[Edaily Shin Yeong-bin Reporter] “The know-how possessed by skilled shipyard workers is tacit knowledge that is difficult to explain in words. Capturing this as data is a key direction for manufacturing physical AI.”

Lee Jun-ho, Chief Operating Officer (COO) of CrowdWorks, Inc.(355390), identified “manufacturing site data” as the core competitive advantage in the era of physical AI. He explained that physical AI should not be limited to competition over the performance of robots or humanoids themselves; rather, digitizing the work methods and decision-making contexts of skilled workers accumulated in actual industrial settings represents a more realistic opportunity.

CrowdWorks, Inc. is an AI data specialist company founded in 2017. With major clients such as Naver, LG Corp., and KTCorporation, it has been engaged in data labeling and training data construction. Last year, its consolidated revenue was approximately 10.5 billion won, down 12.1% year-over-year, and it recorded an operating loss of approximately 13.4 billion won.

CrowdWorks, Inc. recently established the Physical AI Data Lab and is now fully ramping up its efforts in building robot behavior data, egocentric data, and manufacturing site data. As part of the “Physical AI Leading Technology Development” project led by the Ministry of Science and ICT and the Institute for Information & Communications Technology Planning & Evaluation (IITP), the company was tasked with building a multimodal data pipeline for training the Physical AI World Foundation Model.

Lee, the COO, cited the “lack of source data” as the biggest difference between existing AI data and Physical AI data. Large Language Models (LLMs) and vision AI typically involve adding labeling and refinement to existing source data such as text and images. In contrast, Physical AI requires the creation of new behavioral data that captures how robots perceive, move, and make decisions in real physical environments.

“While LLMs and vision AI had source data, Physical AI requires behavioral data,” he said. “Because it involves creating something from nothing, there is a significant technical difference from existing data-building methods.”

This is why CrowdWorks, Inc. named its Physical AI Data Lab a “Data Lab” rather than a “Robot Lab.” It signifies that the organization is not focused on developing robots themselves, but rather on researching how to create high-quality data for robots and AI models to learn from. The COO explained, “The Data Lab is an organization dedicated to research and development on how to create the data necessary for effective learning.”

Teleoperation was the initial physical AI data collection method that garnered attention. This method involves a person remotely operating a robot while simultaneously collecting the robot’s visual, sensor, and behavioral data. However, CrowdWorks, Inc. determined that teleoperation alone had limitations in securing large-scale data. Operator fatigue was high, and technical variables and the difficulty of data verification were significant challenges.

CrowdWorks, Inc. is now focusing on egocentric and exocentric data. Egocentric data is collected from a first-person perspective—similar to the robot’s viewpoint—by attaching cameras to the operator’s head or body. However, first-person video alone makes it difficult to determine elbow angles, posture, and the operator’s full-body movements. To compensate for this, exocentric data from a third-person perspective is collected alongside it.

The COO stated, “People tend to think that robots can learn from any video, but in reality, data from the robot’s point of view is crucial,” adding, “Information on joint angles and posture—which is insufficient with ego data alone—can be supplemented by collecting exo data in conjunction with it.”

CrowdWorks, Inc. is expanding this technology to manufacturing sites such as shipyards. In its shipbuilding AI project, the company is not only digitizing blueprints and documents but also working to convert data generated on-site—such as from cranes and welding operations—into AI training assets. Large shipyard cranes represent a prime application area for Physical AI, where operator controls, signals, surrounding safety conditions, and sensor data all interplay in a complex manner.

In particular, COO Lee emphasized the effort to digitize the know-how of skilled welders at shipyards. While skilled welders are retiring in succession at domestic shipyards, the influx of young workers is limited. On-site reliance on foreign workers is also increasing. There is significant concern that the work methods and judgment criteria acquired through years of experience by these skilled workers may be lost in the process.

The COO explained, “To digitize the know-how of skilled welders, we need an environment that automatically collects sensor data generated during the work process.” He added, “However, sensor data alone makes it difficult to understand why they worked in a certain way, so we need a system that also digitizes the context of the work and the reasoning behind their decisions.”

He distanced himself from the perspective that views Physical AI as centered on humanoid robots. Rather than an approach that simply replaces human workers with robots, he argued that we need to use robots and AI to redesign the entire manufacturing process more efficiently.

The COO said, “Attempting to simply replace humans with robots is likely to fail,” adding, “We need to focus on creating efficient systems and structures using robots.” He went on to emphasize, “For South Korea to become a leader in Physical AI, the ability to convert tacit knowledge from manufacturing sites into data assets is crucial.”