Humanoid Robots: Bipedal Machines and the Bet on Human Shape

A moving statue is an old wish. It is not this control problem

People have built figures that bow, write, or play when a cam comes around. Hero’s devices, clockwork, the Jaquet-Droz automata: predetermined motion, a mechanism, an audience. The engineering page and the robotics page already refused to stretch those into a Unimate. Refuse it here too. A cam that lifts a wooden hand is not the parent of a powered knee that has to keep a battery and a computer from hitting the concrete.

The wish is old. The problem is newer: a machine with its mass up high, feet that leave the ground, and a room that was not fixtured for it. Fascination is not a lineage.

The useful argument is the room you do not want to rebuild

A plant or a warehouse is full of handles, ladders, door levers, and shelf heights that assume a person between roughly five and six feet, with arms in front. A specialist machine is faster when you can bolt it down and bring the work to it. The humanoid sales pitch is the other case: the work is scattered, the building stays, and you would rather teach a body than redesign the aisle.

That pitch can be honest and still be a bad buy. If the job is one motion all day, the specialist still wins, which is what the welding gun and the wire-guided cart already showed. A humanoid earns a trial only where the variety of little jobs is the point, and only after it does them on the bad day, not the filmed day.

Walking is a loop that fails into a fall

A person stays upright because a nervous system corrects before we notice. A robot stays upright only while the loop is fast enough. Joint angles, force under the feet, a gyro, a guess at where the mass is going. The controller moves a motor. The next reading says whether the guess held. Miomir Vukobratović and Juri Stepanenko, in 1972, gave the field a name for one stability check: the zero-moment point, where the ground reaction keeps the foot from tipping. Honda’s own write-up of ASIMO uses that idea. It is a calculation. It is not balance in the sense of a person who knows the floor is wet.

A shove, a cable on the ground, a missed step: the correction is late or it is not, and the machine goes down. Weight that was a specification becomes an impact. The robotics page said a fenced arm is dangerous because of speed and mass. A free biped adds the fall. Do not file that under “extraordinary complexity” and walk away. File it under what hits the floor.

The actuators are the weight they have to lift

A muscle pulls across a joint. A humanoid uses a motor and a gearset, or a hydraulic cylinder, at joints that also have to carry those motors. Honda’s path from the P2 to ASIMO was in large part a fight about mass and about electric actuators, not a fight about a soul. Every kilogram in the leg is a kilogram the hip must swing. Batteries are mass too. A heavier pack can run longer and makes the next step cost more. I will not quote a runtime from a keynote. Runtime is whatever the pack and the gait did that day, on that floor, with that payload.

Heat, gear wear, and a harness that flexes are the shift. A backflip in a video does not service a gearbox.

A five-finger hand is a research bet. A gripper may be the tool

The human hand does a coin and a hammer with the same bones. Copying that takes many joints, many tendons or motors, and many ways to break. Most of the useful contact in a pilot is still a simpler gripper or a suction cup aimed at a tote, which is the warehouse problem again, now on a wrist that can also fall over. Reaching a bench is not using the tool on the bench. A key, a connector, a soft bag: the robotics page already said piece picking fails on the live item. A humanoid wrist does not repeal that.

A camera labels. It does not understand the room

Cameras, depth sensors, force sensors, and an inertial unit report numbers. Software may label a handle, guess a distance, or decide a person-shaped blob is in the aisle. The warehouse page said vision is a sensor with a lot of bits. Same here. Oil, glare, a door that does not match the training pictures, a person bent over: the label can be wrong, and a wrong label on a walking machine is a wrong step, not a wrong cell on a spreadsheet.

Do not say perception connects intelligence to the body. It connects a classifier to an actuator. The computers page already said the instruction is not a mind. A classifier is not a mind either.

A learned policy is still a program with a wider miss

An industrial arm repeats points in a fixture. A humanoid is aimed at rooms that are not fixtured. One way to chase that is a model and a plan: where to stand, which arm, which grasp, what collides. Another way is a policy trained on examples, which may generalize and may also do something no one programmed on purpose. Both are tools. Neither is judgment. “Physical AI” is a sales name for putting those tools on a body. Intelligence has not left the screen. A body has been bolted to a program that can be wrong in the room instead of wrong on a page.

Language in, motion out, is a demo until the motion is the motion you needed, on the object you have, with a stop that works when the language was vague. An instruction a person would shrug at is not clearer because a speaker said it to a robot.

Where this sits in the history of automation

The machine may be told to fetch a tote, walk to a mark, and set it down. That is a task list. If the tote is missing, the honest machine stops. The dishonest failure is the one that completes the gesture on the wrong object and logs success.

The computers page said a network does not make a number true. A humanoid’s report that the task succeeded is that kind of number.

General purpose is a hope. Specialists still have the job

A mill ground grain. A loom wove. A weld robot welds. Those machines are narrow, and that is why they earn their keep. A humanoid is the hope that one platform can be aimed at many small jobs so you do not buy a new machine each time the job changes. Hope is the right word. Reprogramming a platform is not free. You still teach, still fixture what you can, still keep a person for the exception, still own the failures. The robotics page refused “the arm began to move on its own.” This page refuses “one body does the plant.”

A factory aisle is a harder room than a stage, and still not the world

Companies trial these machines in plants and warehouses because the aisles are repetitive and the work is physical, and because the building was laid out for people. Known totes help. They do not remove the wet floor, the extension cord, the pallet in the wrong place, the coworker who steps out. The warehouse page already said piece picking is mostly hands. A humanoid pilot does not silently overturn that. Count the jobs it finished without a rescue. Do not count the jobs in the video.

The 2015 trial is the public record of the hard room

The DARPA Robotics Challenge finals, in 2015, sent robots — many of them bipeds — through a course of driving, a door, a valve, a wall, rough ground, stairs. Falls were ordinary. Finishing was news. Those were disaster-response tasks under a rules sheet, not a Tuesday in a warehouse, and not every machine there was a humanoid. The lesson that travels is the one the films of parkour do not show: contact with an unplanned world is where these bodies lose. A course completed once is not a maintenance log.

Safety is what happens when the loop is wrong

A good demo is not a safety case. You need the speed, the mass, the reach, the battery, the sharp tool if there is one, and the behavior when power drops, when a joint faults, and when a person is closer than the plan assumed. Industrial-arm rules were written for arms, often fenced or speed-limited. A walking body that can topple is a different hazard. The cobot page already said a soft sales word is not a risk assessment. “Humanoid” is not one either. Predictable failure matters more than a clever gait. If it cannot stop without falling on someone, it does not belong in the aisle yet.

A demo is a day. A tool is a year of the same day

Gearboxes wear. Harnesses break. Sensors drift. Batteries age. Heat soaks a joint that was fine for eight minutes on a stage. Reliability is the same demand the warehouse page made of a crane: the flow has to exist on the dull shift, after the visitors left. Maintenance is not a footnote under intelligence. For a machine this jointed, maintenance is most of the ownership.

I will not say they are becoming common, and I will not say they will not. The questions that decide it are dull on purpose. Does it finish the real task. Does it do it again tomorrow. Is it cheaper and safer than the person or the specialist machine for that task, not for a filmed task. Where those line up, a tool gets bought. Where they do not, the shape stays a research body. That is the pattern on the rest of this site. It is not a prediction.

What I want you to keep

The shape is a bet that the room stays built for people. Waseda’s WABOT-1, 1973, was a lab biped, not a first worker and not the ancestor of a product. Honda’s P2 was shown in 1996. ASIMO was shown at the end of October 2000, walked stairs in public, and was not a line worker. Vukobratović’s zero-moment point, 1972, is a stability calculation Honda used, not an explanation of a person. The DARPA finals in 2015 made falls ordinary on a hard course.

A camera labels. A policy misses. A five-finger hand is optional and fragile. A gripper does not repeal piece picking. Safety is the fall and the fault, not the face. A demo is not a shift. And the warehouse was right: a carton still does not need this body. Neither does a welder.