"We plan to deploy robots to worksites faster than anyone in the world and support productivity improvements through efficient robot operations," said Park Sang-yeop, Chief Technology Officer of LG CNS.
Park made the remarks in a recent interview with Seoul Economic Daily at the company's Magok headquarters. LG CNS is focusing on developing integrated robot control, management, and operation platforms, along with smart factory, logistics, and city solutions. The company is also actively collaborating with robot hardware manufacturers and firms developing Robot Foundation Models (RFM), which serve as the brain of robots.
"Current robot technology allows robots to lift standardized boxes from conveyor belts and adjust the position of boxes with barcode stickers to align with scanners," Park said. "In about two years, robots will be able to handle a significant portion of manufacturing processes."
Park's outlook is underpinned by advances in artificial intelligence and agentic AI technologies that enable robots to respond to various situations. The key lies in data training. Park identified "imitation learning" as the fundamental method for robots to acquire and learn task-related data. In this approach, robots move according to human-operated remote controls while continuously collecting motion data. AI can then significantly expand this learned data. The more data available, the more accurately and precisely robots can move.
World Foundation Models represent another major method for robot training. This involves creating virtual environments that replicate actual worksites where robots are deployed for learning. Through repeated training in virtual spaces, robots develop the ability to predict what will happen, enabling them to respond to unexpected situations. However, minimizing the gap between simulation and reality remains a challenge. Other methods include extracting data from videos of human workers or attaching sensors to workers' bodies to collect data for robot training. Pre-trained robots then undergo fine-tuning to adapt to specific manufacturing environments.
Even after completing this process, robots cannot immediately be deployed to worksites. Having excellent textbooks does not guarantee good grades. Robots also suffer from "catastrophic forgetting." Park explained, "A robot that learned 10 things might learn 2 more, but conflicts between existing and new knowledge can cause it to forget what it learned. Even a single mistake during training can sharply reduce accuracy." This is why data curation—accurately identifying data that will help robot training—is as important as collecting large amounts of data.
Even with appropriate training data and suitable hardware, commercializing robots in manufacturing processes is not straightforward. When deploying multiple robots in a process, designing systems that prevent collisions is essential. Systems must also consider battery charging times for optimal robot operation and predict potential malfunctions. As all these factors must be considered when designing new factories, demand for related consulting services is growing alongside robot commercialization.
"LG CNS has extensive experience with smart logistics control systems," Park said. "We can build robot control systems better than anyone."
LG CNS has successfully built automated robot control systems for power infrastructure companies. The company deployed third-generation robots and introduced logistics, production, and facility control systems to boost productivity.
"Return on investment depends on how you design the system," Park said. "Robots can dramatically reduce discomfort in high-voltage electrical work with electrocution risks or chemical processes requiring masks and protective suits."
He added, "Ultimately, humans and robots will have to collaborate. Over time, the proportion of work done by robots will continue to increase."