Humanoid robots clock in at Korean warehouses

Before a humanoid robot made by Figure AI grabbed global attention for sorting and packaging parcels at near-human speed, a similar humanoid worker had already been put to work at a logistics center run by CJ Logistics in Gunpo, Gyeonggi.

As boxes streamed down the conveyor and human workers looked on with a mix of concern and curiosity, the machine went to work — pulling sheets of packing paper from a roll-like dispenser, lifting each passing box, and carefully stuffing it before returning it to the line. Within moments, murmurs of approval spread across the floor.

A semi-humanoid machine developed by Korean robotics firm Robotis features a humanlike upper body while its lower body moves on wheels. Starting in September last year, it underwent a three-month on-site pilot at CJ’s Gunpo fulfillment center in Gyeonggi, marking the logistics industry’s first real-world test of a humanoid robot.

The trial made clear both the promise and the limits of the technology. At roughly 150 million won ($99,700) per unit — including hardware, software development and installation — the robot was expected to justify its cost through performance and agility.

But speed was the first issue. Over three months, it operated at just 30 percent to 50 percent of the pace of an experienced human worker, a notable shortfall.

Accuracy offered a more encouraging picture: The robot processed about 130 boxes an hour on average, working methodically and with relatively few errors.

There were, however, occasional missteps. On one occasion, it dropped cushioning material onto the floor mid-task, which then had to be picked up by a nearby human worker. But observers on site were pragmatic. “Like humans, robots make mistakes,” one said. “This is a level that can be improved.”

The result was a cautious green light. After reviewing the pilot, CJ Logistics purchased two additional units on top of the original test robot, betting that steady iteration — not immediate perfection — would determine the technology’s value.

The targets are pragmatic: Deploy three robots by the second half of the year, raise productivity to roughly 70 percent of a human worker’s pace and achieve 99 percent operational accuracy.

Slow but steady

Upon adoption, CJ Logistics expected the total cost of a humanoid robot, including development expenses, at around 150 million won per unit, according to Koo Seong-yong, head of the automation technology department at CJ Logistics and a KAIST-trained robotics Ph.D.

“Given that labor costs at logistics sites run about 50 million won per worker annually, the robot reaches break-even in roughly three years. That makes it well worth the investment,” said Koo.

The calculation rests on a key assumption: that the robots’ current pace — still roughly half that of experienced human workers — will improve meaningfully.

Matching human performance on a one-to-one basis isn’t necessary. Robots can operate continuously, pausing only for battery charging, which means total throughput could ultimately approach human levels even if individual task speeds remain lower. Ku said he expects the machines to reach about 70 percent of human speed within this year.

This confidence stems from experience. The robots were trained using Nvidia’s physical AI model for robotics, GR00T, while CJ Logistics built a proprietary application layer on top. In the process, the development team encountered results they say even they did not anticipate.

Ku said the company completed in-house software development to deploy robots in logistics operations in just three to four months, calling the short timeline a key achievement. “Traditional robot software development typically takes at least six months and can stretch up to three years,” he said. “But by using newer physical AI techniques, including simulation-based learning, we were able to significantly compress the development cycle.”

CJ Logistics has previously introduced automation systems such as palletizing robotic arms that stack boxes and autonomous mobile robots for internal transport. Its fulfillment centers are currently about 30 percent automated. However, tasks such as packaging and sorting items of varying sizes and materials — operations that run around the clock — still rely heavily on human labor.

Robots at the center of logistics and retail

Logistics and retail are widely seen as the sectors where humanoid robots can be deployed most easily and at the fastest pace. Market research firm Counterpoint Research estimates that 80,000 humanoid robots will be deployed in logistics by 2030, accounting for roughly 25 percent of global installations.

The structure of the industries makes the adoption more straightforward. Compared to manufacturing plants or residential environments, logistics operations tend to involve simpler layouts and more repetitive tasks. The technological threshold is lower, since the level of physical AI required is less demanding than in complex production processes.

Safety concerns are also less pronounced. Shim Won-sup, associate partner at McKinsey & Company, said safety remains one of the key bottlenecks to the wider adoption of humanoid robots. Rather than spreading simultaneously across multiple industries, he expects initial deployments to begin in relatively easier environments such as logistics and delivery, before expanding into other industrial settings and eventually households.

In practice, global logistics companies are already moving ahead. E-commerce and logistics giant Amazon had deployed more than 1 million robots in its operations as of 2025, including autonomous mobile robots and collaborative robots. The company aims to automate 75 percent of its logistics operations by 2030 using a combination of humanoid robots, autonomous robots and AI technologies.

Amazon’s first humanoid robot pilot involved “Digit,” a unit developed by U.S. robotics firm Agility Robotics. The company began testing the model in 2023 and is still evaluating whether to scale its deployment.

Designed specifically for logistics use, the humanoid robot stands 175 centimeters (5.7 feet) tall, weighs 63.5 kilograms (140 pounds) and is built to resemble the proportions of an adult male. It can lift up to 16 kilograms. Seven units of Digit were also deployed at a Toyota plant in Ontario, Canada, earlier this year.

China, which has been aggressively advancing its robotics industry, is also steadily unveiling logistics-focused humanoids. At the CeMAT Shanghai held in Shanghai in October last year, dozens of humanoid robots were showcased.

According to industry officials who attended the event, the contrast with the previous year was striking: While the 2024 exhibition largely featured conventional logistics robots such as autonomous mobile robots and palletizers, humanoids have since emerged as the dominant trend in just one year.

Fieldwork wins, even for robots

At the Robot Intelligence Research Lab at KwangWoon University, visited on Feb. 24, robots were undergoing what amounted to job training. The work is part of a government-backed initiative on humanoid robot demonstration support, carried out by Lotte Global Logistics, domestic robotics firm Robros and researchers at KwangWoon University, Sogang University and Kyung Hee University.

The research team is training the robot to operate an auto-bagger system, with the goal of deploying Robros’ Igris humanoid in logistics operations later this year. Researchers are using virtual reality equipment and wearable exoskeletons to train the robots via teleoperation. Under remote commands, the robots moved in near real time, with little to no lag or interruption.

In front of the auto-bagger system, the robot stood in fully humanlike form. With its 10 fingers, it carefully picked up a T-shirt and placed it into a packaging bag. It also naturally performed the sealing step, pressing buttons on both sides of the machine simultaneously. Workers on site carry out the task in the same way, due in part to concerns over crush-related accidents.

The robot has not yet been deployed in logistics operations, and there is no formal data on its speed or accuracy in real-world tasks. But compared to experienced workers on the floor, it appeared markedly slower.

Like the CJ Logistics pilot team, the researchers do not see an immediate need for robots to match human performance at 100 percent. Park Su-han, a professor in the department of robotics at KwangWoon University who leads the project, said speed can be improved through training over time. “Slowness at this stage is not a problem,” he said. “The real advantage is that robots can work 24 hours a day without fatigue.”

Park said there is a far more important challenge than speed.

“Ultimately, it comes down to how much data you can accumulate,” he said. “Large language models like ChatGPT have vast amounts of training data available everywhere, but that is not the case for robotics. Even with the use of simulation environments and reinforcement learning, real-world deployment data must still be accumulated for robots to take hold in industrial settings.”

Robots behind the counter

With unmanned convenience stores already common, the question arises: Is there any real need for stores where robots actively greet customers? Korea Seven, which operates the 7-Eleven franchise in Korea, is nonetheless planning to deploy humanoid robots in its future store concept.

Lotte Innovate Company began pilot testing in February, introducing a humanoid robot clerk at a “future 7-Eleven” store (AX Lab 3.0) located on the first floor of its Seoul headquarters. The project combines software development with the use of the G1 model from Chinese humanoid robotics firm Unitree Robotics. The robot has been given the name Roi.

Roi moves through the store like a staff member, guiding customers to products and engaging in light conversation. It is connected to Lotte Innovate’s AI platform “iMember,” enabling it to handle casual exchanges, including small talk about everyday topics such as the weather.

Unmanned convenience stores offer efficiency, but lack interaction or humanlike engagement. Korea Seven is instead envisioning a model that combines convenience with a sense of presence and communication, while also delegating store management functions to robots.

The company plans to further train Roi to eventually identify out-of-stock items and restock shelves, expanding its role inside the store.

Lotte Innovate, the systems integration arm of Lotte Group, is also exploring broader applications for humanoid robots across the group’s core retail and logistics businesses. As physical AI emerged as a key industry theme last year, the company consolidated previously dispersed internal teams into a humanoid robot task force, which has since been formalized into a permanent unit.

The humanoid robots currently being tested in convenience stores are two G1 model units. The company is also reviewing additional purchases from domestic and overseas suppliers.

Beyond the convenience store pilot, the G1 robots also underwent training at a stair-climbing challenge at the Lotte World Tower Sky Run held in April, where they prepared to tackle stair ascent as part of their capability development.

Jeon Bong-ji, head of the AI team at Lotte Innovate, said there are numerous potential use cases for humanoid robots across the broader Lotte Group portfolio.

“Before deploying them in manufacturing environments, we see opportunities in service settings such as convenience stores and logistics, as well as in entertainment venues like Lotte World,” he said. “We also believe robots could be used for patrol functions around the Lotte World Tower area.”

BGF Retail, which operates the CU convenience store chain, has also begun preparing for the introduction of humanoid robots. Its logistics subsidiary, BGF Logis, has signed a memorandum of understanding with robotics firm Robotis to jointly develop related technologies. The company also plans to deploy the AI worker model introduced by CJ Logistics in its logistics operations.

Convenience store logistics differ from traditional models. They are not like large-scale retail distribution, which typically moves palletized goods, nor are they like parcel delivery networks that handle everything from large appliances to small accessories.

Instead, they involve a high mix of small, fragmented items — ranging from chewing gum and ready-to-eat meals to tissue products and stationery — which must be broken down and redistributed to individual stores across a nationwide network.

As this requires a far more labor-intensive process, automation is comparatively difficult. Sorting goods by individual store has, until now, been handled manually.

BGF Logis plans to deploy five humanoid robots to conduct pilot tests and validate commercial viability, with the aim of introducing them into manual processes across 24 logistics centers nationwide after 2028.

A CU official said logistics operations are facing labor shortages. “We expect robot adoption to improve efficiency while also helping address these work force constraints,” he said.

A practical guide to humanoid robots

In a January report on the humanoid market, market research firm Gartner said expectations for humanoid robots are significantly outpacing their technological readiness, warning of what it described as a potential bubble. The firm added that wheeled or multifunctional robots are likely to prove more efficient than human-shaped machines.

The assessment broadly holds true in logistics and retail settings. Operators in these sectors are not looking for robots capable of backflips like Boston Dynamics’s Atlas, or kung fu demonstrations seen in some Chinese robot prototypes. Such capabilities are considered to be overengineered for the job.

While physical AI and robotics are clearly spreading across industrial settings, the outlook points to a mix of machines tailored to specific functions. Heavy lifting can be handled by robotic arms, while transport tasks are carried out by autonomous mobile robots. Logistics and retail companies broadly agree that they will begin by introducing robots into select manual tasks that are either hazardous or physically demanding.

Amazon, which has outlined plans to deploy 1 million robots, is also not solely focused on humanoid systems. In a December interview with AFP, Tai Brady, chief technologist at Amazon Robotics, said the starting point in robotics should be the problem to be solved.

Brady believes that the first question in robotics should be determining what is the problem to be solved, and only then should one think about function. From his perspective, form follows function, so starting with form means that the process is reversed.

Ryu Jung-hee, CEO of robotics firm Rlwrld, offered a more nuanced view of whether humanoid robots must resemble humans. “The answer is yes and no,” he said. He noted that while humanoid robot hands must achieve sufficient dexterity to perform human-level tasks, bipedal locomotion is not necessarily superior in all environments. “In terms of mobility, it is hard to argue that two-legged walking is optimal in every situation,” he said. “Humanoid robots may be necessary only in certain settings, such as climbing stairs or ladders.”

BY JANG JOO-YOUNG, KIM HYO-SEONG [[email protected]]