Krafton, the Korean publisher of PUBG, launched a joint world-model research programme with Seoul National University’s Interdisciplinary Program in Artificial Intelligence on October 8, 2026. Over three years the partners will run three research projects and a joint research centre, aimed at turning the virtual-world techniques Krafton has built for games into AI that predicts how the real world moves. The budget was not disclosed.

A world model is an AI system that learns how an environment behaves so it can predict what happens next after an action. It is one of the main research bets behind robot learning, because a good one lets a robot practise in a learned simulation rather than in the physical world.

Key Facts

  • Launched October 8, 2026 with SNU’s graduate Interdisciplinary Program in Artificial Intelligence; three years, three projects plus a joint centre
  • Projects: consistent long-video world models, game motion data for human–object movement, and testing whether AI uses spatial information
  • Krafton CAIO Lee Kang-wook and SNU professor Kang U attended the launch; funding not disclosed
  • Krafton set up Ludo Robotics in the US in late 2025 and a Korean entity in February 2026

What will the projects study?

According to MTN’s report on the launch, the first project is a world model that keeps spatial structure and the appearance of objects consistent over long generated video, a known weakness of video generators. The second uses motion data from games to connect the movements of people and nearby objects naturally. The third tests whether an AI actually uses spatial information, such as the width of a passage or the distance to an object, as the basis for its judgments in virtual scenes.

The third project is the most relevant to robots. A video model can produce plausible frames without representing geometry; a robot planning a path through a doorway cannot. Checking whether a model’s decisions rest on measured space is an evaluation problem that robot-model builders have often handled with task success rates instead. The partners also plan intern programmes, staff exchanges and a corporate research centre inside the graduate programme.

“The simulation and virtual-world building capability Krafton has accumulated through games is an important foundation for world-model research,” said Lee Kang-wook, Krafton’s chief AI officer.

How does this fit Krafton’s robotics push?

It is the latest of several moves. Krafton set up the robotics research company Ludo Robotics in the US in late 2025 and a Korean entity in February 2026, signed an agreement with Hanwha Aerospace in March 2026 to develop physical AI and form a joint venture, and in April made a ₩65 billion strategic investment in the car-sharing company Socar along with an investment in a new autonomous-driving unit. None of these has produced a product yet. World models are also where larger players are spending, from AMD’s purchase of World Labs to Runway’s Praxis-1; we explain the bet in our world-models explainer.

Frequently Asked

What did Krafton and Seoul National University agree?

A three-year joint research programme on world models, launched on October 8, 2026, with three research projects and a joint research centre at SNU’s Interdisciplinary Program in Artificial Intelligence.

What is a world model?

An AI system that learns how an environment behaves so it can predict what happens next after an action. Robot developers use world models to train and test behaviour in simulation.

What else has Krafton done in robotics?

It set up Ludo Robotics in the US in late 2025 and in Korea in February 2026, signed a physical AI agreement with Hanwha Aerospace in March 2026, and invested ₩65 billion in Socar in April 2026.

Sources & Further Reading