How humanoid robots work

The short answer: a humanoid is built roughly like a person so it can use stairs, shelves and tools made for us. It still runs sense, decide, act — but the "act" step has to keep two legs balanced while the hands work, on a battery that lasts a couple of hours. Most viral clips are demos. A smaller set already move totes and car parts for paying customers.

Part 1: Why I wanted to understand humanoids

In my first guide I treated humanoids as one family among many: arms, dog-shaped robots, cars that drive themselves. Then the news feed filled up with them. Agility Robotics started trading on Nasdaq as AGLT. Meta reportedly built a prototype with Fourier. Anyverse's wheeled K15 started work at JD coffee shops in Beijing. I was writing headlines about robots shaped like people without being able to explain, in plain words, how they stay upright or why anyone builds them that way.

So I went back to the primary sources: company blogs, store pages, factory deployment write-ups. This guide is what I learned.

Same disclaimer as always. I'm not an engineer. I'm learning robotics in public, and I'm skeptical of demos by default. In my second guide I wrote about how a real robot can still leave you with the wrong impression when someone is teleoperating it, the clip is sped up, or you're watching the one good take. That checklist matters more for humanoids than for almost anything else I cover. Every number below has a link. Where I couldn't find a source, I left it out.

Part 2: How humanoids work

Why build a robot shaped like a person?

The short reason is the world. Stairs, door handles, shelves, aisles, tools, conveyor heights — almost all of that was sized for human bodies. Boston Dynamics puts it plainly: "The humanoid form factor is a useful design for robots working in a world designed for people" (Boston Dynamics). Agility says the same thing about Digit: two legs, two arms and a torso so the robot can use infrastructure built for people without a full warehouse rebuild (Agility).

That doesn't mean every useful robot should look human. Boston Dynamics also sells Stretch, a wheeled box-mover for warehouses, and Spot, a dog-shaped inspection robot. Wheels are great on flat floors. Four legs are more stable than two. Humanoids are the bet that general work in human spaces is worth the harder balance problem.

One more nuance I like from Boston Dynamics: their new Atlas is meant to look human-ish but move beyond a human range of motion when that helps the job (Boston Dynamics). "Shaped like us" is a starting point, not a costume.

Sense, decide, act — with two legs and two hands

Same loop as the beginner's guide:

1. Sense — cameras, depth sensors, sometimes lidar; joint encoders; an IMU that does the job of your inner ear.
2. Decide — classical control, learned walking policies, and newer vision-language-action models that turn a language goal into motor commands.
3. Act — electric motors at the joints move legs, torso, arms and fingers.

On a humanoid the three steps are tangled. If you pick up a heavy tote, your balance changes. If you take a step, your reach changes. That's why demos that look smooth are either very carefully scoped, or very carefully filmed.

Keeping balance while you walk and carry

Two legs fall over easily. The cheap fix is to freeze walking before you grasp. The harder fix is to keep balancing while the hands work.

Figure's Helix 02 post treats walking and manipulating as one continuous behavior — loco-manipulation — instead of "walk, stop, grasp, walk again" (Figure). Agility's Digit write-ups describe the same real-world problem in warehouses: navigating tight spaces while carrying a tote, then stacking when the conveyor backs up (Agility). Under the hood that usually means an IMU watching tilt, joint encoders watching limb angles, and a control loop that updates motor commands many times a second. For the classical and learned-control side of that loop (model predictive control, sim-to-real), start with how robots work.

Muscles: electric actuators and degrees of freedom

Older Atlas used hydraulics. In April 2024 Boston Dynamics retired that research platform and unveiled a fully electric Atlas aimed at real work, starting with Hyundai factories (Boston Dynamics). Electric motors are quieter, cleaner and easier to maintain than hydraulic fluid. The industry moved that way for a reason.

"Degrees of freedom" just means independently controllable joints. More joints usually means more ways to reach and grasp, and more things that can go wrong. Unitree publishes this number for the G1: the base model has 23 joint degrees of freedom; EDU configs go up to 43 with optional waist joints and hands (Unitree). Figure's Helix post describes controlling a 35-DoF upper-body action space at 200 Hz on its robots (Figure). I don't treat those two numbers as apples-to-apples — different companies count hands and wrists differently — but they give you the scale of the control problem.

Hands are the expensive tradeoff. A simple gripper can move a tote all day. Five fingers can, in theory, open a bottle or pick a pill. Figure's Helix 02 post describes fingertip tactile sensors sensitive enough to feel forces as small as three grams, plus palm cameras for when the head cameras can't see the object (Figure). Agility's Digit 5 takes the other practical path: swappable end-effectors on ISO-standard mounts, so the tool can match the job (Agility). Neither approach is "cheating." They're different bets about what work pays first.

Senses: IMU, vision, and lidar where it's used

Two kinds of sensing matter, the same split as in the beginner's guide:

• Looking inward. Joint encoders say how bent each limb is. An IMU measures tilt and acceleration. Digit 5's own description lists onboard cameras, depth sensors and inertial measurement units (Agility). Without that inward sense, a biped doesn't know it's falling until it's already falling.
• Looking outward. Cameras for color and detail. Depth cameras or lidar for distance. Unitree's G1 ships with a depth camera and 3D lidar (Unitree). Boston Dynamics' Atlas perception blog describes a vision-heavy stack: calibrated cameras, joint kinematics, object detection and pose tracking for factory parts — including shiny metal parts that are hard for cameras (Boston Dynamics). That post is worth reading if you want the unglamorous version of "the robot sees."

I haven't found a primary Tesla page that lays out Optimus's full sensor list the way Waymo does for robotaxis, so I'm not going to invent one. For the car-side camera bet, see how robotaxis work.

Teleop vs autonomy (and why the video checklist still applies)

A lot of humanoid skill still starts with a person. Figure trained Helix on roughly 500 hours of teleoperated demonstration data, then auto-labeled the clips with language (Figure). Agility describes Digit's learning pipeline as blending traditional control, teleoperated demonstrations, policy training and reinforcement learning in simulation (Agility). Teleop isn't fake. It's how the training data gets made.

What is a problem is a clip that lets you believe the robot decided on its own when a human was steering. That's the whole point of how to tell if a robot video is real. For humanoids I watch for:

Figure's Helix 02 dishwasher video is an example of a company going out of its way to say the run is fully autonomous with no human intervention (Figure). Believe the caption, then still ask how often it works.

Who is actually doing paid work?

This is the table I wish every humanoid launch post came with.

Two reading notes. First, "deployed" and "doing useful work" are not the same as "replaced a whole job." Digit's early GXO work is tote handling between AMRs and conveyors — real, repetitive, valuable, and narrow (GXO). Second, wheeled upper-body robots like Anyverse's K15 can also do paid work in human spaces; legs aren't the only path (our story).

Batteries: the boring limit that decides the shift

Humanoids still eat a lot of power for the work they do. Unitree lists the G1's smart battery at 9,000 mAh with about two hours of battery life (Unitree). Agility says Digit 5's battery is designed for about 90 minutes of runtime and a 9-minute charge, a 10:1 run-to-charge ratio, which is how they argue for more than 20 hours of productive work in a 24-hour day with fast top-ups (Agility).

I treat those as company claims about their own hardware, not independent lab tests. The useful habit is: when a humanoid story talks about a full workday, ask whether that means one charge, a swappable pack, or a robot that spends most of the day plugged into a charger between short runs.

What to watch for in the news

When the next humanoid clip hits your feed, I ask:

1. Is someone paying for this shift? Customer names, hours, tote or part counts beat adjectives.
2. Legs or wheels? Both can be "humanoid" in marketing. The hardware choice tells you what floor and what job.
3. Autonomy claim, in writing. "Teleoperated," "autonomous," "1x speed," or silence. Silence is information.
4. Battery math. Runtime, charge time, and whether the robot can swap packs.
5. Hands vs grippers. Fancy fingers are impressive. A tool that matches the tote may ship sooner.
6. Plan or fact? "Expected production later this year" and "moved 100,000 totes" are different kinds of sentence. Tesla's Optimus line is still in the first category (Tesla Q2 2026 update). Digit's GXO tote count is the second (Agility).

FAQ

Why not just use a robot arm on a wheeled base? Often you should. Arms and wheeled lifters are simpler when the job is fixed. Humanoids are the bet that the job will keep changing, or that the building won't be rebuilt for machines.

Are humanoids autonomous now? Some behaviors are. Digit's warehouse shifts and Figure's BMW loading were real customer work. Many viral clips are still teleop, research, or carefully staged. Read the caption the way you would in the video guide.

What's the cheapest humanoid I can actually buy? As of November 6, 2026, Unitree's store lists the G1 at $13,500 before tax and shipping, currently listed as backordered (Unitree store). That's a developer platform, not a household butler.

Is Atlas still the parkour robot? The hydraulic parkour Atlas was retired in 2024. The current Atlas is electric and aimed at factory work with Hyundai (Boston Dynamics).

What I'm still figuring out

I don't know yet whether five-fingered hands or swappable grippers win the first decade of paid humanoid work. Digit's tote numbers make a strong case for the boring tool. Figure's Helix posts make a strong case that dexterity is getting less imaginary. Both can be true for different jobs.

I also don't know how to compare "hours of operation" across companies, because each one defines the metric. Until more of them are public and filing numbers the way Agility now has to, I read every count as the company's own scoreboard.

If I got something wrong here, or a number has moved since I checked, tell me (how to reach me is on the About page). I'll fix it and note the change.

Follow the latest: humanoid robots, Unitree, Tesla, and Agility’s Nasdaq listing.

Last updated November 6, 2026. Prices are list prices on the date shown and change often.

— Nate