South Korean researchers develop task AI enabling robots to learn, execute everyday chores

Related News

• The birth of humanoid robots (4): A Christmas Carol for steel and silicon Graphics by AJP Song Ji-yoon Editor's Note: This is the fourth and final installment in AJP's series on humanoid robotics, examining the anatomy, technologies and economic logic behind one of the most hyped industries of the decade. SEOUL, February 27 (AJP) - A line of humanoid robots flips in unison. They lunge, pivot, brandish swords beside children in a synchronized kung fu routine. In the next sequence, they stagger theatrically, swaying through a stylized "drunken boxing" set — collapsing backward only to rebound with uncanny balance. The spectacle, staged by China's Unitree Robotics at the Spring Festival Gala, ricocheted across the globe. For many viewers it was entertainment. For roboticists elsewhere, it was a reckoning. In South Korea, some scientists watched with admiration; others, with a familiar pang. Decades of painstaking work still sit largely confined to laboratories — brilliant machines, but rarely public performers. Unitree Robotics' G1 models perform group martial arts at China's Temple of Heaven/ Courtesy of Unitree Robotics In a lab at Hanyang University, a different drama unfolds. A humanoid named Alice 4 stands tethered to a rear frame — a metal Pinocchio awaiting animation. With a light tap on a keyboard, the machine jolts to life. It runs in place, almost straining against its restraints, optical sensors fixed straight ahead — on its creator. Han Jae-kwon, professor of robotics at Hanyang University and chief technology officer of Aei Robot, watches without theatrical flourish. "It's not about the kung fu or the backflips," he said. "The essence of a humanoid robot is what it does for work. Replace dangerous, undesirable labor. Help address the population cliff." That, he insists, is the measure that matters. Han Jae-kwon, professor of robotics at Hanyang University and CTO of Aei Robot, poses for a photo beside his team's Alice 4 model at Hanyang University ERICA Campus, Feb. 23, 2026. AJP Han Jun-gu The long arc of Korean humanoids South Korea's humanoid lineage stretches back more than two decades, to an era when bipedal machines were symbols of national ambition. At the Korea Institute of Science and Technology (KIST), a towering metallic figure still hangs in the institute's history hall. Known as "Centaur," it never truly walked upright. Its lower half resembled a mechanical horse — four legs instead of two. Korea's first humanoid was, in effect, a compromise between aspiration and stability. "Centaur was a small project we undertook to understand intelligent machines. Not long after Japan's Honda showcased ASIMO, we were asked to build ubiquitous robot companions capable of performing multiple tasks," recalled You Bum-jae, principal research scientist and former head of humanoid development at KIST. Japan's unveiling of ASIMO had electrified the region. In response, two Korean institutions embarked on parallel paths. At the Korea Advanced Institute of Science and Technology (KAIST), Professor Oh Jun-ho and his team introduced HUBO in late 2004 — a full-scale biped capable of walking, grasping and limited speech. "HUBO is the beginning of that history, we believe," Han said. "There was a sense of national pride — we could do what only Japan had done." Graphics by AJP Song Ji-yoon In early 2005, KIST followed with MAHRU, billed as the world's first network-based humanoid. Standing 150 centimeters tall, MAHRU was designed for domestic assistance — a precursor to what today might be called "physical AI." "MAHRU was a network-based humanoid, capable of understanding vocal commands. It could walk to a microwave, open it, pick up a piece of toast, place it in the toaster, take it out and deliver it to its master," said Yoo, gently patting the robot's original plastic head. "We didn't have advanced AI back then — only recognition skills and programs to support it. But that's essentially how all humanoids aim to function even today: a body running light, a powerful computer supporting it through a network, now equipped with AI." You Bum-jae, principal research scientist and formal leader of humanoid development at KIST, explains Korea's history of humanoids to AJP, Feb. 26, 2026. AJP Han Jun-gu Yet progress proved incompatible with political cycles. As Yoo noted, humanoid development was "not capable of significant results within two-year government projects." By the 2010s, funding for large-scale bipedal programs was pared back. Then disaster intervened. The Fukushima nuclear accident in 2011 prompted the U.S. Defense Advanced Research Projects Agency to launch the DARPA Robotics Challenge in 2012. The competition was designed to push robots into hazardous environments no human should enter. By the finals in June 2015, 24 teams had qualified. Three were Korean — each fielding its own platform rather than relying on Boston Dynamics' Atlas. Team KAIST, led by Professor Oh, won in 44 minutes and 28 seconds. Han's team from Robotis and a team from Seoul National University also competed — both using robots Han had helped design. For a brief moment, Korea stood at the apex of disaster-response robotics. But the aftermath told a more complicated story. Aei Robot's M1 model imitates an engineer's 'double-peace' sign at Hanyang University ERICA Campus, Feb. 23, 2026. AJP Han Jun-gu In the United States, many participants flowed into emerging ventures such as Figure AI or Tesla's Optimus program. In Korea, talent dispersed. Rainbow Robotics, spun off from Team KAIST, pivoted to collaborative industrial arms and was later acquired by Samsung Electronics. Robotis listed publicly and diversified. "The people who competed back then — it's such a waste," Han said. "Many went to the U.S., others became professors but stopped working on humanoids. If all of them had stayed, the situation would be very different today." Research continued, though largely out of public view. A five-year project funded by the Ministry of Trade, Industry and Energy sustained further development at KAIST, including hydraulically actuated humanoids capable of dynamic motion. "But if the media doesn't cover much of it, the public simply doesn't notice," said Park Hae-won, who now leads KAIST's humanoid lab. The motor that changed the race Ironically, the decisive shift did not originate in humanoids at all. At the Massachusetts Institute of Technology, Professor Kim Sang-bae's MIT Cheetah project advanced quasi-direct drive (QDD) motor technology — compact, high-torque electric actuators that allowed quadruped robots to run with unprecedented speed and efficiency. Crucially, the designs were open-sourced. Graphics by AJP Song Ji-yoon Hydraulics, long dominant in high-performance robots, were heavy and maintenance-intensive. Electric QDD systems offered lighter frames and lower costs — prerequisites for commercialization. Korean researchers argue that China moved quickly to absorb this architecture and adapt it to bipedal platforms. "China absorbed the QDD technology very quickly, establishing firm baseline requirements for humanoid development. That includes critical reinforcement learning for robots as well — we now believe the government distributes it to Chinese robotics firms," said Yoo. "That's how so many Chinese humanoids can run from the start, while ours begin with baby steps. We can't share reinforcement learning in Korea — institutions and private firms alike refuse to give up hard-earned data." The divergence, in this telling, is less about talent than about scale and coordination. AI as accelerant The recent resurgence of humanoid ambition in Korea owes much to artificial intelligence. Engineers check on KIST and LG Electronics' joint project KAPEX/ Courtesy of KIST When OpenAI signaled investment interest in Figure AI, the message was clear: large language models might finally supply the cognitive layer humanoids had long lacked. "Bipedal robots could perform tasks, but they required heavy engineering and years of coding. Imagine AI guiding them — robots understanding whatever their operators say and handling tasks without tedious step-by-step instructions. It was sensational," Yoo said. At KAIST, Park's team is assembling a fully domestically developed humanoid under MOTIE funding, targeting full integration by April 2026. At KIST, a joint effort with LG Electronics is producing KAPEX, described as Korea's first AI humanoid platform, with more than 70 degrees of freedom and predominantly domestic actuators. Yet laboratory elegance is not factory reliability. "What you see at trade shows is the most refined version," Park said. "If a robot falls over at the exhibition, imagine how many times it crashed in the lab." The initial commercial target, Park suggests, is not spectacle but small and medium-sized factories — cramped, uneven spaces where wheeled automation struggles. "If you visit Korean SME factories, the floors are uneven, spaces are narrow — wheels can't even get through," he said. "Humanoids could help there, if they can handle tasks without disrupting existing workflows." Park Hae-won, professor of robotics at KAIST's Dynamic Robot Control and Design Laboratory, poses for a photo among HUBO models/ Courtesy of KAIST Han is candid about Korea's relative position. "This year's Chinese Spring Festival Gala — robots were doing gymnastics, flipping like athletes," he said. "Our spirits sink. But the essence of a humanoid is not dancing or kung fu. The real question is: what are you going to do with it?" He argues that Korea's strength lies in manufacturing depth — batteries, semiconductors, precision motors and bearings — the physical half of "physical AI." Equally important is data. "Which country has industrial complexes in every neighborhood?" Han said. "That's all data. If we digitize it quickly and feed it to our robots, ours will outperform the competition." China's scale and cost advantages loom large. The United States is reviewing robotics imports under a Section 232 national security investigation. Korea is unlikely to erect similar barriers. "The only option is to make them cheaper than China," Han said. More pressing than tariffs, however, is demonstration. Korea spends substantially on humanoid research, Han noted, but allocates only a fraction of that to large-scale deployment trials. "What's needed now is getting existing robots into factories, gathering data and proving they work," he said. "That's demonstration projects — not more lab R&D." Yoo tempers expectations. "It's going to take at least five years for robots to be genuinely useful. They can only perform simple tasks like moving items, and even that carries a 10 percent margin of error. Humanoid-tailored AI has yet to arrive — the global race is now on," he said. "We should stop comparing robots against one another and instead draw on each one's unique characteristics to work as a team, talents putting heads together. Of course, that would require a new breed of engineers — ones who understand both software AI and hardware robotics. Attention, money, time — we need all of it." Korea's humanoid story — from HUBO and MAHRU to Alice and KAPEX — is neither triumphalist nor moribund. It is incremental, intermittently brilliant, frequently underfunded. Aei Robot engineers undergo maintenance on the company's humanoids at Hanyang University ERICA Campus, Feb. 23, 2026. AJP Han Jun-gu Whether it can contend with American venture capital and Chinese industrial scale will define the next chapter. Han's closing line carries neither romance nor despair. "Try running in a factory," he said. "They'll tell you to stop." Robots, he implies, must earn their keep the same way humans do — not by acrobatics, but by utility. For steel and silicon, the future will not be written on a gala stage. It will be decided on the factory floor.
• The birth of humanoid robots (1) Robo Sapiens: EVs with brains Graphics by AJP Song Ji-yoon Editor's Note: This is the first installment in AJP's series on humanoid robotics, examining the anatomy, technologies and economic logic behind one of the most hyped industries of the decade. SEOUL, January 20 (AJP) - Humanoid robots have become a buzzword at CES and beyond, captivating both technology and stock markets as artificial intelligence converges with China's formidable manufacturing power. This emerging breed of "robo sapiens" joining humanity — at least in its first generation — is, in essence, a two-legged electric vehicle with a brain. That reframing is not rhetorical. As humanoids move from laboratories toward commercialization, engineers, investors and policymakers are increasingly abandoning the idea of robots as walking computers and instead treating them as mechanical systems: complex assemblies of motors, joints and power units governed by AI. This shift helps explain why automakers have emerged as the industry's most promising producers — and why Hyundai Motor Group's humanoid robot Atlas captured the Best Robot title at CES 2026. "When you break down a humanoid by hardware, you have the head and neck, the torso, the upper body including the arm system, and then the hands at the end," said Park Dong-il, director of the Advanced Robotics Research Center at the Korea Institute of Machinery and Materials. "It's essentially the same anatomy as a human, and each part requires its own actuators, sensors and control mechanisms." The EV connection Graphics by AJP Song Ji-yoon The overlap between humanoid robots and electric vehicles is not merely conceptual. It is physical, financial and increasingly strategic. Both industries depend on high-density lithium-ion batteries, precision actuators and advanced motor technology. Both face supply-chain constraints around rare-earth elements required for permanent magnets. And both are racing to reduce costs through economies of scale. Hyundai Motor Group, which acquired Boston Dynamics for $880 million in 2021, plans to have affiliate Hyundai Mobis supply actuators for the Atlas humanoid robot. The group aims to produce 30,000 robots annually by 2028, leveraging the same manufacturing infrastructure and supplier networks that underpin its vehicle business. Tesla, meanwhile, has promised to unveil the third generation of its Optimus humanoid robot early this year. Elon Musk has suggested the bipedal robot could even become "an incredible surgeon." "Imagine if everyone had access to an incredible surgeon," Musk said. "Of course, we need to make sure Optimus is safe and everything. But I do think we're headed for a world of sustainable abundance." Industry experts identify Hyundai, Tesla and China's XPeng Motors as the current frontrunners among automakers entering robotics. While Xiaomi, BYD and Li Auto have also announced humanoid ambitions — often using autonomous driving data to train AI models — the top three are seen as the only players capable of near-complete in-house robot production. "Google and Nvidia chose Boston Dynamics as a partner for a reason," said Yim Eun-young, an analyst at Samsung Securities. "Hyundai's factories generate real-world behavioral data, continuous datasets that adapt to changing environments, and actual production and logistics sites where robots can be deployed and validated." "Most other robotics firms are startups," she added, "and they lack the infrastructure to accumulate large-scale behavioral data or test machines in real-world conditions." Tesla's Optimus serving popcorn during a presentation in Berlin, Germany, Dec. 20, 2025. AP-Yonhap A definition still in flux Despite more than half a century of development, no international standard defines what constitutes a humanoid robot. Nvidia, whose processors power the vast majority of humanoids currently under development, offers a working definition: "Humanoids are general-purpose, bipedal robots modeled after the human form factor and designed to work alongside humans to augment productivity." Yet in practice, the term encompasses machines with widely divergent designs — from full bipedal bodies to human-like torsos mounted on wheeled bases — many of which do not neatly fit Nvidia's description. "The definition of 'humanoid' itself is still unclear," Park said. "There's no ISO standard. We call robots with only an upper body humanoids, and we also call robots with both upper and lower bodies humanoids." The ambiguity extends further. Must a humanoid have five-fingered hands? Boston Dynamics' production Atlas uses a three-fingered gripper. Must it walk on two legs? Several robots showcased at CES 2026, including LG Electronics' CLOiD, used wheeled bases. And what about the face? From Unitree's G1 to UnixAI's Wanda series, most humanoids opt to remain faceless. Yet companies such as Realbotix and Engineered Arts argue that facial expressions are essential for natural human-robot interaction. According to discussions at the 2025 Humanoids Summit in London, industry groups are now debating whether the term "humanoid" should be replaced altogether with classifications based on capability rather than appearance. Two ways to map the machine Boston Dynamics' Atlas prototype appearing during a CES 2026 presentation in Las Vegas, U.S., Jan. 5, 2026. UPI-Yonhap Korean engineers and social scientists have proposed different frameworks for understanding humanoid technology. The mechanical approach breaks humanoids into physical subsystems — head and neck, torso, arms, hands, waist, legs and feet — each requiring dedicated actuators, sensors and control systems. The logic mirrors how an EV is analyzed through its battery pack, motor, inverter and chassis. Lee Jun-yong, a senior researcher at the Korea Planning & Evaluation Institute of Industrial Technology (KEIT), adopted a different perspective when authoring a government report on humanoid R&D in February 2025. His team consulted futurists, economists and social scientists to envision how humanoids might integrate into society by 2040, then worked backward to identify the technologies required to make those scenarios viable. The framework identified 10 core technologies across four domains: motion control, sensing, human-robot interaction and drive/control systems. "We started by envisioning future society and then identified the technologies needed to reach it," Lee said. "Deciding whether the government should directly lead humanoid development or support private-sector R&D is complicated, because robotics spans so many industries. Household robots that truly help people are still far off, which is why we felt a top-down approach was essential." The market acceleration Graphics by AJP Song Ji-yoon Goldman Sachs Research projected in 2024 that the global humanoid robot market could reach $38 billion by 2035 — more than six times its previous estimate of $6 billion. Manufacturing costs have already fallen by about 40 percent in a single year. Elon Musk offered a far bolder forecast at the Future Investment Initiative conference in Saudi Arabia in October 2024, predicting 10 billion humanoid robots by 2040 at prices between $20,000 and $25,000 each. Many robotics researchers remain skeptical, noting that a single universally capable robot is unlikely within that time frame. Cost remains a major constraint. Investment banks including J.P. Morgan estimate that Boston Dynamics' Atlas will cost at least $130,000 per unit when mass production begins around 2030. The prototype currently costs about $300,000 to build, with the commercial target set at less than half that figure. That remains five to six times higher than Tesla's stated goal of producing Optimus robots at $20,000 to $30,000 each. Still, the industrial logic is taking hold. If humanoid robots are indeed EVs with brains, then the automotive playbook — mass production, supply-chain integration and incremental cost reduction — applies directly. Park cautioned that the field is evolving too rapidly for fixed definitions or confident forecasts. Unitree's humanoid runs on a 400 meter track at the inaugural World Humanoid Robot Games held in Beijing, China, Aug. 15, 2025. Reuters-Yonhap "Research produced a year ago may already be outdated," he said. "The technology changes almost daily. What counts as state-of-the-art depends entirely on who you ask." For now, the working definition remains broad: machines with human-like form designed to operate in environments built for people. Whether that form requires legs, fingers or a recognizable face may ultimately be decided not by standards committees, but by the factories — automotive and otherwise — that build them. Lee's KEIT report ends with a provocative question: will humanoids one day demand labor rights as they replace human workers with tireless, uninterrupted productivity? "Our futurists raised the possibility that robots could eventually replace low-cost labor entirely," Lee said. "If overworked, they might malfunction — or even terminate their own digital existence. Who knows what the future holds?"
• The birth of humanoid robots (2): The Tin Man goes to work Graphics by AJP Song Ji-yoon Editor's Note: This is the second installment in AJP's series on humanoid robotics, examining the anatomy, technologies and economic logic behind one of the most hyped industries of the decade. SEOUL, January 27 (AJP) - A sudden jolt. A shudder. A mechanical doll's waist folds backward halfway, clanking onto the ground. Another teeters from side to side on a soft-soil running track before crashing headfirst, its legs flailing in a motion reminiscent of rigor mortis. Steel Pinocchios were not born overnight. The Defense Advanced Research Projects Agency (DARPA) Robotics Challenge, held from 2012 to 2015, often looked like child's play — engineers cheering on metal constructs of all shapes and sizes as they struggled through tasks that even toddlers might find trivial. Members of team Valor cheer as ESCHER completes the first challenge during day two of the DARPA Robotics Challenge at California, June 6, 2015. Courtesy of the Office of Naval Research Yet none of it was wasted. Every misstep mattered, each failure inching the field closer to semi-autonomous ground robots capable of operating in hazardous environments. "Many participants in the DARPA challenge went on to join Figure AI or help develop Tesla's Optimus humanoid series," said Park Il-woo, project director at the Department of Machinery, Robotics & Equipment at the Korea Planning & Evaluation Institute of Industrial Technology (KEIT). "Even those who failed continued their research, making the challenge a critical milestone in the advancement of bipedal robots." Early models struggled to open doors or walk in a straight line. Today's humanoids are ready to clock in. The Tokyo pioneer The WABOT-2 of Waseda University plays a digital instrument/ Courtesy of Waseda University The first humanoid to step into the world emerged from Waseda University in Tokyo: WABOT-1. The project began in 1970, when four laboratories within Waseda's School of Science and Engineering joined forces to build an anthropomorphic intelligent robot. In 1973, they unveiled the world's first full-scale humanoid robot capable of quasi-dynamic walking, speaking Japanese, and grasping objects. "The movements of early humanoids were, of course, awkward and crude compared with today's cutting-edge technology, but there is no doubt that this was an unprecedented and innovative challenge," said Atsuo Takanishi, director of Waseda University's Humanoid Robotics Institute, in a university news release. Honda Motor entered the race in 1986, spending a decade developing its E-series experimental models before unveiling P2 in December 1996 — the first self-regulating, two-legged robot capable of autonomous walking. That work culminated in ASIMO, introduced in November 2000. Honda's ASIMO gives directions/ Courtesy of Honda Global From MIT to the battlefield While Japan led the early race, a different approach was taking shape across the Pacific. In 1992, roboticist Marc Raibert spun off Boston Dynamics from the Massachusetts Institute of Technology. The company first made its name with quadrupeds. BigDog, developed in 2005, was designed as a robotic pack mule for soldiers. The military application proved short-lived — the U.S. Marines deemed it too loud for combat — but the company's mastery of dynamic balance would prove invaluable. Boston Dynamics' BigDog walking along sealines in Thailand/Courtesy of Boston Dynamics Fukushima's wake-up call The March 2011 Tohoku earthquake and tsunami exposed a painful truth. The disaster crippled the Fukushima Daiichi nuclear power plant, triggering meltdowns and radioactive leaks. Robots should have been the solution — radiation made human intervention deadly — but Japan's machines proved inadequate. DARPA responded by launching its most ambitious robotics program to date. The agency selected Boston Dynamics to build the humanoid platform: Atlas. The DARPA Robotics Challenge Finals, held in June 2015 in Pomona, California, became a spectacle of mechanical ambition and frequent failure. Twenty-three teams competed, their humanoids tasked with driving vehicles, opening doors, and climbing stairs. Graphics by AJP Song Ji-yoon South Korea's Team KAIST emerged victorious with DRC-HUBO, completing all eight tasks in 44 minutes and 28 seconds. Its secret lay in pragmatism: wheels mounted at the knees, allowing the robot to roll rather than walk when speed mattered most. China's rapid ascent While the United States, South Korea, and Japan traded blows in the humanoid arena, China quietly built its own robotics ecosystem. Only one Chinese team participated in the DARPA challenge. Ubtech Robotics, founded in Shenzhen in March 2012, became China's first humanoid robot company to go public when it listed on the Hong Kong Stock Exchange in December 2023. Its Walker series became China's first commercialized life-sized bipedal humanoid robots. By January 2026, Ubtech had produced its 1,000th Walker S2 unit at its Liuzhou factory. Ubtech Robotics' Walker S2 units align for a photo/ Courtesy of Ubtech Robotics Unitree Robotics, established in Hangzhou in 2016, entered the humanoid race in 2023 with the H1 — China's first full-size general-purpose humanoid robot capable of running. In August 2024, Unitree unveiled the G1 humanoid at a starting price of $16,000, sharply undercutting competitors. China now boasts more than 200 humanoid robot manufacturers, according to the China Machinery and Robotics Association. Cross-industry giants such as Xiaomi, XPeng Motors, and BYD have joined specialized startups in a three-pronged development push. From hydraulics to electrons Boston Dynamics continued refining Atlas after the DARPA challenge. The company changed hands several times — acquired by Google in 2013, sold to SoftBank in 2017, and purchased by Hyundai Motor Group for about $880 million in 2021. In April 2024, Boston Dynamics retired its hydraulic Atlas and unveiled a fully electric successor aimed at commercial use. At CES 2026, the company revealed the production version: 1.9 meters tall with a 2.3-meter reach, capable of lifting up to 90 kilograms, featuring 56 degrees of freedom and a four-hour battery life. Hyundai announced plans to produce 30,000 humanoid robots annually by 2028. Graphics by AJP Song Ji-yoon The race ahead The humanoid industry now stands at an inflection point. What began as academic curiosity in Tokyo laboratories and military experiments in American deserts has evolved into a three-way race among the United States, China, and South Korea — each betting billions that bipedal machines will reshape manufacturing, logistics, and eventually daily life. Hyundai's Atlas, Tesla's Optimus, and XPeng's Iron lead the automakers' push into robotics. Chinese startups such as Unitree and Ubtech are driving costs down at breakneck speed, while legacy players focus on precision and reliability. Hisense's Harley stands on display at the CES 2026 held in Las Vegas, Jan. 7, 2026. AP-Yonhap "In the early days, companies like Toyota and Honda aimed to deploy humanoids in industrial settings, but the programming burden was so immense that practical application remained out of reach," Park said. "Now, with generative AI making it far easier to train robots and teach them new motions, the United States and China — which led those methodological breakthroughs — are pulling ahead." The robots that stumbled and crashed at the DARPA challenge a decade ago are now walking factory floors. The question is no longer whether humanoids will work alongside humans, but how soon — and who will build them.