South Korean Startup Uses Worker Motion Data to Train AI for Humanoid Robots

The company is also collecting data from logistics workers at CJ, recording how they grip, lift and handle goods, and from staff at a Japanese convenience store chain, tracking how they organize food displays. RLWRLD's engineers have identified replicating the dexterity of human hands as a priority for developing humanoids that can function effectively in varied environments.

The startup last week unveiled its robotics foundation model, an AI system intended for use in robots. David Park, who has worked nine years at the Lotte Hotel Seoul, is one of about 10 food and beverage team members participating in the recording sessions.

He said he has been wired up with cameras about once a month. During one session he folded a banquet napkin into a tight layered square before wiping wine glasses, knives and forks while colleagues prepared for actual services nearby.

RLWRLD is part of a group of South Korean firms and manufacturers competing in the global market for physical AI, which refers to machines equipped with AI and sensors that can perceive, decide and act with some autonomy in real-world settings. The country is seeking to use its strengths in semiconductors and manufacturing to become an AI powerhouse.

The government last month announced a $33 million project to capture the skills of master technicians into a database for AI-powered manufacturing to boost productivity and address an aging workforce. Hyundai Motor plans to introduce humanoids from its Boston Dynamics unit at global factories starting with its Georgia plant in 2028.

Samsung Electronics aims to convert all manufacturing sites into AI-driven factories by 2030 using humanoids and task-specific robots. RLWRLD expects industrial AI robots to be deployed at scale around 2028.

Labor groups in South Korea have expressed concerns that widespread adoption of robots could reduce jobs for skilled workers. The Korean Confederation of Trade Unions has warned that robot deployments risk severing the pipeline for skilled labor. Its policy director said mastery of skills remains a human achievement and called for engagement with workers on AI implementation.

The startup's approach involves converting worker camera footage into machine-readable data, then having engineers repeat tasks while wearing cameras, virtual reality headsets and motion-tracking gloves. This data trains test robots, often guided by pilots using wearable devices.

The process records details such as joint angles and applied force. In a demonstration at a lab located in a suite at the Lotte Hotel, a wheeled robot with humanlike metal hands lifted and placed cups at a minibar, though it knocked over a dish at one point.

More advanced test footage shows a humanoid opening a box, placing a computer mouse inside, closing the box and setting it on a conveyor belt. RLWRLD is focusing on AI for five-fingered hands that mimic human touch, which could be particularly useful in homes requiring closer human interaction.

Lotte Hotel hopes robots will be capable of cleaning guest rooms and other behind-the-scenes tasks by 2029. The company also plans to offer robot rental services for hospitality and other industries, with possible future expansion to homes. Park said humanoids might eventually handle 30 to 40 percent of the workload in back-of-house event preparation areas but would have difficulty replacing roles involving direct human interaction.