PUBG Maker Krafton Bets Its Gaming AI Can Power Real Humanoid Robots

Krafton, the South Korean studio behind PUBG: Battlegrounds, posted second-quarter revenue of 1.29 trillion won — about $890 million, up 95% year over year — and it's using that firepower to push video game AI out of virtual worlds and into physical robots.

The vehicle is Ludo Robotics, Krafton's Palo Alto subsidiary. Its experimental system, Ludi, runs on Unitree's G1, a roughly $16,000 humanoid, and combines existing AI models: a vision-language model from Alibaba's open-source Qwen family handles perception and decisions, while Nvidia's GR00T robotics foundation model, fine-tuned by Ludo, controls physical tasks like picking up objects.

CH Kim, CEO of both Krafton and Ludo Robotics, frames the pivot as a natural extension of two decades building game AI. "We've been developing AI in the form of nonplayer characters, or NPCs, for our video games even before the recent wave of interest in the [AI] sector," Kim told Fast Company. "We reached a point where we started thinking about what would happen if we brought these AI characters out of virtual worlds and into the real world. That's essentially physical AI."

Ludo Robotics is set to present Ludi 0.2, its latest experimental system, at the Bay Area Robotics Symposium at the University of California, Berkeley. The update improves conversational memory and gesture recognition, though Kim concedes Ludi remains a research demonstration — not reliable enough for everyday use.

What does PUBG's AI teammate teach a robot?

Krafton developed PUBG Ally with Nvidia, calling it "the world's first voice-enabled coplayable character (CPC)." During a limited beta in PUBG's arcade mode, players could bring along Ella, a companion responding to voice commands. The reception was mixed. Inven Global, citing Krafton's own report, put her combat performance at 2.58 out of 5. Half of respondents saw her as a tool; only 18.5% called her a teammate.

Still, Kim says the technology transfers. "The foundational technology we've applied to PUBG Ally is very similar to the technology we're applying to Ludi," he explained. "Our longer-term goal is to expand this work into developing our own proprietary foundation model."

Game engines give Krafton a practical edge: they simulate how objects move and interact, generating synthetic training data. Ludo trained its reasoning model on 14,432 computer-generated interaction sequences. Physical manipulation was learned separately, from recordings of people remotely controlling the robot in a mock household.

How does Ludo's stack measure up?

Ludo's Interaction Core28 benchmark tests AI across 28 simulated household scenarios — delivering objects, interpreting ambiguous requests, responding when someone speaks mid-task. Before fine-tuning, the smaller Qwen model completed 10 scenarios; after, 17. An OpenAI frontier model with an optimized prompt completed 26, exposing a substantial gap. All results come from simulation, which makes comparisons with physical robots difficult.

For context: Figure, which raised more than $1 billion last year at a $39 billion valuation, reported a 56% task-completion rate across 420 real-world home trials — without testing interrupted chores or shared spaces. Skild AI secured $1.4 billion in January for robot-control AI, and 1X sells its $20,000 NEO household humanoid with remote human backup.

What changes for the robotics stack?

Ludi 0.1 and 0.2 both rely on multiple foundation models wired through an agent harness — one reasons, others execute. That breaks down when circumstances change mid-task, says CTO Kangwook Lee, who is also Krafton's chief AI officer. "Each receives a final instruction to execute without understanding the broader context, which becomes a significant limitation when circumstances change," he explained. "This constant back-and-forth disrupts the connection between reasoning and physical execution."

Ludi 1.0, scheduled for 2027, aims to fix this with a unified foundation model that retains context and adapts without handing control back and forth. "If something changes in the environment, the robot should be able to build on its previous reasoning and adjust its actions much faster," Lee said.

Compute is the other constraint. Ludi 0.1 ran locally on a workstation beside the robot — no cloud — and Ludo picked a smaller, 4-billion-parameter model to cut delays. "One of the biggest challenges is developing those capabilities within the specifications available on the robot itself," Kim said.

What's the money and timeline?

Krafton's broader AI spending includes $70 million for a computing cluster and a reported $33 million in HyperAccel, a South Korean AI chip startup. Ludo's team includes chief scientist Rob Nowak and human-robot interaction lead Yea-Seul Kim, both from the University of Wisconsin-Madison.

Kim says Ludo focuses on research this year, then identifies a real-world deployment environment in 2027. "We need to find an environment where those research results can be applied and where we can potentially create a data flywheel or a training flywheel," he said. Krafton may eventually raise outside capital: "Robotics is a very long-term project. We may need to raise additional funds from outside the company to properly pursue this area."

For now, Ludo builds intelligence, not hardware — it plans to integrate its software into other manufacturers' robots. In a game, players laugh off an AI teammate's blunders; in a kitchen, the same mistake costs the human the very supervision they hoped to offload.