Tales of human-like dolls craving to grow to be actual individuals flip up all over the place. Pinocchio needs to be an actual boy. The robotic baby in Spielberg’s A.I. needs to be beloved like a human son. The story retains getting retold as a result of individuals assume the trajectory is clear. Construct one thing that appears human, hold bettering it, and someday the copy turns into indistinguishable from the unique.
What’s taking place on the bottom is stranger than that. At CES 2026, Boston Dynamics’ Atlas demonstrated wrists that bent backward and a torso that spun a full 180 levels. Elsewhere, humanoid robots are starting to diverge in much more placing methods. Some can swap their very own batteries by reaching each arms behind their backs. Others stroll on reverse-jointed legs. The human silhouette remains to be there, however the actions inside it have gone some place else fully.
There’s an apparent objection right here. Hasn’t copying nature labored earlier than? Generally. Gecko toe pads gave engineers the thought for dry adhesives. Sharkskin texture confirmed up in aggressive swimsuits. However in each instances, engineers borrowed the physics beneath, not the form. Those who tried to repeat pure varieties wholesale normally hit a wall.
For hundreds of years, individuals tried to construct ornithopters that flapped like birds, however none grew to become a sensible path to human flight. The Wright brothers received off the bottom not as a result of they merely imitated, however as a result of they moved past flapping and targeted on the rules of carry and management.
If evolution has spent tens of millions of years refining a design, why don’t engineers simply copy it? That query went to the Hubo Lab at KAIST. The lab constructed HUBO, the robotic that gained the 2015 DARPA Robotics Problem, and immediately it’s led by Prof. Park Hae-won. His staff’s latest work offers a way of the vary. Humanoid legs that dash at 12.6 kilometers per hour. A quadruped robotic that walks straight up vertical partitions. A one-legged hopper that launches into mid-air somersaults and lands on the identical leg.
From the middle of the again row, clockwise Hae-Gained Park, Dongyun Kang, Hajun Kim, JongHun Choe, Min-Su Kim
Picture: KAIST
Mimicking nature shouldn’t be at all times the appropriate reply.
At 12.6 kilometers per hour, an individual has to interrupt right into a run. A robotic constructed by Prof. Park Hae-won’s staff at KAIST can dash at that velocity on two legs. It glides by means of motions that appear like Michael Jackson’s moonwalk and picks its manner over tough terrain with a duck-like waddle.
One place to begin is biology. Roboticists have been borrowing nature’s tips for many years. Prof. Park’s robots do appear like they arrive from that custom. However he works the opposite manner round. As an alternative of learning an animal to construct one, he picks an issue and builds a machine to resolve it.
“In the event you’re growing expertise for high-speed motion, wheels will be an environment friendly selection,” Prof. Park stated. “There’s no must mimic the movement of a cheetah.”
A automobile on wheels outruns a cheetah. Evolution by no means got down to construct the quickest runner. It constructed the one more than likely to outlive.
“Learning pure organisms offers us a way of the extent of efficiency that may be reached when one thing is nicely designed,” Prof. Park stated. “It serves as a helpful reference for setting route throughout analysis and improvement.” He added “It’s necessary to view nature as one reference level. Relatively than replicating it instantly, it’s extra applicable to make use of it as a supply of concepts.”
Humanoids face the identical query. A human physique runs on muscular tissues, tendons, and chemical vitality. A robotic runs on steel frames, motors, and electrical energy. To repeat human motion faithfully you’d want synthetic muscular tissues, however motors nonetheless are inclined to outperform commercially accessible synthetic muscular tissues in lots of sensible metrics. So why handicap a robotic by forcing it to maneuver like a physique it doesn’t have?
MARVEL, a quadruped robotic from Prof. Park’s lab, was designed for grimmer work. Researchers needed a robotic that might transfer freely throughout the metal constructions of shipyards, bridges, and enormous storage tanks. Locations the place upkeep crews threat deadly falls.
Gecko ft or insect claws would possibly sound like the appropriate mannequin for a wall-climbing robotic. However actual industrial metal is rusted, layered in outdated paint, and caked with grime. Gecko-style adhesion would doubtless battle to carry heavy gear on surfaces like that.
As an alternative, Researchers constructed MARVEL with electro-permanent magnets in its ft. Standard electromagnets drain energy constantly to remain on. Electro-permanent magnets work otherwise. A quick electrical pulse rearranges the inner alignment of the magnet’s poles, switching the grip on or off. MARVEL’s ft lock and launch in about 5 milliseconds.
As soon as the magnets have interaction, the wall itself turns into the robotic’s floor. Three legs keep anchored whereas the fourth steps ahead. MARVEL travels at 0.7 meters per second on vertical partitions and at 0.5 meters per second whereas hanging the wrong way up from a ceiling. Its adhesive pressure reaches almost 54 kilograms, which is sufficient to carry not simply its personal weight but in addition heavy instruments.
“In the event you method a shipyard robotic from a biomimetic perspective, you would possibly conclude that it ought to resemble a human employee and deal with instruments the identical manner,” Prof. Park stated. “In the end, what issues is designing a system that matches the working atmosphere and the duty at hand.”
AI alone can not construct an ideal robotic.
Designing the physique is simply half the issue. AI and reinforcement studying have modified how robots study to maneuver, however what works in simulation nonetheless has to carry up on actual {hardware}.
Prof. Park’s staff trains its robots by means of reinforcement studying. The AI controls the robotic’s physique and figures out find out how to stroll by trial and error, falling and getting again up the way in which a toddler does. Doing that hundreds of instances on actual {hardware} would take ceaselessly. So researchers practice in simulation as a substitute.
Contained in the simulation, Prof. Park’s staff runs roughly 400 copies of the identical robotic without delay. Every copy falls and recovers below completely different circumstances, and what all of them study feeds right into a single AI community in actual time. Time itself will be compressed. What would take a few yr of bodily follow matches into roughly 4 hours on a high-performance pc. Prof. Park stated half a day of reinforcement studying is sufficient to get a robotic strolling.
The catch is {that a} robotic educated in simulation doesn’t at all times survive contact with actuality. A robotic that tumbles like a gymnast on display screen can lose its stability and topple the second it’s positioned on an actual flooring. Roboticists name this the sim-to-real hole. Simulations can’t seize each wrinkle of real-world physics, and the variations are sufficient to throw off an AI that realized in an easier world. Closing that hole is the place the KAIST staff’s {hardware} experience is available in.
One method Researchers took was to make the true robotic behave extra like its simulated twin. A giant motive AI struggles to regulate a bodily robotic is friction within the joints. Standard robots use off-the-shelf reducers with excessive gear ratios to amplify motor output. That provides the robotic highly effective pressure. On the identical time, inside friction makes all the pieces stiff, like pedaling a bicycle caught in excessive gear.
“In a gear system with a excessive discount ratio, it’s very arduous to pressure it to show from the surface,” Prof. Park stated. “In the event you connect a linkage and strike it with a hammer, the resistance is so intense that the gear tooth may shatter.”
Most simulations don’t account nicely for that friction. An AI that realized to stroll in a near-frictionless digital world loses its stability the second it hits the stiff resistance of an actual joint. So Prof. Park’s staff constructed its personal actuator that lower the gear ratio to roughly one-tenth of typical ranges whereas boosting the motor’s personal output. It’s a quasi-direct drive design, an idea first proposed at MIT. Much less friction within the {hardware} meant the true robotic moved extra just like the simulated one. After the adjustment, AI’s coaching truly carried over.
KAIST staff additionally labored the issue from the opposite route. As an alternative of constructing the {hardware} match the simulation, they made the simulation match the {hardware}. As a result of Prof. Park’s staff designed and constructed its personal motors, they’d detailed knowledge on how these motors truly behave.
That knowledge issues. Most simulations assume torque stays the identical regardless of how briskly the motor spins. Actual motors don’t work that manner. Spin sooner, accessible torque drops. Decelerate, accessible torque climbs. Coaching an AI on the simplified model will drive it to push the {hardware} past its limits. Prof. Park’s staff fed their precise torque-limit curves into the coaching, so the AI realized the place the motor’s ceiling was and stayed below it.
The place all of this comes collectively is KAIST’s hopping robotic. The entire machine is one leg. No arms, no second foot to catch itself. That type of stability drawback is brutal to resolve. In the intervening time Prof. Park had already gotten quadruped leg robotic strolling to work. As an alternative of transferring to 2 legs subsequent, he went straight to 1. As a result of If the algorithm can deal with the toughest case first, then two legs gained’t be an issue.
KAIST Humanoid v0.5
Researchers loaded all the pieces about the true robotic into the simulation. Its shifting heart of gravity, its inertia, and the bodily limits of its actuators. From there they ran almost the identical reinforcement studying algorithm they’d used for the quadruped. The AI found out find out how to stability on one leg. It began leaping. Earlier than lengthy it was doing mid-air somersaults, touchdown cleanly every time.
“Constructing the hopping robotic confirmed that our reinforcement studying algorithm and {hardware} design will be utilized below a variety of circumstances,” Prof. Park stated. “It gave us a chance to discover how our motor expertise and reinforcement studying methods would possibly prolong to the event of robots in many alternative varieties.”
Prof. Park doesn’t purchase the concept software program can clear up all the pieces. He’s watched junior researchers spend days debugging code when the true drawback was a unfastened screw or a damaged solder joint. When a robotic gained’t stroll, individuals attain for the algorithm first. They tweak the parameters, rerun the simulations, rewrite the management logic. In the meantime the precise fault is sitting proper there within the {hardware}. No quantity of code will tighten a screw. {Hardware} information isn’t going away simply because AI received good.
“Regardless of how refined the management expertise, there are limits to what will be achieved if the {hardware} can not sustain,” Prof. Park stated. “In robotic improvement, management and {hardware} are each essential. Neither will be thought-about in isolation.”
Can humanoid robots grow to be a part of our on a regular basis lives?
The cash pouring into humanoid robots proper now’s staggering. However loads of applied sciences have regarded simply as promising and gone nowhere. Honda spent over 20 years on ASIMO earlier than quietly retiring it. A robotic that walks throughout a stage at a commerce present shouldn’t be the identical factor as a robotic that survives a shift on a manufacturing facility flooring.
Prof. Park’s humanoid is being constructed for the manufacturing facility flooring. The goal payload is 25 kilograms or extra. Most humanoids available on the market prime out nicely under that. He selected that quantity due to the place South Korea is correct now. The nation runs one of many world’s largest manufacturing sectors, however the workforce is graying quick. Younger individuals aren’t lining up for welding jobs or assembly-line shifts. The slack is being picked up by older expert staff and overseas laborers, and there aren’t sufficient of both. A robotic that may solely carry mild objects is ineffective in that atmosphere. The quasi-direct drive actuators and customized motors his researchers have been constructing exist for precisely this sort of work.
The manufacturing facility flooring isn’t the one attainable market, although. Prof. Park introduced up drones. For many years solely the army and some infrastructure inspectors bothered with them. Then YouTube creators began wanting aerial photographs and went in search of one thing that might fly a digicam. Drone firms shipped an affordable quadcopter with an honest digicam mount. Inside a number of years a shopper drone business had grown up round a necessity that hardly existed earlier than. Prof. Park thinks humanoids may go the identical manner. The use that really drives adoption may be one no one within the business has imagined but.
On the shut of the interview Prof. Park stated, “I consider robots ought to complement individuals, not compete with them. My hope is that robots will in the end be used to counterpoint individuals’s lives and free them to pursue extra fulfilling work.”
The story was produced in partnership with our colleagues at In style Science Korea.
2025 PopSci Better of What’s New
Leave a Reply