoid robot and the machine stops looking like one product. It becomes a compact ecosystem of structural frames, precision transmissions, electric motors, bearings, battery cells, power converters, computers, cameras, encoders, touch sensors, cables, seals, cooling hardware, and safety controllers. Every piece competes for the same limited mass, volume, power, thermal capacity, and service access.

The medical analogy is useful because a humanoid is organized like a body. The frame is its skeleton. Motors are muscles. Reducers and bearings form the joints. Tendons appear as cables, belts, and screws. Cameras are eyes. An inertial measurement unit behaves like an inner ear. Encoders and force sensors create proprioception. The cable harness is a nervous system. The battery and power bus provide metabolism and circulation. Cooling maintains homeostasis. Safety electronics provide reflexes that must remain dependable even when the main AI computer is confused.

This atlas decomposes those systems to the component level, shows how signals and forces pass through the body, identifies the companies that build each layer, and separates public supplier relationships from commercially suitable products. That final distinction is essential. Most humanoid makers do not publish complete bills of materials, and a supplier that makes an appropriate component is not automatically inside Tesla Optimus, Figure 03, Apollo, Atlas, Digit, or any other named robot.

The Humanoid Body in One Page

Skeleton

Robot Equivalent

Torso, pelvis, links, housings, fasteners

Primary Job

Carry loads and preserve joint geometry

Typical Failure

Fatigue, loosened interfaces, deformation, impact damage

Muscles

Robot Equivalent

Motors and integrated actuators

Primary Job

Convert electrical energy into motion

Typical Failure

Heat, winding damage, demagnetization, controller failure

Joints

Robot Equivalent

Reducers, bearings, shafts, brakes

Primary Job

Multiply torque and carry complex loads

Typical Failure

Backlash, wear, grease loss, bearing damage

Tendons

Robot Equivalent

Cables, belts, linkages, ball screws

Primary Job

Transmit force while controlling mass placement

Typical Failure

Stretch, fatigue, routing friction, loss of calibration

Brain

Robot Equivalent

GPU, CPU, memory, storage

Primary Job

Perception, planning, learned policy, coordination

Typical Failure

Thermal throttling, memory pressure, timing instability

Spinal cord

Robot Equivalent

Real time controller, fieldbus, joint drives

Primary Job

Fast deterministic balance and torque control

Typical Failure

Clock drift, packet loss, missed deadlines

Eyes and inner ear

Robot Equivalent

Cameras, depth, lidar, IMU

Primary Job

Scene geometry, motion, orientation, localization

Typical Failure

Glare, occlusion, contamination, vibration, drift

Touch and proprioception

Robot Equivalent

Encoders, force torque sensors, tactile arrays

Primary Job

Measure joint state, contact, slip, and load

Typical Failure

Calibration drift, elastomer wear, sensor disagreement

Heart and metabolism

Robot Equivalent

Battery, BMS, converters, charging

Primary Job

Store, protect, and distribute energy

Typical Failure

Cell imbalance, overheating, contactor or interconnect failure

Nerves

Robot Equivalent

Harnesses, connectors, flexible circuits

Primary Job

Move power and data through a moving body

Typical Failure

Copper fatigue, fretting, abrasion, electromagnetic noise

Skin

Robot Equivalent

Covers, foam, textiles, seals, optical windows

Primary Job

Protect people and internal components

Typical Failure

Tearing, contamination, seal drag, hidden damage

Reflexes

Robot Equivalent

Safety PLC, E stops, brakes, STO, contactors

Primary Job

Create an independent path to a safe state

Typical Failure

Unsafe common cause failure or nuisance stops

A commercial humanoid is a system of systems. No single specification predicts whether the complete body will survive productive work.

How to Read the Supplier Atlas

The humanoid supply chain is unusually opaque. Robot companies often say that a subsystem is designed internally while still purchasing cells, semiconductors, image sensors, connectors, bearings, magnets, passive components, and manufacturing services. In house design describes architectural control. It does not mean every physical part is fabricated by the robot maker.

Every company named in this guide belongs to one of four evidence classes. The labels prevent a common market research error: converting product suitability into an unsupported customer claim.

Evidence Classes Used Throughout This Guide

Confirmed integration

Meaning

The robot maker or component company identifies a specific component or supplier on a named platform.

What Can Be Said

The relationship and configuration can be described with its source and date.

Confirmed partnership

Meaning

Both parties disclose development, supply, evaluation, or adoption, but the exact production configuration may remain private.

What Can Be Said

The partnership can be stated, while installed content must remain qualified.

Commercial supplier

Meaning

The company offers a component explicitly designed or technically suitable for humanoid robotics.

What Can Be Said

It belongs in the supplier landscape, not in a named robot bill of materials.

OEM design, vendor private

Meaning

The robot maker claims control of the design but does not identify underlying vendors.

What Can Be Said

Describe the architecture as internal and keep merchant component suppliers undisclosed.

Black Scarab Weekly

Get the next Black Scarab deep dive

Keep reading now, then get every new deep dive and news report in Thursday's briefing.

Weekly reporting and deep dives. Unsubscribe anytime. See our privacy notice.

1. The Skeleton: Structure, Load Paths, and Packaging

The frame carries body weight, payload, ground impact, joint reaction loads, and the violence of a fall. It also fixes the coordinate system that every encoder and control model assumes. If a shoulder carrier or leg link bends under load, the controller sees an error that no amount of software can fully erase.

A strong frame is not necessarily a heavy frame. Added mass increases motor torque, battery drain, impact energy, and the strength required everywhere below it. Mass in a hand is costly to the wrist, elbow, and shoulder. Mass in the torso is costly to the hips, knees, ankles, and feet. Designers therefore use machined or cast aluminum around concentrated loads, steel for compact shafts and hardened interfaces, thin wall links or composites for limbs, and molded polymers or textiles for nonstructural covers.

The most efficient structures perform several jobs. A joint housing can carry load, locate bearings, spread motor heat, seal grease, and route cables. A battery enclosure can reinforce the torso. A limb shell can guide airflow and keep fingers away from pinch points. This multifunctional packaging saves mass, but it can make repair harder when one damaged part performs too many unrelated functions.

Figure says Figure 03 shifted from the predominantly machined construction of Figure 02 toward die casting, stamping, and injection molding. Its battery housing uses stamped steel, die cast aluminum, and structural adhesive as a load bearing torso member. Boston Dynamics emphasizes common Atlas actuator modules and simple links. 1X wraps a low inertia tendon driven body in a custom lattice polymer exterior. These are different answers to the same anatomy problem: stiffness, low mass, safe contact, manufacturability, and service access must coexist.

Humanoid robot skeleton with torso frame, pelvis, limb links, bearings, shafts, and structural interfaces separated for inspection.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

Structural Regions and Material Logic

Torso and pelvis

Typical Construction

Cast or machined aluminum, stamped steel, hybrid composite

Engineering Priority

Bending stiffness, battery integration, heat spreading

Supplier Landscape

Jabil and other manufacturing partners; BASF and Covestro material systems; OEM custom structures

Joint carriers

Typical Construction

Machined or cast aluminum with steel inserts

Engineering Priority

Bearing alignment, torque reaction, fatigue

Supplier Landscape

Precision casting and machining ecosystem; robot OEM design

Arm and leg links

Typical Construction

Thin wall metal, carbon composite, hybrid structures

Engineering Priority

Low distal inertia and impact resistance

Supplier Landscape

BASF and Covestro materials; specialist composite and metal fabricators

Covers and padding

Typical Construction

Polycarbonate blends, TPU, foam, lattice polymer, textile

Engineering Priority

Pinch protection, contact softness, access

Supplier Landscape

BASF, Covestro, 3M, Henkel, contract molding and soft goods firms

Fasteners and bonds

Typical Construction

High strength bolts, inserts, thread lockers, structural adhesive

Engineering Priority

Joint retention and repeatable service

Supplier Landscape

Bossard, Böllhoff, Henkel, 3M, and broad industrial supply base

BASF and Fourier have disclosed a material development relationship. Other companies in this table are category suppliers unless a direct platform relationship is separately stated.

2. The Joint Is an Organ, Not a Motor

A rotary joint is the densest mechanical organ in the body. It may contain a frameless motor, inverter, rotor encoder, reduction mechanism, output bearing, output encoder, torque sensor, brake, controller, seals, grease, temperature sensing, cooling fins, structural interfaces, and a hollow path for cables. The quality of this integration often determines the robot's weight, control quality, noise, heat, durability, and repair time.

The motor creates speed and torque. The reducer trades speed for higher output torque. The bearing carries radial load, axial load, and overturning moment. The encoders measure rotor and joint position. A torque sensor or calibrated compliant element estimates force. The drive electronics switch battery power through the motor phases. The housing aligns everything and becomes the main thermal path into the frame.

Peak torque attracts attention, but continuous torque pays for work. A knee can briefly produce a spectacular number and still overheat during repetitive lifting. Continuous output depends on winding temperature, inverter losses, gearbox friction, housing conduction, airflow, duty cycle, ambient temperature, and how often adjacent joints generate heat at the same time.

Exploded humanoid joint actuator with motor, controller, gears, bearings, encoders, torque sensing, brake, hollow bore, and housing.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

Exploded Joint Anatomy

Frameless motor

Function

Creates electromagnetic torque

Design Question

How much continuous torque per kilogram and per liter?

Representative Companies

maxon, Kollmorgen, TQ RoboDrive, Moog, Schaeffler

Servo drive

Function

Controls phase current and motor torque

Design Question

Can it close fast loops efficiently inside a hot joint?

Representative Companies

Synapticon, Elmo Motion Control, Copley, TI, Infineon, ST

Reducer

Function

Converts motor speed into joint torque

Design Question

What balance of backlash, shock life, efficiency, and backdrivability is required?

Representative Companies

Harmonic Drive, Nabtesco, Leaderdrive, Schaeffler, Nidec

Output bearing

Function

Carries combined body and payload loads

Design Question

Can one compact bearing handle radial, axial, and moment loading?

Representative Companies

Schaeffler, SKF, THK, NSK, NTN, Timken

Dual encoders

Function

Measure rotor and actual output position

Design Question

Can the controller detect reducer compliance, wear, and disagreement?

Representative Companies

RLS, HEIDENHAIN, Renishaw, ams OSRAM, Celera Motion

Torque sensing

Function

Measures joint load or flexure deflection

Design Question

Is direct sensing worth its mass, cost, and calibration burden?

Representative Companies

Bota Systems, ATI, Schaeffler, OEM internal designs

Brake

Function

Holds selected axes when power is removed

Design Question

Which joints must resist collapse without motor current?

Representative Companies

mayr, Miki Pulley, Kendrion, OEM internal selection

Housing and seals

Function

Aligns, protects, lubricates, and rejects heat

Design Question

Can the cartridge survive dust, impact, heat, and field replacement?

Representative Companies

Trelleborg, Parker, Freudenberg, OEM housings

3. Muscles: Motor and Actuator Families

Most humanoids use permanent magnet electric motors, but motor geometry and transmission choice create very different behavior. A high speed inner rotor motor paired with a large reduction can be compact and powerful. An outer rotor torque motor can deliver more torque at lower speed. A quasi direct drive uses a torque dense motor and a low reduction ratio to improve backdrivability. A linear actuator uses a screw and linkage to resemble a muscle pulling across a joint.

Frameless motors remove the separate motor case and install the rotor and stator directly into the joint housing. That saves volume and mass, but the robot designer inherits responsibility for alignment, bearing selection, contamination, electromagnetic compatibility, and heat removal. Integrated actuator suppliers package more of that risk into a tested module.

Public examples show the range of strategies. Unitree describes internal rotor permanent magnet joint motors with hollow routing and dual encoders. Apptronik describes proprietary linear and rotary force controlled actuators. 1X says it winds and manufactures its Revo2 motors internally for a tendon driven system. Boston Dynamics says Atlas standardizes much of its body around only two powerful actuator types. PAL describes custom linear actuators in Kangaroo. There is no universal humanoid muscle.

Actuation Architectures

High ratio strain wave joint

Strength

Compact, precise, low backlash, large reduction in one stage

Tradeoff

Flexspline fatigue, compliance, lower backdrivability

Likely Body Region

Shoulder, elbow, wrist, waist

Planetary joint

Strength

Efficient, familiar, durable, scalable

Tradeoff

Backlash control and multiple stage complexity

Likely Body Region

Hip, knee, shoulder, general joints

Cycloidal or planocentric joint

Strength

High stiffness, torque, and shock tolerance

Tradeoff

Geometry, mass, vibration, and precision manufacturing

Likely Body Region

Highly loaded hips and knees

Quasi direct drive

Strength

Low reflected inertia, force transparency, natural contact

Tradeoff

Larger motor and higher current demand

Likely Body Region

Arms, hands, ankles, dynamic legs

Linear screw actuator

Strength

High force density and muscle like packaging

Tradeoff

Screw cost, linkage geometry, side loading, service complexity

Likely Body Region

Knees, hips, torso

Tendon drive

Strength

Moves motors away from fingers and distal limbs

Tradeoff

Stretch, friction, routing wear, calibration

Likely Body Region

Hands, forearms, soft body architectures

Public Actuator Evidence

Schaeffler and Humanoid

Publicly Supported Claim

Development and supply agreement for strain wave actuators, primarily for upper body joints

Evidence Class

Confirmed partnership

maxon and Reachy 2

Publicly Supported Claim

Exact EC, ECX, MILE, ENX, and GPX components are documented in Orbita joints

Evidence Class

Confirmed integration

Unitree H1

Publicly Supported Claim

Unitree branded M107 joint with hollow shaft, dual encoders, and crossed roller output bearing

Evidence Class

OEM design, underlying vendors private

Apptronik Apollo

Publicly Supported Claim

Proprietary linear and rotary actuator architecture

Evidence Class

OEM design, underlying vendors private

1X NEO

Publicly Supported Claim

In house Revo2 motor winding, actuator production, and tendon architecture

Evidence Class

OEM design, underlying vendors private

Synapticon ACTILINK

Publicly Supported Claim

Integrated humanoid joint family with servo drive and planetary or shaft gearing

Evidence Class

Commercial supplier

Harmonic Drive

Publicly Supported Claim

Integrated actuator and strain wave component families for compact robotics

Evidence Class

Commercial supplier

4. Reducers, Bearings, Brakes, and Proprioception

The reducer transforms the motor's useful operating speed into the torque required at the limb. Strain wave gearing uses a wave generator, flexible spline, and circular spline. Planetary gearing divides load among several planets. Cycloidal systems distribute contact across lobes or pins. Linear drives convert rotation into translation through ball screws or planetary roller screws.

The bearing is the cartilage of the machine. A reducer can transmit torque, but the output bearing carries the robot. Crossed roller bearings are attractive because alternating rollers support radial, axial, and moment loads within one compact ring. Thin section bearings preserve hollow bores. Angular contact pairs control axial and radial loads. Needle bearings save radial space in linkages.

Two encoders answer different questions. The motor encoder measures the rotor before the gearbox and supports commutation and fast velocity control. The output encoder measures the actual joint after backlash, compliance, flex, and wear. Comparing both creates better accuracy and a valuable health signal. RLS and HEIDENHAIN market dual encoder architectures that package both measurements around a hollow joint.

A holding brake is usually spring applied and electrically released. When power disappears, selected gravity loaded joints can remain controlled instead of collapsing. Brakes add weight, drag, heat, cost, and another failure mode, so many designs use them selectively or rely on counterbalance, controlled descent, mechanical stops, and compliant collapse strategies.

Transmission and Bearing Supplier Map

Strain wave reducers

Representative Suppliers

Harmonic Drive, Schaeffler, Leaderdrive, Nabtesco

Public Robot Evidence

Schaeffler has a public supply relationship with Humanoid. Other named platform relationships remain private in cited materials.

Planetary reducers

Representative Suppliers

maxon, Schaeffler, Neugart, Wittenstein, Nidec

Public Robot Evidence

maxon GPX models are documented in Reachy 2.

Cycloidal reducers

Representative Suppliers

Nabtesco, Sumitomo Drive Technologies, Spinea

Public Robot Evidence

Commercially suitable; current target platform relationships are generally undisclosed.

Ball and roller screws

Representative Suppliers

Schaeffler, THK, Bosch Rexroth, Ewellix, HIWIN

Public Robot Evidence

Schaeffler publicly positions a complete linear motion portfolio for humanoids.

Crossed roller bearings

Representative Suppliers

THK, Schaeffler, SKF, NSK, IKO

Public Robot Evidence

Unitree confirms the bearing type in M107 but not the supplier.

Encoders

Representative Suppliers

RLS, HEIDENHAIN, Renishaw, ams OSRAM, Celera Motion

Public Robot Evidence

RLS documents integration on PAL REEM C; other current humanoid ties vary.

Holding brakes

Representative Suppliers

mayr, Miki Pulley, Kendrion

Public Robot Evidence

Suitable compact robotic products; named current humanoid customers not publicly confirmed here.

5. Hands: A Robot at the End of Each Arm

A dexterous hand is not an accessory. It is another dense robot containing palm structure, phalanges, tiny bearings, miniature motors, gear stages or screws, tendons, pulleys, springs, encoders, force sensing, tactile arrays, palm cameras, flexible circuits, local control, and protective skin. Every gram in the hand multiplies the effort required from the wrist, elbow, and shoulder.

A fully actuated hand offers independent control but creates extraordinary wiring, thermal, and reliability demands. An underactuated hand lets several joints conform mechanically around an object with fewer motors. A tendon driven hand moves motors into the palm or forearm to reduce finger mass, while accepting tendon stretch, creep, routing friction, and recalibration. For repetitive logistics, a two or three finger gripper may outperform a beautiful humanlike hand.

Touch closes the manipulation loop. Sparse pressure pads confirm contact. Three axis taxels estimate normal and shear force. Optical tactile sensors observe deformation of an elastomer to recover contact geometry, texture, and slip. More sensory channels can improve dexterity, but they also create more wiring, data, calibration, wear surfaces, and machine learning burden.

Figure describes internal fingertip sensors that detect forces near three grams and palm cameras that see around occlusions. 1X describes a 25 degree of freedom hand with quasi direct tendon drives. Sanctuary AI integrates proprietary tactile sensing into Phoenix. NVIDIA's H2 reference design pairs a Unitree chassis with Sharpa Wave tactile hands. These are direct architectural disclosures. They do not establish that independent hand suppliers appear in other undisclosed robots.

Humanoid robotic hands showing finger mechanisms, tendons, motors, tactile pads, palm camera, wrist sensing, and forearm electronics.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

Hand and Touch Companies

Shadow Robot

What It Provides

Tendon driven Dexterous Hand and DEX EE with many sensors and tactile options

Evidence Class

Commercial supplier

SCHUNK

What It Provides

SVH five finger hand and a dedicated humanoid hand development company

Evidence Class

Commercial supplier and active humanoid development

Tesollo

What It Provides

DG 5F dexterous five finger hand

Evidence Class

Commercial supplier

Wonik Robotics

What It Provides

Allegro Hand research platform

Evidence Class

Commercial supplier

qb robotics

What It Provides

SoftHand underactuated compliant hand

Evidence Class

Commercial supplier

Inspire Robots and Seed Robotics

What It Provides

Hands offered as PAL Kangaroo options

Evidence Class

Confirmed platform options

XELA Robotics

What It Provides

uSkin three axis tactile arrays for fingertips, phalanges, palms, and grippers

Evidence Class

Commercial supplier with hand level integrations

GelSight

What It Provides

Optical elastomer tactile sensors for high resolution contact geometry

Evidence Class

Commercial supplier

Figure and Sanctuary AI

What It Provides

Internally developed fingertip and haptic sensing

Evidence Class

OEM design, underlying vendors private

Hand specifications change quickly. Compare active degrees of freedom, joint count, force, speed, sensor coverage, replacement cost, and demonstrated cycle life rather than one headline number.

6. Feet and Balance: Where the Model Meets the Floor

A humanoid foot is simultaneously a load path, a sensor, an impact absorber, a friction surface, and sometimes a charging interface. The ankle must manage pitch and roll while the sole transmits ground reaction force into the body. The controller needs to know whether the heel, toe, or edge is touching, where the center of pressure sits, and whether the contact is beginning to slip.

A typical foot may contain a structural metal plate, ankle actuators, joint encoders, a six axis force torque sensor, pressure or contact sensors, compliant heel and toe elements, wiring, sealed connectors, replaceable tread, and shock isolation. Some designs add an active toe. Figure 03 adds inductive charging coils in its feet, so the same organ that controls balance also receives two kilowatts of charging power.

The sole is a consumable component with an outsized effect on control. Dust, oil, water, wear, temperature, and floor material change friction. A balance policy trained against a fresh sole on clean concrete can become unreliable when tread is polished or contaminated. Vibram's disclosed sole collaboration with Agility Robotics is a useful reminder that footwear science is part of humanoid engineering.

Exploded humanoid ankle and foot showing actuators, bearings, force torque sensing, pressure sensors, compliant tread, toe, wiring, and charging coil.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

What the Foot Must Know

Ground reaction force

Hardware

Ankle or foot force torque sensor

Control Use

Estimate body support and load transfer

Failure Consequence

Poor balance, damaging steps, uncertain contact

Center of pressure

Hardware

Force torque sensor or distributed pressure array

Control Use

Keep the projected mass inside the support region

Failure Consequence

Edge loading and falls

Contact state

Hardware

Pressure pads, switches, torque estimates

Control Use

Change walking phase and stabilize landing

Failure Consequence

Early or late gait transition

Slip

Hardware

Shear sensing, IMU, vision, kinematic mismatch

Control Use

Reduce tangential force and recover stance

Failure Consequence

Unexpected foot motion and collapse

Ankle angle and torque

Hardware

Dual encoders, current sensing, joint torque sensor

Control Use

Whole body state estimation

Failure Consequence

Accumulating pose error

Sole condition

Hardware

Inspection, force signature, maintenance record

Control Use

Adjust friction model and schedule replacement

Failure Consequence

Gradual reliability loss that software may misdiagnose

7. Energy Metabolism: Cells, Pack, BMS, and Charging

The battery has to support two different loads at once. Compute, sensors, and communications create a persistent background draw. Walking, catching balance, lifting, and recovering from contact create fast current peaks. Decelerating joints can return regenerative energy to the bus. Pack design must therefore consider peak current, voltage sag, regeneration acceptance, heat, and fault containment, not just average energy consumption.

A complete pack begins with cells connected in series for voltage and in parallel for capacity and current. Cell holders, busbars or fusible wire bonds, thermistors, and compression structures form modules. A battery management system measures cell voltage and temperature, estimates state of charge and health, balances cells, and commands protection. The high power path adds a service disconnect, primary fuse, precharge circuit, contactors, current sensing, and a distribution bus. Converters create lower voltage rails for compute, sensors, fans, safety, and logic.

Figure's F.03 battery is the clearest public pack anatomy from a leading humanoid maker. Figure describes a 2.3 kWh structural torso pack, custom BMS, active cooling, potting, thermal propagation protection, a flame arresting vent, and cell wire bonds tuned as fusible links. It claims five hours of peak performance and two kilowatt charging. The underlying cell maker remains private.

Unitree lists a 13 series, 9,000 mAh smart quick release battery for G1 and an 864 Wh pack for H1. PAL lists a 1,080 Wh TALOS pack with up to 100 A discharge. 1X lists an 842 Wh NEO pack. These numbers are not directly comparable without voltage, task, duty cycle, payload, software, and thermal context. Amp hours alone are especially misleading across different pack voltages. Use watt hours for energy and study the current profile separately.

Humanoid torso battery system with cylindrical cells, structural pack, BMS, contactors, fuses, busbars, converters, cooling, and charging hardware.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

Battery Autopsy

Cell chemistry and format

What It Does

Stores electrochemical energy

What Must Be Verified

Energy density, power, cycle life, temperature, propagation behavior, availability

Series and parallel topology

What It Does

Sets voltage, capacity, and current sharing

What Must Be Verified

Worst cell behavior, balancing, busbar resistance, pack voltage window

BMS

What It Does

Monitors and protects cells

What Must Be Verified

Measurement accuracy, open wire detection, fault coverage, state estimation, communications

Fuse and disconnect

What It Does

Interrupts severe faults and permits service isolation

What Must Be Verified

Fault current rating, access, arc control, replacement process

Precharge and contactors

What It Does

Connects the pack without destructive inrush

What Must Be Verified

Weld detection, redundant isolation, coil power, safe failure

Current sensor

What It Does

Measures load and regenerative flow

What Must Be Verified

Bandwidth, offset drift, isolation, bidirectional range

Thermal system

What It Does

Moves cell heat and limits propagation

What Must Be Verified

Cooling under peak load, vent path, sensor coverage, blocked airflow

Structural enclosure

What It Does

Contains cells and may carry torso load

What Must Be Verified

Impact, puncture, fall, ingress, service and transport tests

Charge system

What It Does

Restores energy by cable, dock, or inductive interface

What Must Be Verified

Power, cooling, alignment, unattended operation, fleet scheduling

Battery and Power Company Map

CATL

Capability

Cells and battery systems; CATL identifies Galbot S1 as battery powered by CATL

Evidence Class

Confirmed named platform

LG Energy Solution

Capability

Cylindrical and pouch cells, modules, and packs marketed for robots

Evidence Class

Commercial supplier; named humanoid customers private

Samsung SDI

Capability

Robot battery development with Hyundai and Kia Robotics Lab

Evidence Class

Confirmed development partnership

Molicel

Capability

High power cylindrical cells including P45B and P50B

Evidence Class

Commercial supplier; Molicel says P50B entered humanoid applications

Figure

Capability

Pack, BMS, enclosure, cooling, and manufacturing designed internally

Evidence Class

OEM design; cell vendor private

TI, Analog Devices, NXP, Infineon

Capability

Cell monitoring, protection, balancing, current sensing, and power control silicon

Evidence Class

Commercial supplier candidates

Sensata and TE

Capability

High current contactors and switching

Evidence Class

Commercial supplier candidates

Littelfuse and Eaton

Capability

Fusing and circuit protection

Evidence Class

Commercial supplier candidates

Vicor, Murata, TDK Lambda

Capability

High density DC conversion

Evidence Class

Commercial supplier candidates

A suitable BMS chip, cell, or contactor does not prove use in a named robot. The exact pack supply chain is private for most humanoids.

8. The Brain Is a Hierarchy, Not One GPU

A humanoid needs several layers of computation with different timing contracts. The main AI computer processes cameras, language, mapping, foundation models, and learned policies. A real time controller converts intentions into coordinated body targets. Distributed joint controllers close position, velocity, and torque loops. Motor control units switch phase current at far higher frequency. An independent safety controller monitors limits and can command a safe state.

The lower a control layer sits, the faster and more deterministic it must become. A semantic model may run around tens of updates per second. Figure says Helix 02 produces full body joint targets at 200 Hz, while its whole body controller runs at 1 kHz. PAL describes EtherCAT control at 2 kHz and as high as 5 kHz on TALOS. Motor current loops commonly operate in the tens of kilohertz. These examples illustrate a principle, not a universal timing specification.

The GPU should not directly switch every motor, and the AI computer should not be the only device capable of stopping the robot. Variable inference latency, model updates, operating system load, and thermal throttling are incompatible with the strictest reflex loops. Intelligence can propose motion. Deterministic controllers and safety hardware must enforce physical limits.

NVIDIA Jetson Thor currently defines the high end merchant compute category. NVIDIA lists as much as 2,070 sparse FP4 TFLOPS, 128 GB of memory, 273 GB per second of bandwidth, and configurable power from 40 to 130 W. NVIDIA has publicly connected Thor with Boston Dynamics, Agility, Figure, 1X, and its Unitree H2 based reference design at different stages of adoption. Qualcomm, AMD, Intel, and custom OEM silicon create other paths. TOPS and TFLOPS remain poor comparisons unless precision, sparsity, memory bandwidth, model shape, power, and latency are held constant.

Humanoid compute hierarchy with edge AI module, GPU board, real time controller, network switch, safety electronics, and distributed joint control boards.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

Control Rate Pyramid

Semantic perception and VLA reasoning

Illustrative Rate

Roughly 10 to 30 Hz

Work

Interpret scenes, language, objects, and task context

Failure Tolerance

Can tolerate some variable latency but not stale world state

Motion policy

Illustrative Rate

Roughly 100 to 200 Hz

Work

Generate whole body or limb targets

Failure Tolerance

Timing variation degrades smoothness and response

Whole body control

Illustrative Rate

Roughly 500 to 1,000 Hz or higher

Work

Balance, contact, constraints, coordinated torque

Failure Tolerance

Missed deadlines can destabilize the robot

Joint servo

Illustrative Rate

Kilohertz class

Work

Position, velocity, torque, and impedance loops

Failure Tolerance

Requires deterministic local execution

Motor current control

Illustrative Rate

Tens of kilohertz

Work

Field oriented control and phase current regulation

Failure Tolerance

A missed deadline can create poor torque or a hardware fault

Safety supervision

Illustrative Rate

Independent and deterministic

Work

Limits, E stops, STO, contactors, watchdogs

Failure Tolerance

Must remain available when AI compute fails

Rates are architectural ranges. Actual values differ by robot, task, motor, network, and controller design.

Compute and Control Suppliers

AI compute

Representative Companies

NVIDIA, Qualcomm, AMD, Intel

Public Evidence

Thor adoption or evaluation is publicly disclosed for several makers. Intel and NVIDIA options appear in Unitree and Fourier products.

Real time MCU

Representative Companies

TI, STMicroelectronics, Infineon, NXP, Renesas

Public Evidence

All publish robotics or humanoid design resources. Exact OEM chips are usually private.

Servo control

Representative Companies

Synapticon, Elmo Motion Control, Copley, maxon

Public Evidence

Integrated motion products are commercially available; platform relationships vary.

Industrial control

Representative Companies

Beckhoff, Siemens, B&R, Codesys ecosystem

Public Evidence

Useful for test, safety, and cell integration; onboard use is platform dependent.

Safety control

Representative Companies

Pilz, Beckhoff TwinSAFE, Siemens, SICK, Synapticon

Public Evidence

Agility confirms safety PLC and FSoE but does not identify the PLC vendor.

9. Two Nervous Systems and One Electrical Circulation

A useful humanoid architecture separates three traffic classes. High bandwidth perception carries camera, lidar, audio, and tactile streams toward the main computer. Deterministic control carries synchronized targets and encoder feedback among the whole body controller and joint drives. Safety traffic carries E stop state, limit checks, watchdogs, and certified stop commands. Power distribution runs alongside all three but must be electrically managed so motor switching noise does not corrupt delicate sensors.

Cameras may use MIPI, GMSL2, USB, or Gigabit Ethernet. Motion systems often use EtherCAT or CAN FD. EtherCAT distributed clocks support synchronized control across many axes, while Safety over EtherCAT can carry certified safety data through a black channel architecture. Wireless links are valuable for telemetry, updates, remote assistance, and fleet coordination, but they should not be the sole path for the fastest balance or emergency functions.

The main DC bus commonly feeds local inverters at each joint. Separate converters create rails for compute, sensors, networking, fans, and logic. Grounding, isolation, shield termination, common mode current, and return path design matter because each motor drive is a fast electrical switch sitting next to encoders, cameras, microphones, and touch sensors.

Transparent humanoid body showing power, motor, sensor, camera, control, and safety cable routes through hollow joints and flex zones.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

Data and Power Planes

Perception

Typical Interfaces

MIPI CSI, GMSL2, USB, Gigabit Ethernet

Priority

Bandwidth, synchronization, low latency

Should Not Depend On

A safety stop decision that must survive AI failure

Motion

Typical Interfaces

EtherCAT, CAN FD, deterministic Ethernet

Priority

Bounded latency, distributed clocks, fault detection

Should Not Depend On

Cloud availability or nondeterministic model inference

Safety

Typical Interfaces

Hardwired I/O, FSoE, safety fieldbus, redundant channels

Priority

Independent authority and diagnostic coverage

Should Not Depend On

The normal application process

Telemetry

Typical Interfaces

WiFi, private cellular, Ethernet

Priority

Fleet monitoring, logs, remote support

Should Not Depend On

Local balance and immediate hazard response

High power

Typical Interfaces

Battery bus and motor phase conductors

Priority

Low loss, safe interruption, insulation

Should Not Depend On

Sensitive signal grounds or unshielded sensor routing

Low voltage

Typical Interfaces

Converted rails for compute, sensing, logic, and fans

Priority

Stable power, sequencing, isolation

Should Not Depend On

Uncontrolled regenerative or motor noise

10. Cable Harnesses: The Quiet Reliability Bottleneck

A humanoid cable is asked to bend, twist, slide, and survive impact millions of times in a space too small for the ideal bend radius. Shoulders, elbows, wrists, hips, knees, ankles, and fingers are hostile zones. Copper work hardens. Insulation abrades. shields fracture. Connectors fret. Service loops snag. A single intermittent conductor can look like a software fault, an encoder failure, or a motor problem.

Good architecture reduces how many conductors cross each joint, routes through hollow shafts, controls the neutral bend axis, separates motor phases from encoder and camera lines, adds real strain relief, and places connectors at replaceable module boundaries. Flexible printed circuits help in thin regions. Hybrid connectors can combine power and data. Local electronics reduce wire count, but they add heat and can create another failure node.

Figure's BMW deployment makes this tangible. Figure says the Figure 02 forearm became its leading hardware failure area. A distribution board and dynamic cabling connected the main computer to wrist actuators. Figure 03 removed the board and dynamic wrist cabling so motor controllers communicate more directly with the main computer. That is an anatomy lesson with commercial consequences: every connector and flexing conductor has a failure probability.

Harness Design Checklist

How does the joint move?

Good Practice

Model bend, twist, sliding, and full service pose

Failure Being Prevented

Hidden overtravel and conductor fatigue

Where is the neutral axis?

Good Practice

Constrain the cable path and respect minimum bend radius

Failure Being Prevented

Kinking and unpredictable strain

What crosses the joint?

Good Practice

Reduce conductors through local networking or hollow routing

Failure Being Prevented

Bulk, snagging, and excess terminations

How is load removed from pins?

Good Practice

Use clamps, boots, overmold, and strain relief

Failure Being Prevented

Cracked solder joints and pulled contacts

How is noise controlled?

Good Practice

Separate phases, twist pairs, shield, and design returns

Failure Being Prevented

Encoder jitter, camera errors, tactile noise

How is a limb replaced?

Good Practice

Place keyed, touch safe connectors at module boundaries

Failure Being Prevented

Long repair time and incorrect reconnection

How is life validated?

Good Practice

Test representative motion, temperature, contamination, and current

Failure Being Prevented

Passing a simple flex test that misses real failures

Harness and Connector Suppliers

Molex

Relevant Capability

Compact board, signal, power, circular, flexible, and hybrid interconnect portfolio for humanoids

Evidence Class

Commercial supplier

TE Connectivity

Relevant Capability

Power, signal, data, hybrid connectors, relays, contactors, filters, and sensors positioned for humanoids

Evidence Class

Commercial supplier

Amphenol

Relevant Capability

High speed, power, circular, sealed, and rugged interconnects

Evidence Class

Commercial supplier

Hirose

Relevant Capability

Compact board, FPC, coaxial, and humanoid application interconnect map

Evidence Class

Commercial supplier

Samtec and LEMO

Relevant Capability

High density board and rugged circular connections

Evidence Class

Commercial suppliers

igus and LAPP

Relevant Capability

Continuous flex and torsion rated robotic cables

Evidence Class

Commercial suppliers

Figure

Relevant Capability

Public wrist harness simplification between F02 and F03

Evidence Class

Confirmed OEM architecture lesson; cable vendors private

11. Eyes, Depth, Lidar, and the Inner Ear

Humanoids need overlapping views rather than one pair of eyes. Head cameras support navigation and broad manipulation. Wrist or palm cameras look around the arm and into the final centimeters of a grasp. Depth sensing estimates geometry. Lidar provides robust ranging over a wide field. Fisheye cameras reduce blind zones. An inertial measurement unit measures angular velocity and linear acceleration when vision blurs, becomes occluded, or loses features.

Stereo vision triangulates depth from two synchronized views. Active stereo adds projected texture in difficult indoor scenes. Time of flight measures photon travel time. Lidar scans or distributes laser measurements around the robot. Each modality fails differently. Stereo can struggle with textureless, transparent, reflective, very close, or sunlit surfaces. Lidar provides sparse geometry and may struggle with reflectivity or close range. Cameras provide color and semantic detail but depend on exposure, cleanliness, timing, and calibration.

Unitree's public repositories identify a RealSense D435i head camera and Livox MID 360 lidar in G1 configurations, with a RealSense D405 available at the wrist. This is rare named sensor evidence. Figure says its current vision architecture doubles frame rate, cuts latency to one quarter, expands field of view by 60 percent per camera, and adds palm cameras, but it keeps sensor vendors private. Boston Dynamics lists 360 degree camera coverage on Atlas. Exact optical bills of materials remain private across most platforms.

Humanoid head and perception modules showing stereo cameras, depth optics, lidar, side cameras, wrist and palm cameras, IMU, microphones, and speaker.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

Perception Organ Comparison

Global shutter RGB

What It Adds

Texture, color, semantics, motion without rolling distortion

Primary Weakness

Lighting, exposure, occlusion, lens contamination

Representative Companies

Sony, onsemi, OmniVision, Basler, IDS

Stereo depth

What It Adds

Passive or active metric geometry from paired views

Primary Weakness

Texture, sunlight, reflective and transparent surfaces

Representative Companies

RealSense, Stereolabs, Luxonis, Orbbec

Time of flight

What It Adds

Compact direct depth measurement

Primary Weakness

Multipath, ambient light, range and resolution

Representative Companies

Analog Devices, Infineon, Orbbec, Sony

Lidar

What It Adds

Wide field geometry across varied illumination

Primary Weakness

Sparse semantics, reflectivity, close range, cost

Representative Companies

Ouster, Livox, Hesai, RoboSense

Event camera

What It Adds

Very low latency brightness change and high dynamic range

Primary Weakness

Different data representation and integration complexity

Representative Companies

Prophesee

IMU

What It Adds

Fast body rotation and acceleration

Primary Weakness

Bias, vibration, thermal drift, integration error

Representative Companies

Bosch Sensortec, TDK InvenSense, Analog Devices

Palm and wrist cameras

What It Adds

Close view of grasp and occluded work area

Primary Weakness

Cable flex, impacts, small baseline, contamination

Representative Companies

OEM internal modules; RealSense, Orbbec, Stereolabs candidates

12. Proprioception, Force, Touch, Hearing, and Voice

Proprioception lets the robot estimate its own body. Motor and output encoders provide joint position. Motor current estimates commanded torque. Physical torque sensors measure strain in a calibrated element. Wrist and ankle force torque sensors recover the full three dimensional force and moment at a contact. Temperature, voltage, and current reveal stress that geometry alone cannot see.

Motor current is a useful force estimate but contains friction, gearbox losses, temperature effects, cogging, and dynamic error. Direct torque sensing adds weight and calibration but improves compliant interaction. A series elastic actuator intentionally introduces a known deflection that acts as both spring and force measurement. No method is free: stiffness, bandwidth, shock survival, and accuracy trade against each other.

PAL TALOS is a useful research anatomy reference because it exposes joint torque sensing, six axis force torque sensors at wrists and ankles, high rate inertial sensing, encoders, temperature, and EtherCAT control. ATI and Bota Systems supply compact multiaxis force torque technology. XELA, GelSight, Touchlab, Shadow Robot, and other specialists address the far denser problem of touch across fingers and skin.

Audio is both an interface and a sensor. A microphone array estimates direction, supports beamforming, and helps the robot hear through factory noise, motor whine, fans, and its own footsteps. Acoustic echo cancellation is required if the robot speaks while listening. Unitree discloses a four microphone array and 5 W speaker on G1. Figure describes a larger speaker and repositioned microphone. The underlying acoustic component vendors remain private.

Sensing the Self and Contact

Rotor position

Hardware

Magnetic, inductive, or optical motor encoder

Best Use

Commutation and fast speed control

Caveat

Does not include reducer error

Joint output angle

Hardware

Absolute output encoder

Best Use

True link pose after the transmission

Caveat

Resolution does not equal installed accuracy

Joint torque

Hardware

Strain flexure, series elastic deflection, or observer

Best Use

Impedance and collision response

Caveat

Temperature, friction, and calibration matter

Wrist load

Hardware

Six axis force torque transducer

Best Use

Compliant manipulation and payload estimation

Caveat

Overload survival and tool offset calibration

Ground load

Hardware

Ankle or foot force torque sensor

Best Use

Balance and center of pressure

Caveat

Shock load and mounting stiffness

Finger contact

Hardware

Pressure pad or tactile taxel array

Best Use

Grip force and contact location

Caveat

Coverage gaps and drift

Slip and texture

Hardware

Three axis or optical tactile sensor

Best Use

Dexterous grasp correction

Caveat

Elastomer wear, compute, and contamination

Voice direction

Hardware

Synchronized MEMS microphone array

Best Use

Barge in, localization, interaction

Caveat

Self noise and echo cancellation

13. Thermal Homeostasis: Turning Peak Performance Into Shift Work

Every important humanoid organ creates heat. Copper losses warm motor windings. Inverters dissipate switching and conduction losses. Reducers and bearings create friction. The AI computer consumes continuous power. Battery resistance generates heat during hard discharge, regeneration, and charging. Brakes, converters, radios, and even dense tactile electronics add smaller thermal loads.

The usual path runs from winding or semiconductor, through insulation and thermal interface material, into a housing, frame, heat pipe, cold plate, or fin, then into ambient air. The most elegant design uses structural metal as a heat spreader. The most serviceable design still lets a technician replace the module without destroying its thermal interface or calibration.

Passive cooling is quiet and has fewer failure points, but it depends on surface area, temperature difference, and airflow around the body. Forced air removes more heat but adds fans, noise, filters, ingestion risk, and parasitic power. Liquid cooling can concentrate heat rejection but introduces pumps, seals, manifolds, and leak risk. Many humanoids mix methods: passively cooled joint housings, forced air compute, and an actively cooled battery.

Boston Dynamics says Atlas actuators use passive cooling and external fins, with a fan in the head. Figure integrates active cooling into the battery structure. Unitree highlights heat dissipation in its internal rotor joint design. The broader lesson is simple: peak torque, peak compute, and fast charging cannot be used continuously unless the complete body can reject the combined heat.

Humanoid thermal map with heat in motors, compute, battery, and joints plus cooling fins, heat pipes, cold plates, fans, and thermal interface materials.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

Thermal Organs and Their Limits

Motor winding

Cooling Path

Stator to housing to link and air

What Degrades First

Continuous torque and insulation life

Design Test

Worst joint pose at repeated duty cycle

Servo inverter

Cooling Path

Semiconductor to board and housing

What Degrades First

Current limit, switching efficiency, capacitor life

Design Test

Peak phase current at hot ambient

Reducer and bearing

Cooling Path

Lubricant and metal interfaces to housing

What Degrades First

Efficiency, grease life, backlash

Design Test

Long duration oscillation and shock

AI compute

Cooling Path

Heat spreader, heat pipe, fin stack, fan

What Degrades First

Clock rate and inference latency

Design Test

All sensors and largest deployed model at once

Battery discharge

Cooling Path

Cells to holders, gap filler, cold plate, enclosure

What Degrades First

Voltage sag, cell life, protection margin

Design Test

Walking, lifting, recovery, and regen profile

Fast charging

Cooling Path

Cells and charge electronics to active cooling

What Degrades First

Charge rate and fleet availability

Design Test

Hot pack after work entering unattended charge

External skin

Cooling Path

Internal heat through padding and covers

What Degrades First

Safe touch temperature and hidden hotspots

Design Test

Blocked vents, dirty filters, damaged foam

Thermal Supplier Landscape

Thermal interface materials

Representative Companies

Henkel, 3M, Parker Chomerics, Laird

Typical Contribution

Pads, gels, gap fillers, adhesives, electrically insulating heat paths

Cold plates and liquid hardware

Representative Companies

Boyd, Modine, Parker

Typical Contribution

Custom plates, manifolds, compact liquid cooling

Fans and blowers

Representative Companies

Delta Electronics, ebm papst, Sunon, Nidec

Typical Contribution

Compute and enclosure airflow

Heat pipes and vapor chambers

Representative Companies

Boyd, Wakefield, Asia Vital Components

Typical Contribution

Move heat from compact compute to usable surface area

Thermally conductive polymers

Representative Companies

Covestro, BASF, Celanese

Typical Contribution

Light enclosures and heat spreading structures

OEM integration

Representative Companies

Figure, Boston Dynamics, Unitree, 1X, Apptronik

Typical Contribution

Body level co design; exact vendors generally private

14. Skin, Seals, and the Independent Safety Reflex

The visible shell does more than create a face. Hard covers block debris and distribute impact. Foam lowers contact pressure. Textiles hide gaps while remaining washable and replaceable. Bellows and boots prevent pinch access. Seals retain lubricant and protect bearings and electronics. Optical windows protect lenses without destroying image quality. The final skin must survive repeated cleaning, abrasion, sweat, oils, dust, and collisions.

Softness can reduce injury severity, but it does not replace control safety. A padded robot can still trap a person, drop a payload, fall, or move a joint with high energy. Safe design combines limited inertia, rounded geometry, compliant control, speed and force limits, verified stopping, protective sensing, emergency stops, safe torque off, brakes, contactors, diagnostics, and operational procedures.

The safety reflex must be able to override the main AI system. A safety PLC or equivalent controller reads independent E stops, limit channels, protective devices, drive status, and watchdogs. It can request a controlled stop, remove motor torque through STO, apply brakes where designed, and open the high power path. Redundancy without diversity can still fail from a shared connector, power rail, software defect, or physical impact, so common cause analysis matters.

Agility publicly describes a safety PLC, on robot and pendant E stops, a Category 1 stop, Safety over EtherCAT, and functions designed to Performance Level d on Digit. Boston Dynamics lists Atlas as IP67 with field replaceable modules. Figure uses multidensity foam and washable textiles. 1X uses a soft lattice polymer exterior. These details indicate different safety philosophies, but certification must always be tied to the exact application, configuration, standard, and integration.

Humanoid protective layers from frame to foam and textile skin alongside an independent safety controller, E stop, contactor, brake, and human sensing zone.
Original Black Scarab illustration. Representative engineering architecture. Internal layouts vary by manufacturer and this is not a proprietary robot teardown.

The Safety Reflex Arc

Hazard sensing

Hardware or Logic

Vision, lidar, safety scanner, tactile, force limits

Purpose

Detect people, obstacles, contact, or abnormal load

Verification Question

What faults can create a missed detection?

Independent input

Hardware or Logic

Dual channel E stop, enable device, hard limits

Purpose

Give people and hardware direct authority

Verification Question

Can one broken wire mask the demand?

Safety decision

Hardware or Logic

Safety PLC, safe MCU, certified drive logic

Purpose

Evaluate limits and select stop behavior

Verification Question

Is it independent from the AI application?

Controlled stop

Hardware or Logic

Motion controller with monitored deceleration

Purpose

Reduce kinetic energy without losing balance

Verification Question

What happens during network or compute loss?

Torque removal

Hardware or Logic

STO at the drives

Purpose

Prevent active torque production

Verification Question

Is dangerous stored energy still present?

Mechanical retention

Hardware or Logic

Holding brake, counterbalance, stops

Purpose

Manage gravity loaded axes

Verification Question

Will the body or payload fall after STO?

Power isolation

Hardware or Logic

Redundant contactors, discharge path, weld detection

Purpose

Open the high energy bus

Verification Question

Can a welded contact be detected and contained?

Recovery

Hardware or Logic

Reset, diagnostics, supervised restart

Purpose

Prevent unexpected return to motion

Verification Question

Who confirms the workspace is safe?

Standards Context

ISO 10218 series

Where It Helps

Industrial robots and robot systems

Important Limit

A mobile dynamically balancing humanoid may introduce hazards beyond a traditional fixed arm

ISO TS 15066

Where It Helps

Collaborative industrial robot applications

Important Limit

Application limits and contact evaluation still depend on the complete system

ISO 13482

Where It Helps

Personal care robot safety

Important Limit

Scope differs from industrial and medical applications

IEC 61508

Where It Helps

Functional safety foundation

Important Limit

A capable component does not certify the final robot

ISO 13849

Where It Helps

Safety related control system performance

Important Limit

Performance Level requires architecture, diagnostics, validation, and application evidence

IEC 61800 5 2

Where It Helps

Safe drive functions such as STO

Important Limit

Drive safety does not resolve falling, payload, navigation, or grasp hazards

The applicable framework depends on where and how the robot is used. Buyers should demand a documented hazard analysis and certification scope for the exact deployed configuration.

15. Comparative Anatomy of Leading Humanoids

Humanoids that look similar can have radically different internal priorities. Some maximize industrial payload and ingress protection. Some minimize weight for home contact. Some use swappable packs for continuous operations. Some trade work readiness for low cost research access. Published specifications also use different definitions, so payload, runtime, degree of freedom count, and compute cannot be compared without reading the method.

The table below uses publicly disclosed 2026 information and preserves uncertainty. A blank or private specification is more honest than carrying an older generation number into a current product.

Representative Platform Anatomy

Figure 03

Body and Power

1.73 m, 61 kg, 20 kg payload, 2.3 kWh structural battery, stated 5 hour runtime

Sensing and Compute

Internal vision and tactile architecture; palm cameras; compute vendor private

Anatomical Signature

High internal design control, structural pack, wireless foot charging, manufacturing redesign

Apptronik Apollo

Body and Power

1.72 m, 72.6 kg, 25 kg lift and carry; swappable batteries

Sensing and Compute

Head and torso perception; chest computer; exact vendors private

Anatomical Signature

Linear and rotary force controlled actuators, modular power, Jabil manufacturing relationship

Agility Digit

Body and Power

About 1.75 m, 63.5 kg, 15.9 kg carry, 4 hour battery, 28 degrees of freedom

Sensing and Compute

Jetson in current generation path; Thor planned for later generation

Anatomical Signature

Logistics first body, interchangeable end effectors, autonomous docking, disclosed safety stack

Boston Dynamics Atlas

Body and Power

1.9 m, 90 kg, 56 degrees of freedom, 4 hour nominal runtime, IP67

Sensing and Compute

360 degree camera coverage, tactile hands, Jetson Thor integration

Anatomical Signature

High mobility, modular common actuators, passive joint cooling, automated battery swap

Unitree G1

Body and Power

1.32 m, about 35 kg, 23 to 43 joint motors, about 2 hours, 13 series battery

Sensing and Compute

Eight core CPU, depth camera, lidar, optional Orin class compute

Anatomical Signature

Low entry price, compact research body, dual encoder hollow joint routing

Unitree H1 2

Body and Power

1.78 m, 70 kg, 27 degrees of freedom, 864 Wh pack

Sensing and Compute

Intel control compute, up to three optional Orin NX modules, lidar and depth

Anatomical Signature

Speed and high torque leg architecture with growing manipulation capability

1X NEO

Body and Power

1.68 m, 29.9 kg, 842 Wh, stated 4 hour runtime, soft lattice body

Sensing and Compute

Jetson Thor product path and remote expert assistance

Anatomical Signature

Low inertia tendon drive, soft home oriented exterior, 25 degree of freedom hand

Fourier GR 2

Body and Power

1.75 m, 63 kg, up to 53 joints, 950 Wh, about 2 hours

Sensing and Compute

Intel motion computer, ARM perception computer, stereo camera, IMU

Anatomical Signature

Research orientation, internal FSA actuators, optional tactile hands

UBTECH Walker S2

Body and Power

1.76 m, about 70 kg excluding hands, 52 degrees of freedom, dual batteries

Sensing and Compute

X86 plus Jetson Orin, stereo, depth, fisheye, IMU, force torque sensing

Anatomical Signature

Autonomous three minute battery swap and continuous industrial operating thesis

PAL TALOS

Body and Power

1.75 m, 95 kg, 32 degrees of freedom, 1,080 Wh battery

Sensing and Compute

Joint torque, wrist and ankle force torque, high rate IMU and EtherCAT

Anatomical Signature

Research grade whole body torque control with unusually transparent sensing anatomy

Tesla Optimus Gen 3

Body and Power

Current production design specifications not publicly complete

Sensing and Compute

Tesla vision and planning architecture; current robot chip and sensor BOM private

Anatomical Signature

Maximum vertical integration thesis with incomplete public component detail

Payload definitions and runtime test conditions differ. Figures are not a ranking and should be verified against the current configuration before procurement.

16. Master Component and Supplier Atlas

The humanoid value chain is not one new industry. It joins automotive batteries and electronics, industrial motion control, machine vision, precision bearings and transmissions, mobile robotics, safety engineering, contract manufacturing, and AI compute. The table maps the major organs to companies that can credibly occupy each layer.

Only the evidence column establishes whether a supplier belongs to a named robot. Where the status says commercial supplier, the company has an appropriate portfolio but the target OEM relationship is not claimed.

Mechanical and Motion Supplier Atlas

Integrated actuators

Companies

Schaeffler, Synapticon, maxon, TQ RoboDrive, Kollmorgen, Moog

Evidence and Role

Schaeffler has disclosed Humanoid partnerships; maxon components are confirmed in Reachy 2; others are commercial suppliers

Frameless motors

Companies

maxon, Kollmorgen, TQ RoboDrive, Moog, Schaeffler

Evidence and Role

Commercial motor suppliers; current flagship OEM content usually private

Strain wave gearing

Companies

Harmonic Drive, Schaeffler, Leaderdrive, Nabtesco

Evidence and Role

Schaeffler and Humanoid is a confirmed partnership; others are commercial suppliers

Planetary gearing

Companies

maxon, Schaeffler, Neugart, Wittenstein, Nidec

Evidence and Role

maxon GPX is documented in Reachy 2; broader platform ties vary

Cycloidal gearing

Companies

Nabtesco, Sumitomo Drive Technologies, Spinea

Evidence and Role

Commercial supplier category

Linear screws and actuators

Companies

Schaeffler, THK, Bosch Rexroth, Ewellix, HIWIN, Rollvis

Evidence and Role

Commercial suppliers; OEM linear actuator details remain private

Bearings

Companies

Schaeffler, SKF, THK, NSK, NTN, Timken, IKO

Evidence and Role

Commercial suppliers; Unitree confirms crossed roller architecture but not its vendor

Encoders

Companies

RLS, HEIDENHAIN, Renishaw, ams OSRAM, Celera Motion, Netzer

Evidence and Role

RLS is documented on PAL REEM C; most modern OEM ties private

Brakes

Companies

mayr, Miki Pulley, Kendrion

Evidence and Role

Commercial compact safety brake suppliers

Hands

Companies

Shadow Robot, SCHUNK, Tesollo, Wonik, qb robotics, Inspire Robots, Seed Robotics

Evidence and Role

PAL confirms Inspire and Seed options; others are commercial platforms

Soles and elastomers

Companies

Vibram, BASF, Covestro, Trelleborg

Evidence and Role

Vibram and Agility is confirmed; BASF and Fourier is a development relationship

Seals and boots

Companies

Trelleborg, Freudenberg, Parker

Evidence and Role

Commercial supplier category; modern target OEM content private

Electrical, Compute, and Sensor Supplier Atlas

Battery cells

Companies

CATL, LG Energy Solution, Samsung SDI, Molicel, Panasonic Energy, EVE Energy

Evidence and Role

CATL and Galbot confirmed; Samsung and Hyundai Kia development confirmed; most OEM cell vendors private

BMS and power silicon

Companies

TI, Analog Devices, NXP, Infineon, Renesas, STMicroelectronics

Evidence and Role

All offer suitable components or humanoid reference material; exact OEM chips private

Contactors and protection

Companies

Sensata, TE, Littelfuse, Eaton

Evidence and Role

Commercial supplier category

DC conversion

Companies

Vicor, Murata, TDK Lambda, Delta Electronics

Evidence and Role

Commercial supplier category

AI compute

Companies

NVIDIA, Qualcomm, AMD, Intel

Evidence and Role

NVIDIA has several disclosed humanoid relationships; Intel and NVIDIA options appear in named platforms

Real time control

Companies

TI, Infineon, STMicroelectronics, NXP, Renesas

Evidence and Role

Commercial processors and reference designs; OEM selection private

Servo drives

Companies

Synapticon, Elmo Motion Control, Copley, maxon

Evidence and Role

Commercial motion control suppliers

RGB and depth cameras

Companies

RealSense, Stereolabs, Luxonis, Orbbec, Basler, IDS

Evidence and Role

RealSense modules are documented in Unitree configurations; others are commercial suppliers

Image sensors

Companies

Sony, onsemi, OmniVision, SmartSens

Evidence and Role

Merchant sensor suppliers; module and OEM relationships often private

Lidar

Companies

Livox, Ouster, Hesai, RoboSense

Evidence and Role

Livox MID 360 is documented in G1 software; Ouster has a confirmed FieldAI relationship

IMU

Companies

Bosch Sensortec, TDK InvenSense, Analog Devices

Evidence and Role

Commercial suppliers; exact base IMU vendors mostly private

Force torque

Companies

ATI Industrial Automation, Bota Systems

Evidence and Role

Commercial suppliers; PAL and IIT disclose sensing architectures but not always an outside vendor

Tactile

Companies

XELA, GelSight, Touchlab, Shadow Robot, Pressure Profile Systems, Tekscan

Evidence and Role

Commercial suppliers; Figure and Sanctuary use internal designs; Sharpa is confirmed on NVIDIA reference

Audio

Companies

Infineon, TDK, Knowles, Seeed Studio

Evidence and Role

Commercial microphones and arrays; named robot die vendors private

Harness and connectors

Companies

Molex, TE, Amphenol, Hirose, Samtec, LEMO, igus, LAPP

Evidence and Role

Commercial humanoid and robotics portfolios; OEM content private

Thermal

Companies

Henkel, 3M, Boyd, Parker, Delta, ebm papst, Sunon

Evidence and Role

Commercial material and cooling suppliers

Safety

Companies

Pilz, Beckhoff, Siemens, SICK, Synapticon

Evidence and Role

Agility confirms the architecture but not its safety PLC vendor

Confirmed Relationships Worth Remembering

Schaeffler

Robot Company or Platform

Humanoid

Confirmed Scope

Development and supply of strain wave actuators for humanoid joints

maxon

Robot Company or Platform

Pollen Robotics Reachy 2

Confirmed Scope

Named motors, encoders, and GPX gearboxes in Orbita joints

RLS and Renishaw

Robot Company or Platform

PAL Robotics REEM C

Confirmed Scope

Named magnetic encoder families in robot joints

Vibram

Robot Company or Platform

Agility Robotics Digit

Confirmed Scope

Purpose developed sole

Jabil

Robot Company or Platform

Apptronik Apollo

Confirmed Scope

Worldwide manufacturing, test, procurement, and supply chain support

BASF

Robot Company or Platform

Fourier

Confirmed Scope

Evaluation and development of polymers for robot components

CATL

Robot Company or Platform

Galbot S1

Confirmed Scope

Battery relationship identified by CATL

RealSense and Livox

Robot Company or Platform

Unitree G1 configurations

Confirmed Scope

D435i or D405 cameras and MID 360 mounts identified in official repositories

NVIDIA

Robot Company or Platform

Atlas, Digit, 1X, Figure, H2 reference

Confirmed Scope

Adoption, planned adoption, evaluation, or confirmed reference integration varies by platform

Sharpa

Robot Company or Platform

NVIDIA H2 reference design

Confirmed Scope

Dual Wave tactile five finger hands

Relationship scope can change by robot generation and configuration. A confirmed partnership does not establish every component inside the platform.

17. What Robot Makers Build and What They Buy

The highest differentiation sits at the interfaces. A company may buy an excellent motor, reducer, encoder, and drive, yet still fail if the joint is heavy, hot, noisy, difficult to assemble, or impossible to service. That is why leading OEMs tend to control actuator architecture, battery integration, hand design, structural packaging, calibration, controls, safety, and the learning system even when they purchase many underlying parts.

Merchant components remain essential. Battery cells require enormous electrochemical scale. GPUs and semiconductors require advanced fabrication. Precision bearings, connectors, seals, magnets, passive components, and image sensors benefit from specialized suppliers. Contract manufacturers can industrialize winding, electronics, tooling, test, and procurement without owning the robot architecture.

Figure explicitly says its actuators, batteries, sensors, electronics, and structures are designed internally. Apptronik owns its actuator approach and works with Jabil on manufacturing. Agility controls Digit and operates RoboFab. Boston Dynamics designs and builds Atlas while integrating merchant compute. 1X manufactures motors and actuators around its tendon and soft body thesis. Sanctuary describes a mixed internal and partner ecosystem. The likely mature industry is neither total vertical integration nor pure assembly. It is selective control of the organs that determine behavior, cost, and reliability.

Practical Make Versus Buy Logic

Actuator packaging and control

Why

Defines mass, torque, heat, feel, and service

Usually Buy or Co Develop

Motor materials and semiconductor devices

Why

Specialized scale and process technology

Hand and tactile integration

Why

Manipulation data and mechanics are strategic

Usually Buy or Co Develop

Selected hand modules during research

Why

Faster learning before internal volume design

Battery pack and BMS integration

Why

Safety, structure, current profile, and charging are body specific

Usually Buy or Co Develop

Electrochemical cells

Why

Cell manufacturing requires enormous scale

Safety architecture

Why

Must match body dynamics and deployment hazards

Usually Buy or Co Develop

Certified components and external validation

Why

Evidence and independent assurance matter

Whole body controls and calibration

Why

Creates the platform's motion quality

Usually Buy or Co Develop

Industrial buses, encoders, and MCUs

Why

Standards and supplier ecosystems reduce reinvention

Data, models, and fleet learning

Why

Compounds with every deployment

Usually Buy or Co Develop

Cloud and training infrastructure

Why

Capital efficiency and broad tooling

Structural and thermal co design

Why

Every gram and watt propagates through the body

Usually Buy or Co Develop

Materials, adhesives, fans, and thermal interfaces

Why

Specialist materials science and volume supply

End of line test

Why

Calibration quality defines shipped behavior

Usually Buy or Co Develop

Contract manufacturing execution

Why

Partners can scale fixtures, procurement, and labor

18. Where the Cost and Value Concentrate

Actuation is usually the largest mechanical value pool because a humanoid may contain dozens of high precision joints. Schaeffler estimates that actuators can represent roughly half of the bill of materials in many humanoids. Its illustrative analysis places significant value in rotary actuators, linear actuators, and hands. This is a supplier estimate, not a universal cost model.

Agility's public transaction materials placed a current Digit generation bill of materials near $125,000. Unitree advertises a far lower G1 entry price because it targets a smaller research platform with different payload, durability, safety, support, compute, and configuration. Prototype platforms, scaled consumer products, and certified industrial systems should never be placed on one cost curve without normalizing what the buyer receives.

Cost reduction will come from architecture as much as supplier negotiation. Fewer actuator sizes, common left and right parts, cast and molded structures, automated winding, shorter harnesses, reduced connector count, modular calibration, higher electronics yield, right sized compute, and field replaceable cartridges can each remove recurring cost. A cheap component that increases assembly time or field failures can raise total cost.

The Cost Down Levers That Matter

Common actuator families

Factory Effect

Fewer tools, fixtures, tests, and stocked parts

Field Effect

Simpler spares and training

Hidden Risk

Oversized joints can waste mass and power

Cast, stamped, and molded structure

Factory Effect

Lower unit cost after tooling

Field Effect

More consistent geometry

Hidden Risk

Tooling changes become expensive

Shorter harness and fewer connectors

Factory Effect

Less manual routing and test

Field Effect

Fewer intermittent failures

Hidden Risk

Overcentralization can make module replacement harder

Automated calibration

Factory Effect

Higher throughput and repeatability

Field Effect

Better replacement consistency

Hidden Risk

Poor reference fixtures can scale error

Right sized compute

Factory Effect

Lower module and cooling cost

Field Effect

Lower power and heat

Hidden Risk

Insufficient memory or model headroom

Replaceable cartridges

Factory Effect

Module level test before assembly

Field Effect

Shorter mean time to repair

Hidden Risk

Connector and sealing complexity

Supplier dual sourcing

Factory Effect

Resilience and price pressure

Field Effect

Part availability

Hidden Risk

Calibration and software variation across alternates

Illustrative Value Concentration by System

Actuation and joints50 % directional share

Directional supplier estimate. Includes motors, transmissions, bearings, sensing, and integration.

Perception and compute18 % directional share

Camera, lidar, inertial sensing, AI compute, networking, and memory can vary widely by platform.

Structure and skin14 % directional share

Materials and processes move from machined prototypes toward cast, stamped, molded, and textile parts.

Power system10 % directional share

Cells are only part of the pack; BMS, protection, cooling, structure, and charging add value.

Harness, safety, and controls8 % directional share

Often underestimated in cost models despite a large effect on assembly and uptime.

This visualization synthesizes broad industry ranges and Schaeffler's actuation estimate. It is not a disclosed bill of materials for any named robot and should not be used as a procurement quote.

19. Failure Anatomy: What Breaks Before the Demo Ends

A laboratory demonstration rewards peak capability. A commercial deployment rewards ordinary repetition. The failure that stops a fleet may be a wrist cable, worn sole, dirty lens, drifting tactile pad, hot inverter, loose fastener, or calibration record rather than the foundation model.

Failures also cascade. Bearing wear changes joint friction. The torque observer becomes biased. The controller compensates with more current. Motor and inverter temperature rise. Battery consumption increases. Motion quality falls. If health monitoring watches only individual components, it can miss the system level pattern.

Failure Mode Atlas

Frame

Failure Mode

Crack, loosened fastener, interface deformation

Early Signal

Changing alignment, vibration, torque offset

Design or Maintenance Response

Fatigue analysis, witness marks, torque audit, nondestructive inspection

Motor

Failure Mode

Winding insulation or magnet damage

Early Signal

Rising resistance, heat, torque loss

Design or Maintenance Response

Thermal model, current derating, winding test

Reducer

Failure Mode

Backlash, tooth or flexspline fatigue, grease loss

Early Signal

Dual encoder disagreement, noise, ripple

Design or Maintenance Response

Shock limits, lubrication plan, health trending, cartridge replacement

Bearing

Failure Mode

Pitting, brinelling, contamination

Early Signal

Vibration, temperature, friction

Design or Maintenance Response

Correct preload, sealing, impact control, replacement interval

Brake

Failure Mode

Failure to release or hold

Early Signal

Current change, drag, position creep

Design or Maintenance Response

Proof test, redundant strategy, safe collapse analysis

Encoder

Failure Mode

Drift, contamination, lost counts, mounting slip

Early Signal

Residual against second encoder and kinematics

Design or Maintenance Response

Absolute reference, plausibility checks, recalibration

Torque sensor

Failure Mode

Thermal zero shift or overload

Early Signal

Bias with unloaded joint

Design or Maintenance Response

Temperature compensation, overload stops, automatic zero validation

Tendon

Failure Mode

Stretch, creep, fray, pulley wear

Early Signal

Position hysteresis and rising motor travel

Design or Maintenance Response

Tension monitoring, replaceable routing, scheduled service

Tactile skin

Failure Mode

Puncture, delamination, drift, contamination

Early Signal

Dead taxels, baseline shift, inconsistent slip

Design or Maintenance Response

Replaceable pads, calibration, cleaning protocol, sensor redundancy

Foot sole

Failure Mode

Wear, hardening, contamination

Early Signal

Changed pressure signature and slip frequency

Design or Maintenance Response

Replaceable tread, friction inspection, policy adjustment

Battery cell

Failure Mode

Imbalance, resistance rise, internal fault

Early Signal

Voltage spread, heat, reduced usable energy

Design or Maintenance Response

Cell monitoring, propagation barriers, pack retirement criteria

Contactor

Failure Mode

Weld or failure to close

Early Signal

Auxiliary contact mismatch, abnormal bus voltage

Design or Maintenance Response

Precharge, weld detection, redundant isolation

Harness

Failure Mode

Copper fatigue, fretting, abrasion, loose retention

Early Signal

Intermittent faults tied to joint pose

Design or Maintenance Response

Motion representative flex test, strain relief, module connectors

Camera

Failure Mode

Dirty window, exposure failure, calibration movement

Early Signal

Reduced confidence, reprojection residual

Design or Maintenance Response

Cleaning, health image, redundant view, calibration check

Lidar

Failure Mode

Blocked aperture or timing fault

Early Signal

Coverage holes and range anomalies

Design or Maintenance Response

Window cleaning, plausibility fusion, safe degraded mode

IMU

Failure Mode

Bias, vibration coupling, clock offset

Early Signal

State estimator residual and drift

Design or Maintenance Response

Rigid mount, isolation, thermal calibration, synchronized time

Compute

Failure Mode

Thermal throttling, memory exhaustion, storage fault

Early Signal

Latency tail and missed deadlines

Design or Maintenance Response

Resource partitioning, watchdog, degraded model, cooling service

Network

Failure Mode

Clock drift, packet loss, connector fault

Early Signal

Jitter, cyclic error counters

Design or Maintenance Response

Distributed clock monitoring, redundant path where justified

Safety system

Failure Mode

Nuisance trip or dangerous missed demand

Early Signal

Diagnostic event or failed proof test

Design or Maintenance Response

Independent validation, proof testing, common cause analysis

Software update

Failure Mode

Timing, calibration, or behavior regression

Early Signal

Longer latency and changed fault distribution

Design or Maintenance Response

Versioned configuration, staged rollout, rollback, acceptance test

20. Manufacturing, Calibration, and Service Anatomy

A robot is not ready for scale when one expert can assemble it. It is ready when normal technicians can build, calibrate, test, repair, and return it to service with predictable tools and time. Manufacturing quality becomes control quality because encoder alignment, bearing preload, tendon tension, motor phase calibration, camera extrinsics, foot geometry, and torque zeros all enter the software model.

Module level testing is powerful. A joint cartridge can be tested for current, torque, friction, backlash, encoder agreement, brake operation, leakage, temperature, and vibration before it enters the body. A limb can store its own serial number and calibration. A replacement then becomes a controlled exchange rather than a mechanical rebuild inside the customer site.

Boston Dynamics emphasizes Atlas commonality and field replaceable modules. Figure redesigned around production processes and eliminated known forearm failure architecture. Jabil provides Apptronik manufacturing, test, procurement, and supply chain support. Agility operates RoboFab with stated capacity as high as 10,000 robots annually. These examples show that the humanoid race is becoming a manufacturing and service race.

End of Line Examination

Joint characterization

What Is Measured

Torque constant, friction, backlash, encoder offset, temperature

Why It Matters in the Field

Controller receives the real joint rather than an ideal model

Limb calibration

What Is Measured

Axis geometry, kinematics, force zero, cable behavior

Why It Matters in the Field

Replacement limbs preserve whole body accuracy

Vision calibration

What Is Measured

Intrinsics, extrinsics, synchronization, focus, optical cleanliness

Why It Matters in the Field

Depth, grasp, and navigation share one coordinate system

Battery and power test

What Is Measured

Capacity, impedance, protection, precharge, contactor, insulation

Why It Matters in the Field

High energy faults are found before shipment

Network timing

What Is Measured

Cycle time, jitter, clock agreement, error counters

Why It Matters in the Field

Balance and sensing remain deterministic

Safety validation

What Is Measured

E stop, limits, STO, brakes, contactors, restart

Why It Matters in the Field

The actual built unit reaches the intended safe state

Whole body acceptance

What Is Measured

Walking, contact, payload, thermal soak, recovery

Why It Matters in the Field

Interactions among otherwise good modules are exposed

Configuration record

What Is Measured

Hardware revision, firmware, calibration, model, service history

Why It Matters in the Field

Fleet failures can be traced and updates controlled

Service Metrics That Belong Beside Payload

Mean time between mission affecting failures

What It Reveals

How long the complete robot performs useful work

Mean time to diagnose

What It Reveals

Whether faults identify the real organ instead of generating ambiguous alarms

Mean time to repair

What It Reveals

Modularity, access, tools, connectors, calibration, and spare strategy

Intervention rate

What It Reveals

How frequently a person rescues the autonomy or hardware

Spare consumption per thousand hours

What It Reveals

True wear and reliability economics

Energy per useful task

What It Reveals

Body efficiency rather than idle runtime

Thermal derating time

What It Reveals

Difference between peak demonstration and sustained work

Calibration retention

What It Reveals

Whether transport, impact, and module replacement change behavior

21. A Buyer's Dissection Checklist

A humanoid should be purchased around a paid task, not around a body shape. Start with the work object, route, force, cycle, environment, exception rate, and safety boundary. Then trace every organ required to complete that task repeatedly.

A buyer does not need the confidential supplier name for every screw. The buyer does need enough architecture, test evidence, spares, telemetry, and support terms to understand failure consequences and operating cost. The strongest vendors can explain what the robot does when a camera is blocked, a tactile pad drifts, a joint overheats, a network link drops, a battery faults, or the main computer reboots.

Technical Due Diligence

Task definition

Questions to Ask

What exact objects, weights, poses, cycle times, routes, and exceptions are included?

Actuators

Questions to Ask

What are continuous torque, thermal duration, shock limits, backlash, life test, and replacement time?

Hands

Questions to Ask

What grasps are validated, what tactile coverage exists, and how often are fingers, tendons, or pads replaced?

Battery

Questions to Ask

What is usable energy, peak current, runtime at this task, charge time, cycle life, transport status, and fault containment?

Compute

Questions to Ask

Which models run onboard, at what latency and power, with what memory headroom and degraded mode?

Perception

Questions to Ask

What happens in glare, darkness, dust, reflective packaging, transparent objects, occlusion, and dirty optics?

Control

Questions to Ask

Which loops run locally, what are their rates, how is time synchronized, and what happens during packet loss?

Safety

Questions to Ask

Which standard and configuration are assessed, what safe states exist, and who owns site integration validation?

Reliability

Questions to Ask

What are fleet hours, task hours, intervention rate, component failures, and thermal derating data?

Service

Questions to Ask

Which modules are field replaceable, what tools and calibration are required, and where are spares stocked?

Cybersecurity

Questions to Ask

How are identity, signed software, remote access, logs, secrets, rollback, and vulnerability response managed?

Data rights

Questions to Ask

Who owns video, tactile, intervention, failure, and derived training data from the site?

A Disciplined Pilot Sequence

Bench anatomy

Proof Required Before Advancing

Component limits, hazard analysis, service access, and interface documentation are understood

Controlled workcell

Proof Required Before Advancing

One task completes safely with measured cycle time, quality, intervention, and heat

Representative shift

Proof Required Before Advancing

Battery, charging, thermal behavior, wear, and support survive the real schedule

Exception campaign

Proof Required Before Advancing

The robot handles or safely escalates common failures and edge cases

Maintenance trial

Proof Required Before Advancing

Site staff diagnose, exchange modules, recalibrate, and return the robot to work

Small fleet

Proof Required Before Advancing

Fleet software, spares, traffic, remote assistance, and correlated failures are measured

Scale decision

Proof Required Before Advancing

Cost per useful task and operational risk beat the best simpler alternative

22. The Next Anatomy

Humanoid design is converging on a small set of contested choices rather than one universal body. High ratio precision joints compete with lower ratio force transparent joints. Rotary cartridges compete with linear muscle architectures. Hard industrial shells compete with soft consumer bodies. Swappable batteries compete with autonomous opportunity charging. Merchant AI modules compete with custom silicon. Fully actuated hands compete with robust underactuated grippers.

The near term bottlenecks are physical. High force roller screws, precision reducers, torque dense motors, durable touch, flex tolerant harnesses, safe high power packs, fast calibration, and serviceable hands have to reach automotive style repetition without automotive style volumes. Supply chain depth will matter as much as model intelligence.

The strongest platform may not be the one with the highest peak torque, most degrees of freedom, or largest AI number. It may be the one whose bearings, cables, soles, seals, battery modules, and joint cartridges survive thousands of ordinary cycles with predictable maintenance. Intelligence makes the robot capable. Anatomy makes the capability available every day.

Black Scarab Takeaway

A humanoid is best understood as a moving network of electromechanical organs. The actuator is an integrated muscle and joint. The hand is a miniature robot. The foot is a changing contact model. The battery is a structural, thermal, electrical, and safety system. The GPU sits at the top of a hierarchy, not at the center of every reflex. The cable harness is a reliability system. The skin is both protection and service hardware.

The component opportunity is therefore much larger than the visible robot brands. Motors, reducers, bearings, screws, encoders, drives, cells, BMS silicon, connectors, cameras, lidar, IMUs, force sensors, tactile arrays, thermal materials, safety controls, contract manufacturing, calibration, and spare logistics all become part of the market.

The discipline is to follow evidence. A product page proves capability, not customer adoption. A partnership proves a relationship, not the exact production bill of materials. An in house claim proves architectural control, not self sufficiency. That careful anatomy is what lets investors, operators, suppliers, and buyers see the humanoid industry as it actually is.

Sourcing and Verification

This guide was compiled from official robot product pages, engineering updates, technical documentation, supplier product materials, public manufacturing announcements, safety references, and disclosed commercial relationships available through September 7, 2026. Platform specifications, component options, supplier relationships, certification scope, pricing, and availability can change by generation and configuration. Verify current details directly with the robot maker and component supplier before engineering, investment, or procurement decisions.

All anatomy images are original Black Scarab editorial illustrations. They synthesize common engineering architectures to make the systems understandable. They do not reveal or claim the confidential internal layout of any named commercial robot.

Black Scarab Weekly

Catch up on physical AI in one email

Every new Black Scarab deep dive and news report, collected into one clear Thursday briefing.

Weekly reporting and deep dives. Unsubscribe anytime. See our privacy notice.

Related Insights

Physical AI Investing

Top 15 Physical AI Venture Capital Firms and Funds in 2026

A founder and investor guide to the venture firms funding robotics, embodied AI, autonomy, industrial automation, edge intelligence, and the systems that connect AI to the physical world.

Read related insight

Robotics Infrastructure

Antioch Deep Dive: Can Simulation Become the Test Layer for Physical AI?

Inside Antioch's simulation platform, hybrid fidelity model, cloud evaluation workflow, NVIDIA and Nebius integrations, Amazon and Ring evidence, economics, alternatives, and the questions buyers should ask before trusting a digital twin.

Read related insight

Transportation & Logistics

Plus One Robotics Deep Dive: The Economics of Supervised Warehouse Autonomy

Inside Plus One Robotics and its supervised autonomy model: how PickOne sees parcels, how Yonder resolves exceptions, what InductOne and DepalOne sell, what the evidence establishes, and how buyers should calculate value.

Read related insight

Next Step

Explore a commercial opportunity in Mexico

If you are exploring how a physical AI technology could fit the Mexican market, Black Scarab can help assess the opportunity, map relevant stakeholders, and define a credible commercial next step.

Discuss an OpportunityBack to Insights