Skip to content
LumenaryAcademy
Frontier 80 min

Humanoid robots

Why build robots in our shape; what is inside them — actuators, hands, batteries — and what their energy budget allows; how they are controlled; who builds them, where they work in 2026, and what is still missing.

Code tested · 5AI-draftedAwaiting review by a roboticist

Your guides: Chloe · The Builder, Margaret · The Experimenter, Gideon · The Safety Engineer

MotivationH

Why the human form?

In 1973 Ichiro Kato's group at Waseda University in Tokyo switched on WABOT-1, a full-scale robot that walked a few slow steps, gripped objects and spoke a little Japanese. For the next forty years humanoids were research platforms: Honda's P2 (1996) and ASIMO (2000), Japan's national HRP series, KAIST's HUBO in Korea, IIT's iCub in Italy, NASA's Valkyrie. At the 2015 DARPA Robotics Challenge, won by KAIST's DRC-HUBO, many of the leading robots were humanoids, and the most memorable footage was of them falling over. In 2026 there are, by analysts' estimates, tens of thousands of humanoids being built a year, a few working paid shifts in car plants and warehouses, and investment measured in billions. What changed, and is the shape actually a good idea?

The arguments for the human form are practical. Our world is built for human bodies: stairs, door handles, shelves at shoulder height, tools and workstations sized for hands. A robot that fits it can, in principle, be dropped into existing workplaces without rebuilding them. Second, data: people can teleoperate a humanoid naturally by moving their own bodies, and the vast supply of video of people doing things is easier to learn from when the robot's body resembles theirs. Third, generality: one platform for many tasks spreads its development cost over many uses.

The arguments against are just as real. Wheels are far more efficient than legs on flat floors. A specialised machine — a palletiser, a conveyor, a robot arm in a cell — beats a generalist at any single task on speed, reliability and price. Legs add a way to fail that wheels do not have: falling. And human-like hands are the hardest part to build. Several companies hedge, putting a humanoid upper body on a wheeled base. This lesson looks at what is inside these machines, computes what their batteries allow, and surveys honestly, with dates, who builds them and what they do.

DefinitionsH

Bodies, joints and actuators

A typical full-size humanoid in 2026 stands 1.3–1.8 m tall, weighs 35–90 kg and has around 25–30 actuated joints in its body, plus 6–22 in each hand.

neck: 2shoulder: 3 eachelbow: 1 eachwrist: 2–3 eachwaist: 1–3hip: 3 eachknee: 1 eachankle: 2 eachbody: about 25–30hands: 6–22 each
Where a typical humanoid's degrees of freedom are. Counts vary by design; many robots have fewer at the wrist, waist or neck, and hands add 6 to 22 each.

Each joint is driven by an actuator: today almost always an electric motor, its drive electronics, a gearbox and sensors in one unit. Boston Dynamics' hydraulic Atlas (2013–2024) showed what high power density allows, but leaks, noise and cost pushed the field to electric drives; the electric Atlas replaced it in April 2024. Four properties matter for a humanoid:

Definition— Actuator properties

  • Torque density: peak and continuous torque per kilogram of actuator. Every kilogram in a leg must itself be carried and accelerated.
  • Gear ratio NN: the gearbox multiplies the motor's torque by NN (less friction losses) and divides its speed by NN.
  • Backdrivability: whether pushing on the output turns the motor. A backdrivable joint can feel and comply with forces through the motor's current; a highly geared one hides them behind friction.
  • Reflected inertia: the rotor's moment of inertia JJ appears at the joint as N2JN^2 J. In a collision or a foot strike, that inertia must be stopped by whatever the joint hits.

The common designs trade these against each other. Quasi-direct-drive joints — a large-diameter motor with a low-ratio planetary gear (around 6–10:1), a lineage running from MIT's Cheetah robots to Unitree's machines — are backdrivable and survive impacts, which suits legs. Strain-wave (harmonic) gears at 50–160:1 are compact and precise, common in arms and wrists, but are stiff, not easily backdriven and vulnerable to shock. Screw-driven linear actuators, with ball or planetary roller screws, push like a muscle across a joint and appear in several companies' legs. Series elastic actuators add a deliberate spring, measuring force by its deflection.

The computation below asks what each design pays to hold a 5 kg box at arm's length. An electric motor holding a load still draws current, and current in the windings makes heat (P=I2RP = I^2 R), so holding still costs power — unlike a locked brake. The motor constant KmK_m (newton-metres per square-root watt) captures this: a motor producing torque τm\tau_m dissipates (τm/Km)2(\tau_m / K_m)^2 watts. The numbers are illustrative values for this size of joint, not any company's specification.

python
import numpy as np

g = 9.81
# A shoulder holding the arm straight out with a 5 kg box in the hand.
arm_mass, arm_com = 4.0, 0.28          # kg, and metres from the shoulder (upper arm, forearm, hand)
box_mass, box_dist = 5.0, 0.55
tau = g * (arm_mass * arm_com + box_mass * box_dist)
print(f"shoulder torque needed: {tau:.1f} N m")

# Four ways to build the joint. Joule heat in a motor's windings is (motor torque / Km)^2,
# where Km is the motor constant (N m per square-root watt); a bigger motor has a bigger Km.
# The rotor's inertia, seen through a gearbox of ratio N, appears at the joint multiplied by N^2.
designs = [  # name, ratio N, gearbox efficiency, Km, rotor inertia (kg m^2)
    ("quasi-direct drive, large motor", 9, 0.93, 0.60, 6.0e-4),
    ("planetary, compact motor", 30, 0.88, 0.15, 1.2e-4),
    ("strain-wave, compact motor", 50, 0.80, 0.15, 1.2e-4),
    ("strain-wave, compact motor", 100, 0.75, 0.15, 1.2e-4),
]
Rth, max_rise = 1.5, 80.0              # K per W from windings to air; tolerable temperature rise in K
limit = max_rise / Rth
print(f"heat a motor can shed continuously: about {limit:.0f} W\n")
print(f"{'design':32s} {'N':>4s} {'motor torque':>13s} {'heat':>7s}   {'all day?':8s} {'reflected inertia':>18s}")
for name, N, eta, Km, J in designs:
    tau_m = tau / (N * eta)
    heat = (tau_m / Km) ** 2
    print(f"{name:32s} {N:4d} {tau_m:9.2f} N m {heat:6.0f} W   {'yes' if heat <= limit else 'no':8s} {N**2 * J:12.3f} kg m^2")

# Holding with both arms for an hour, with the quasi-direct-drive joints.
N, eta, Km = 9, 0.93, 0.60
both_arms_wh = 2 * (tau / (N * eta) / Km) ** 2 * 1.0
print(f"\nboth shoulders, quasi-direct drive, one hour of holding: {both_arms_wh:.0f} Wh of heat")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Holding the box needs 38 N m at the shoulder. The quasi-direct-drive joint, ideal for walking and for feeling contact, makes 57 W of heat doing nothing, just over what it can shed indefinitely; the 100:1 strain-wave joint makes only 11 W but presents 24 times the reflected inertia to anything it bumps into. Both shoulders holding for an hour turn 114 Wh into heat — about 5% of a 2.3 kWh battery. This is why humanoid designers mix actuator types joint by joint, add brakes or clutches to joints that hold loads, and watch winding temperatures in software.

Hands

Hands are where humanoids differ most. Many in service use simple grippers or six-actuator hands whose fingers are driven through linkages (for example, the Inspire hands common on Chinese humanoids and used in Gemini Robotics 2's tests); others put more actuators in the forearm and pull the fingers through tendons, approaching the 20-odd degrees of freedom of a human hand; Gemini Robotics 2 used 22-degree-of-freedom hands from Sharpa. Hands are also the part that breaks: Figure reported that the forearm was the most frequent hardware failure in its BMW deployment, and redesigned it for Figure 03. Dexterous hands and touch covers hand kinematics and tactile sensing in depth.

ComputationC

The energy budget

A humanoid carries its energy on its back, usually a lithium-ion pack of 1–3 kWh; Figure states 2.3 kWh and five hours "at peak performance" for Figure 03, an average draw of about 460 W. Where does the power go? The standard measure of locomotion efficiency answers part of the question.

Definition— Cost of transport

The (specific) cost of transport of a machine of mass mm moving at speed vv while consuming power PP is the dimensionless number

c=Pmgv,c = \frac{P}{m g v},

the energy spent per unit weight per unit distance. A person walking has a metabolic cost of transport of about 0.2; Collins and colleagues (2005) estimated Honda's ASIMO at about 3.2, sixteen times worse, and built passive-dynamics-based walkers that matched the human figure.

Companies rarely publish their robots' cost of transport, so the budget below brackets it between 0.5 and 1.0, alongside the loads that do not depend on walking: an onboard GPU module (NVIDIA's Jetson AGX Thor, a common choice in 2025–26, is rated 40–130 W), sensors and fans, the small but constant effort of balancing, and the arms. Every figure is an assumption stated in the code; change them and see what moves.

python
g = 9.81
mass = 70.0                 # kg, robot
battery_kwh = 2.3           # the pack size Figure has published for Figure 03
usable = 0.9                # keep 10% in reserve to protect the cells
pack_wh_per_kg = 180        # pack-level specific energy of a good lithium-ion pack
print(f"battery: {battery_kwh} kWh, about {battery_kwh * 1000 / pack_wh_per_kg:.1f} kg "
      f"({battery_kwh * 1000 / pack_wh_per_kg / mass:.0%} of the robot's mass)")

def walking_power(cot, speed, payload=0.0):
    """Electrical power to walk: cost of transport x weight x speed."""
    return cot * (mass + payload) * g * speed

base = {"computer (GPU module)": 100, "sensors, radios, fans": 40, "balancing in place": 60}
scenarios = {
    "standing idle": {},
    "standing, sorting light parts": {"arms working": 120},
    "standing, holding a 5 kg box": {"arms holding": 114},      # from the actuator calculation
    "walking 1.2 m/s, CoT 0.5": {"legs": walking_power(0.5, 1.2)},
    "walking 1.2 m/s, CoT 1.0": {"legs": walking_power(1.0, 1.2)},
    "walking 1.2 m/s, CoT 1.0, 10 kg tote": {"legs": walking_power(1.0, 1.2, 10), "arms holding": 114},
}
print(f"\n{'scenario':40s} {'power':>7s} {'runtime':>9s}")
for name, extra in scenarios.items():
    total = sum(base.values()) + sum(extra.values())
    hours = battery_kwh * 1000 * usable / total
    print(f"{name:40s} {total:5.0f} W {hours:7.1f} h")

# For comparison: a person's metabolic cost of transport when walking is about 0.2.
print(f"\nperson, 70 kg, walking 1.2 m/s at a metabolic CoT of 0.2: {0.2 * 70 * g * 1.2:.0f} W")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Three lessons sit in the table. First, a humanoid that mostly stands and handles light parts can last a shift of six hours or so, consistent with the published 2.3 kWh and five hours; one that walks all day cannot, and at a cost of transport of 1.0 a loaded robot runs flat in about 1.6 hours. Second, the battery itself is 12.8 kg, 18% of the robot, and every extra kilogram of battery must be walked around too. Third, at walking pace a person spends about 165 W where the robot spends about 410–820 W on its legs alone. So deployments are designed around the battery: robots that dock to charge between tasks (Agility's Digit), robots that swap their own batteries (UBTech's Walker S2 does so in about three minutes, by the company's account), and tasks chosen to keep walking short.

ExplanationH

Whole-body control and learned policies

A humanoid has 30-odd joints, touches the world through two feet and two hands, and is never statically stable while walking. Its controller is layered by time scale.

task layer: VLA model or task planner1–10 Hz · what to dowhole-body controller: optimisation or learned policy50–1000 Hz · balance, contacts, joint targetsjoint servo loops · 1–40 kHz · current and torquemotors, encoders, IMU, force sensorsstateestimatorIMU, encoders,foot contacts
The control stack of a 2026 humanoid. Slow layers decide what to do; fast layers keep the robot upright and make the joints obey. A state estimator fuses the IMU, joint encoders and foot contacts for every layer.

Optimisation-based whole-body control was the dominant approach through the early 2020s and remains common. A model-predictive controller on a simplified model — the robot's centre of mass and the contact forces at its feet, the centroidal model — plans a second or so ahead (Model predictive control, Legged locomotion). Then, a thousand times a second, a quadratic program chooses joint accelerations, joint torques and contact forces that best achieve several prioritised tasks — keep the centre of mass over the feet, put the swing foot here, put the hand there — subject to the full rigid-body dynamics of Robot dynamics, M(q)q¨+h(q,q˙)=STτ+JcTfcM(q)\ddot q + h(q, \dot q) = S^{\mathsf T}\tau + J_c^{\mathsf T} f_c, friction cones on the contact forces, and torque limits. It is interpretable and precise, and it needs good models.

Learned locomotion has largely taken over walking since about 2023. A neural-network policy maps the estimator's state and a velocity command to joint targets at 50–100 Hz; it is trained by reinforcement learning across thousands of simulated robots in parallel (Isaac Lab, MuJoCo) with randomised masses, frictions and delays, then transferred to hardware (Reinforcement learning for robots, Sim-to-real). Radosavovic and colleagues' 2023 transformer policy for Agility's Digit, trained only in simulation and walking outdoors in Berkeley, was an early public demonstration. Such policies recover from shoves and stumbles that would defeat a hand-built controller, but their behaviour is harder to certify.

Learning from human motion followed. Motion-capture recordings of people are retargeted to the robot's skeleton and a policy is trained to track them, which yields natural-looking walking, dancing and kicking. Stanford's HumanPlus (2024) let a person "shadow" a humanoid in real time from a single camera; NVIDIA and CMU's HOVER (2024) trained one controller for many command modes; ASAP (2025) corrected the simulator using real-world data to make agile skills transfer; BeyondMimic (2025) combined motion tracking with diffusion-guided control. The same machinery serves teleoperation for collecting demonstrations, from virtual-reality headsets (Open-TeleVision, 2024) to full-body motion-capture suits.

On top sits the task layer: increasingly a vision-language-action model of the previous lesson. Figure's Helix drives the upper body at 200 Hz while a separate learned controller handles the legs; Google DeepMind's Gemini Robotics 2 (July 2026) is the most prominent claim of a single model controlling walking, crouching and manipulation together.

ExplanationH

Who builds humanoids (as of late 2026)

The field moves monthly; the list below is a factual snapshot of prominent organisations, not a ranking, with the dates of the events mentioned.

RegionOrganisations and robots
United States and CanadaAgility Robotics (Digit, legged with a simple torso; logistics); Apptronik (Apollo; pilots with Mercedes-Benz and GXO from 2024; a Gemini Robotics hardware partner); Boston Dynamics, owned by Hyundai (electric Atlas from 2024; product version shown January 2026); Figure AI (Figure 02 in 2024, Figure 03 in October 2025); Tesla (Optimus, for its own factories first); 1X, founded in Norway as Halodi Robotics (NEO, aimed at homes); Sanctuary AI in Canada (Phoenix, with an emphasis on dexterous hands)
ChinaUnitree (H1, G1 and the small R1; quasi-direct-drive legs; listed on the Shanghai STAR Market in 2026); UBTech (Walker S series, trialled in car plants); AgiBot (several humanoid lines and the AgiBot World dataset); Fourier (GR series, the robot used in NVIDIA's GR00T N1 paper); Galbot (wheeled dual-arm robots); XPeng (Iron); and many younger firms
EuropePAL Robotics, Spain (TALOS and KANGAROO research humanoids); IIT, Italy (iCub and ergoCub); Neura Robotics, Germany (4NE1); Pollen Robotics, France, part of Hugging Face since 2025 (Reachy 2, wheeled); Hexagon (AEON, 2025); Humanoid, United Kingdom (HMND 01, 2025)
Japan and KoreaHonda (ASIMO, 2000, retired in 2022); AIST (HRP series); Toyota (T-HR3) and the Toyota Research Institute, which works with Boston Dynamics on learned behaviours for Atlas; Kawasaki (Kaleido); Rainbow Robotics, Korea, from KAIST's HUBO lineage, in which Samsung became the largest shareholder at the end of 2024

By market analysts' estimates, Chinese companies made the large majority of humanoids shipped in 2025 and the first half of 2026 — more than 90% in several counts, led by AgiBot and Unitree — helped by a dense supply chain for motors, gears and batteries and by research, education and demonstration buyers. For scale, the International Federation of Robotics counted about 542,000 industrial robots installed in 2024 and 4.7 million in operation: humanoids are still a small fraction of robots at work.

Pilot deployments as of 2026

  • Car plants. Figure reports that Figure 02 robots worked ten-hour shifts at BMW's plant in Spartanburg, South Carolina, for about eleven months in 2024–25, loading sheet-metal parts onto a welding fixture (three parts within 5 mm in about two seconds) for more than 30,000 cars, over some 1,250 hours of runtime; BMW began using Figure 03 for parts sequencing there in 2026. UBTech's Walker S robots have been trialled in Chinese car factories including those of BYD and Zeekr. Hyundai plans to use Atlas in its own plants, and Boston Dynamics says its whole 2026 output of the product Atlas, which swaps its own batteries, is committed to Hyundai's robotics application centre and Google DeepMind.
  • Logistics. Agility's Digit has moved totes for GXO under a robots-as-a-service contract since 2024, and in February 2026 Toyota Motor Manufacturing Canada signed an agreement for seven Digits to unload parts totes at its Woodstock, Ontario plant.
  • Homes. 1X opened orders for NEO in October 2025 at USD 20,000 or a monthly subscription, and opened a factory in California in 2026. When NEO cannot do a household task by itself, a 1X operator wearing a VR headset can take over remotely during sessions the owner schedules — an honest answer to today's limits that raises its own questions of privacy.

Most of these are pilots with a handful of robots on one line or site, doing one or two tasks. What is still missing, by most practitioners' accounts:

  • Reliability. A factory measures availability in nines; humanoids still fall, overheat and break hands, and published mean time between failures is rare.
  • Speed. Many demonstrations run slower than a person; robot videos are often sped up, and responsible companies say so.
  • Generality. Each pilot task still needs task-specific data and engineering. The promise of the form — one robot, many jobs — is not yet demonstrated at scale.
  • Safety certification. There is not yet a published safety standard for industrial robots that must balance actively (below), so most deployments keep humanoids apart from people.
  • Cost. Covered next.
ComputationC

Cost and safety

Cost. List prices span two orders of magnitude. Unitree's G1 launched in 2024 from 99,000 yuan (about USD 14,000) and its smaller R1 in 2025 from 39,900 yuan (about USD 5,600, cut further in 2026), sold largely for research and education; NEO is priced at USD 20,000; industrial humanoids are mostly leased as robots-as-a-service, with prices not published. Actuators dominate the bill of materials, and falling actuator costs are the main reason prices are dropping. But purchase price is only part of the cost of a working robot: integration, maintenance, supervision and downtime matter as much, and an exercise asks you to compute a cost per working hour.

Safety. A fenced industrial arm has a simple safe state: cut power to the motors and apply the brakes (Embedded control and functional safety). A standing humanoid has no such state. Cut its power and it falls, as an inverted pendulum pivoting about its feet:

python
import numpy as np

g = 9.81
# If a standing humanoid's motors are switched off, it tips over like a rigid inverted pendulum
# pivoting about its feet: theta'' = (g / L) sin(theta), with theta measured from vertical.
def fall(L, theta0_deg, dt=1e-4):
    theta, omega, t = np.radians(theta0_deg), 0.0, 0.0
    marks = {}
    while theta < np.pi / 2:                      # until it lies flat
        omega += (g / L) * np.sin(theta) * dt     # semi-implicit Euler
        theta += omega * dt
        t += dt
        for deg in (10, 30, 60):
            if deg not in marks and theta >= np.radians(deg):
                marks[deg] = t
    return t, marks, omega

L, mass = 0.95, 70.0                              # height of the centre of mass (m), robot mass (kg)
for start in (0.5, 2.0, 5.0):
    t, marks, omega = fall(L, start)
    print(f"start tilt {start:3.1f} deg: 10 deg at {marks[10]:.2f} s, 30 deg at {marks[30]:.2f} s,"
          f" flat at {t:.2f} s; head speed at impact {omega * 1.7:.1f} m/s")

energy = mass * g * L
print(f"\nenergy released as the centre of mass drops {L} m: {energy:.0f} J")
print(f"time constant sqrt(L/g) = {np.sqrt(L / g):.2f} s")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

From a small tilt a 70 kg robot is on the ground in 1.1–1.9 seconds, with its head moving at several metres per second in this stiff-plank model (a real robot crumples, which helps a little) and 650 J released — enough to injure someone standing nearby. The time constant L/g≈0.31\sqrt{L/g} \approx 0.31 s is why balance control must run fast and why "stop" for a humanoid has to mean a controlled action: crouch, kneel or sit while still powered, then switch off.

Common mitigations in 2026 pilots: zones separated from people, speed limits near them, wheeled or statically stable bases for work beside people, controlled-crouch behaviour on any fault, padded and rounded designs, and lower mass for home robots (NEO is reported at about 30 kg). For learned controllers, runtime monitors and safety filters add a further layer (Evaluation, safety and deployment). If you build or program robots yourself, the safe place to practise whole-body control is simulation and small, light, slow robots; full-size humanoids are machines for trained teams behind barriers.

Common misconceptionsH

Common misconceptions

Misconception

"Humanoids are about to replace people across factories." As of 2026 humanoids work in pilots of a few robots doing one or two tasks each, while about half a million conventional industrial robots are installed every year. The human form's promise is generality, which is exactly what has not yet been shown at scale.

Misconception

"A robot holding still uses no energy." An electric motor holding a load draws current and makes heat; the lesson's shoulder makes 57 W holding a 5 kg box. Brakes, clutches and high gear ratios reduce this, at other costs.

Misconception

"The human form is the most efficient design." On flat floors, wheels beat legs on energy, speed and safety. Legs earn their place on stairs, kerbs and clutter; many 2026 designs put a humanoid upper body on wheels for exactly this reason.

Misconception

"If something goes wrong, just hit the emergency stop." For a balancing robot, cutting power is itself dangerous: it falls in about a second. A humanoid's stop must be a controlled behaviour, which is one reason a dedicated standard is being written.

Misconception

"The video shows the robot doing it by itself." Some public demonstrations have been teleoperated or assisted, and some videos are sped up. Responsible companies label this; when they do not, ask.

Exercises

Exercises

1. Humanoid robots

Warm-upHuman

A 60 kg humanoid walks at 1.01.0 m/s with a cost of transport of 0.80.8 (electrical energy per unit weight per unit distance). Taking g=9.81g = 9.81 m/s², how many watts does walking draw, to the nearest watt?

2. Humanoid robots

Warm-upHuman

A humanoid has a 2.0 kWh battery, of which 90% is usable. On a delivery route it draws 450 W for walking, 120 W for its computer and 30 W for everything else. How many minutes can it work before it must recharge?

3. Humanoid robots

CoreHuman

Many legged robots use quasi-direct-drive actuators (a large motor with a gear ratio of about 6–10:1) in their legs rather than compact motors with 100:1 strain-wave gears. What is the main reason?

Choose one

4. Humanoid robots

CoreHuman

A fenced industrial arm is made safe by cutting power to its motors and engaging brakes. Why is this not an adequate safe state for a walking humanoid?

Choose one

5. Humanoid robots

CoreComputational

Write two functions for an electric joint with gear ratio ratio, gearbox efficiency eta and motor constant km (N m per square-root watt). joint_heat(load_torque, ratio, eta, km) returns the Joule heat in watts while statically holding load_torque at the joint, using motor torque τm=τ/(Nη)\tau_m = \tau / (N\eta) and heat (τm/Km)2(\tau_m / K_m)^2. max_hold_torque(ratio, eta, km, heat_limit) returns the largest joint torque that can be held indefinitely without exceeding heat_limit watts.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

6. Humanoid robots

CoreComputational

Write cost_per_hour(price, years, hours_per_year, maintenance_frac, power_kw, energy_price), the cost of one working hour of a robot: its purchase price spread evenly over its working life, plus yearly maintenance (a fraction of the price each year) spread over the year's working hours, plus electricity. All money is in the same currency unit.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

7. Humanoid robots

CoreHuman

Which statements about humanoid robots are accurate as of 2026? Select all that apply.

Choose all that apply

Further reading and resources