Reports

Humanoid Robotics: The Technology Landscape

Report·Humanoid·Mid-2026
Humanoid Robotics:
The Technology Landscape

A read of the current commercial humanoid field along three axes: form factor, hardware, and intelligence architecture.

3 Insights
01
A humanoid is two independent choices: how it moves (locomotion) and how it manipulates (end effectors).
02
Hardware bottlenecks are shifting. Actuators dominate cost today, 40 to 50% of the build, and fall with scale, but compute and battery are becoming the binding constraints as AI moves to foundation models.
03
Intelligence is moving up the stack, from hand-engineered control with learned perception toward foundation-model-driven systems. That shift trades predictability for generalization, and rests entirely on perception.
The frame
Three Axes

A humanoid robot is the product of three design layers. Each can be evaluated independently. Together they define what the platform can do, who can deploy it, and where it breaks.

01
Form Factor
How the robot moves and how it manipulates. Two independent design choices.
02
Hardware
Actuators, sensors, compute, battery. Each layer is its own constraint.
03
Intelligence
The software stack that perceives, plans, and acts. Three architectural patterns.
01 / Section
Form Factor Taxonomy

The term humanoid robot covers more variation than is sometimes acknowledged. Two independent dimensions distinguish current commercial platforms.

Form Factor Taxonomy
01 / Locomotion and manipulation
Two Dimensions

How it moves (locomotion) and how it manipulates (end effectors).

Axis A · Locomotion
Full bipedal
Wheeled
Bimanual stationary
Wheeled bipedal
Axis B · End effectors
Multi-finger dexterous hands
Task-specific effectors
Hybrid configurations
Form Factor Taxonomy
01 / Four approaches to moving through space
Locomotion
A.01
Full bipedal

Approximates human morphology. Mature on controlled surfaces. Open problems: novel terrain, fall recovery, reactive walking near humans.

Apptronik·ApolloFigure AI·Figure 03Agility Robotics·DigitUnitree·G1
A.02
Wheeled

Human-form upper body on a wheeled base. Cannot climb stairs. Trades leg locomotion for stability, efficiency, and uptime.

SoftBank Robotics·PepperSanctuary AI·Phoenix 8Rainbow Robotics·RB-Y1
A.03
Bimanual stationary

Two arms, no locomotion. Mounted in place or carried on an AMR. Focus entirely on dexterous manipulation.

Tokyo Robotics·ToroboEngineered Arts·AmecaABB·YuMi
A.04
Wheeled bipedal

Both legs and wheels. Climbs stairs when needed, runs on wheels on flat ground. Emerging hybrid form factor.

Hexagon Robotics·AEON
Form Factor Taxonomy
01 / Three approaches to manipulation
End Effectors
B.01
Dexterous hands
11 to 22 DOF per hand

Five-digit, anthropomorphic. Grasps arbitrary objects, uses tools, performs fine motor tasks. Mechanically complex.

Figure AI·Figure 03Apptronik·ApolloXPENG·IRON1X·Neo
B.02
Task-specific

Simple grippers, suction, hinged effectors. Trades versatility for reliability, lower cost, predictable behavior. Tuned for known objects.

Agility Robotics·DigitBoston Dynamics·AtlasTokyo Robotics·Torobo
B.03
Hybrid

Task-specific effectors with optional dexterous attachments. A middle ground between purpose-built reliability and general-purpose dexterity.

Unitree·G1 + upgradeHexagon Robotics·AEON
02 / Section
Hardware Architectures

Current humanoids share broad architectural patterns but differ in component choices that meaningfully affect capability and cost.

Hardware Architectures
02 / Four layers, four binding constraints
4 Hardware Layers

Each hardware layer is its own engineering problem with its own limit.

02.A
Actuators
The muscles that execute every action. Each combines a motor, gearbox, encoder, and controller.
02.B
Sensors
RGBD, LiDAR, IMU, force-torque, tactile. Perception failures cascade into action failures.
02.C
Compute
Runs perception, planning, and control models in real time. The brain.
02.D
Battery
Shared across locomotion, sensing, compute, and actuation. Defines deployment economics.
Hardware Architectures
02 / The cost driver
Actuators

A full-size humanoid uses 28 to 44 actuators. Each is itself a small system. Their cost dominates the bill of materials, and their cost trajectory shapes humanoid pricing.

40 to 50% of build cost$16,000 to $40,000 per robot today$100 to $300 per actuator at scale
Anatomy · One actuator
ControllerCONTROL LOOPMotorBLDC, HIGH RPMGearboxTORQUE TRADEJointEND EFFECTEncoderPOSITION FEEDBACK
Hardware Architectures
02 / A generational shift: Orin to Thor
Compute

As humanoid AI moves from classical control with learned perception toward foundation models in the planning layer, the Protocol team’s read is that on-device compute is becoming the binding constraint on the next generation.

Jetson AGX OrinJetson AGX Thor
AI computeup to 275 TOPS2,070 FP4 TFLOPS
Power draw15 to 60 W40 to 130 W
AI compute vs Orinbaseline7.5x higher
Energy efficiency vs Orinbaseline3.5x better
Sensor fusion latencynot specifiedunder 10 ms
Designed forclassical control with learned perceptionmany concurrent foundation models
Hardware Architectures
02 / The current bottleneck, the next inflection
Battery

Current platforms cluster in a narrow capacity range. Hot-swap and inductive charging are the workarounds. Solid-state is the next inflection.

PlatformCapacityRuntimeCharging
Unitree H10.864 kWhunder 4 h staticstandard
Tesla Optimus Gen 22.3 kWh~2 h dynamicstandard
Apptronik Apolloper pack4 h per packhot-swappable
Inflection · Solid-state
over 500 Wh/kg74 GWh demand by 2035

Solid-state lithium-metal could theoretically double current energy density. A humanoid carrying half the weight, or running a full eight-hour shift on the same pack.

03 / Section
Intelligence Architecture

The software stack that decides what the robot does. Sits on top of the compute hardware. Three layers, three current patterns.

Intelligence Architecture
03 / Three layers, three patterns, data flows top to bottom
The Core Framework
Classical control
Foundation model
Teleoperated
Sensors
Cameras, LiDAR, IMU
Cameras, LiDAR, IMU
Cameras, LiDAR, IMU
Perception
Neural network
Foundation model (VLA)
Human operator (VR)
Planning
Hand-engineered
Control
Hand-engineered
Classical (MPC, ZMP)
Classical (MPC, ZMP)
Actuators
Motor commands
Motor commands
Motor commands
LearnedClassicalHuman
Protocol
Research toward a safer autonomous world
protocolz.org
Request full PDF
01 / 15