Technology that once moved game characters is moving into physical robots

NC’s AI research did not begin on a factory floor. In 2011, NCSOFT created what it describes as the first dedicated AI organization in Korea’s game industry. In 2016, Blade & Soul’s Tower of Infinity introduced reinforcement-learning combat AI that learned from battle situations and selected actions in response to an opponent’s movement.

In February 2025, the research organization spun off as the independent company NC AI. By 2026, the company’s public technology story reaches well beyond games into digital twins, world models, robot foundation models, manufacturing, shipbuilding, defense and humanoids.

On September 17, NC AI was also reported to have joined tasks in both Jeonbuk and Gyeongnam under a national Physical AI R&D program. The total program is worth KRW 1.4131 trillion over 2026–2030. That number is the scale of the overall regional program, not a contract awarded to NC AI alone.

Remove the industry labels and the problem structures begin to overlap

Game AI needs to understand the state of a virtual environment, observe other actors, anticipate what happens next and choose an action. It also creates 3D spaces, represents human motion as data and repeatedly tests decisions and behavior in simulation.

Physical AI operates under much harder real-world constraints, but some of the underlying problem structures are familiar: use cameras and sensors to understand a scene, recognize human and machine movement, predict future states and determine a robot’s next action. Digital twins and simulation can be used before real-world deployment.

This does not make game technology identical to manufacturing technology. Safety, sensor error, mechanical constraints, quality, throughput and failure cost change the problem materially. The point is narrower: some problem-solving capabilities can remain valuable even after the industry label changes.

Hanwha Ocean, POSCO DX, Hyundai Rotem and LG Electronics show where the capabilities are going

With Hanwha Ocean, NC AI is developing a robot foundation model for autonomous welding in shipbuilding. The task requires visual-language intelligence to interpret welding conditions under intense arc light and dust, then connect perception to precise robot motion.

With POSCO DX, the company is co-developing a general robot foundation model for humanoids to understand factory environments and act on assigned tasks. With Hyundai Rotem, NC AI is working on world-model technology within a digital-twin simulation project for integrated operation of multiple unmanned robots in future battlefield environments.

In a government humanoid project with LG Electronics, NC AI is responsible for initial training data and a Human-to-Robot Retargeting pipeline that transfers human motion to a robot while accounting for joint limits, body proportions, balance and other physical constraints.

Banseog calls this Capability Migration

By industry classification, NC began in games, Hanwha Ocean is shipbuilding, POSCO DX is industrial IT and automation, and Hyundai Rotem spans rail and defense. By company name, these can look like separate talent markets.

By capability, different connections appear. 3D environments can connect to digital twins. Character motion can connect to Human-to-Robot Retargeting. Reinforcement learning can connect to robot action learning. Simulation can connect to Sim-to-Real training. Vision, language and world modeling can reappear in factory understanding and physical-state prediction.

Banseog calls this Capability Migration: the value of a capability is not fixed by the industry in which it was first built. What matters is whether the same underlying problem-solving machinery becomes useful against a new bottleneck.

The Physical AI talent pool may extend well beyond robotics companies

If a Physical AI organization needs 3D, simulation, motion, vision, reinforcement learning, multimodal systems, world models, controls and manufacturing integration, some of those capabilities may have been developed for years outside robotics.

Games, VFX, autonomous driving, industrial simulation, digital twins, computer graphics and defense simulation can all contain relevant capability pockets. A search restricted to ‘five years in robotics’ can therefore miss people whose problem-solving experience is closer than their industry label suggests.

That does not mean a game developer automatically becomes a robotics engineer. Robot safety, controls, mechanical constraints, factory operations, process engineering and real-time sensor integration may still need to be learned. The useful task is to separate what transfers from what does not.

BANSEOG VIEW | The industry that created a capability does not have to be the industry that ultimately values it most

NC’s AI research began inside games in 2011. By 2026, the company is applying related capabilities to shipyard welding, manufacturing humanoids, defense simulation, humanoid motion and Physical AI projects tied to manufacturing and semiconductor processes.

That is a larger story than ‘a game company moved into robotics.’ When a capability developed in one market meets a bottleneck in another, the value of the technology — and the map of companies competing for the people behind it — can both change.

The next talent question should not stop at ‘Which industry did this person come from?’ It should also ask: ‘Which problems has this person repeatedly solved, and where might those capabilities become more valuable next?’

Physical AI requires tracking capability paths, not just job titles

Banseog Physical AI does not stop at company lists. It tracks real job postings and technology changes to see which roles are growing, which skills are being added and which experiences from adjacent industries are beginning to connect.

That is why Apptronik’s humanoid hiring can surface MES and Kubernetes, while Samsung SDS can make orchestration across robots and enterprise systems more important than the robot itself. Physical AI talent markets are difficult to understand through robotics labels alone.

Career Radar follows the same logic: instead of reading only a current job title, it looks for where existing capabilities sit closest to emerging roles. As Capability Migration accelerates, the next destination of a skill may matter as much as the title attached to it today.

BANSEOG VIEW

Banseog View — Capability Migration

NC’s AI work began in a game-industry research organization in 2011, accumulating reinforcement learning, 3D, motion, simulation and vision capabilities before expanding into shipbuilding, manufacturing, defense, humanoids and semiconductor-related Physical AI work.

An industry change does not transfer every skill automatically. But capabilities built around overlapping problem structures can gain new value in a different market.

Physical AI recruiting should therefore ask not only whether someone comes from robotics, but what problems they have repeatedly solved and where those capabilities overlap with a new industrial bottleneck.

SOURCES

Primary sources and references

KRW 1.4131 trillion is the total size of the Jeonbuk–Gyeongnam Physical AI R&D program for 2026–2030, not NC AI’s contract value. The Hanwha Ocean, POSCO DX, Hyundai Rotem and LG Electronics examples are separate publicly disclosed collaborations or projects and should not all be read as subprojects of the Sep 17 regional program. Game AI technology is not treated as identical to manufacturing or robotics technology. ‘Capability Migration’ is Banseog analysis based on the disclosed technology history and use cases.