The robotics industry wants humanoids to become part of everyday life, taking on work in homes and offices. Billions of dollars are flowing toward machines capable of increasingly complex physical tasks. The humanoid robot maker Figure raised more than $1 billion last year at a $39 billion valuation, while the robotics software startup Skild AI secured $1.4 billion in January to develop artificial intelligence that can control different kinds of robots.
But working alongside people requires understanding incomplete instructions and adapting when someone changes their mind midway through a task. A robot that needs constant explanations could leave its owner supervising the very work they hoped to hand over.
Krafton believes its video game experience could help solve that problem. The South Korean developer behind PUBG: Battlegrounds has spent nearly two decades creating virtual worlds where software responds to unpredictable player decisions. Its Palo Alto subsidiary, Ludo Robotics, is now applying that expertise to AI that helps robots understand human requests and decide how to act.
CH Kim, CEO of both companies, sees Krafton’s experience developing AI teammates as a starting point. Developed with the chipmaker Nvidia, PUBG Ally is what the companies describe as “the world’s first voice-enabled coplayable character (CPC),” an AI teammate that goes beyond traditional nonplayable characters (NPCs) by cooperating with players, adapting to gameplay and remembering previous interactions.
Ludo Robotics hopes to bring those capabilities into the physical world.
“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. With the latest large language models and AI technologies, we’ve been able to give our in-game NPCs greater intelligence,” Kim tells 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, and that’s how we came to work on Ludi.”
Ludo Robotics is set to present Ludi 0.2, its latest experimental robotic intelligence system, at the Bay Area Robotics Symposium at the University of California, Berkeley. The new version aims to improve conversational memory and gesture recognition, though Kim acknowledges that Ludi remains a research demonstration, not yet reliable enough for everyday use.
What PUBG Can Teach a Humanoid Robot
During a limited beta in PUBG: Battlegrounds’ arcade mode, players could bring along Ella, a PUBG Ally companion that responded to voice commands. In a recent blog, Krafton’s AI chief described players joking with her and treating some mistakes as entertainment. Yet her usefulness as a teammate was less convincing. The gaming news outlet Inven Global, citing Krafton’s PUBG Ally report, put combat performance at 2.58 out of 5. Half of respondents viewed the character as a tool, while just 18.5% considered her a teammate.
“The foundational technology we’ve applied to PUBG Ally is very similar to the technology we’re applying to Ludi,” Kim explained. “We’re building on those capabilities as we move forward, and our longer-term goal is to expand this work into developing our own proprietary foundation model.”
Krafton believes its gaming expertise offers a competitive advantage in humanoid robotics. Game engines simulate how objects move and interact, providing virtual environments that can generate training data. Kim says Ludo is already using synthetic data from these worlds and plans to incorporate more real-world information as its research advances.
Ludi 0.1, introduced in August, uses an architecture combining several existing AI models. A vision-language model from Alibaba’s open-source Qwen family analyzes camera images and transcribed speech to decide what the robot should do next. Nvidia’s GR00T robotics foundation model, fine-tuned by Ludo, controls physical tasks such as picking up objects. The system runs on Unitree’s G1, a roughly $16,000 humanoid robot, with Ludo’s software coordinating the components.
The company trained its reasoning model on 14,432 computer-generated interaction sequences. The robot learned physical manipulation separately from recordings of people remotely controlling it in a mock household. Although Ludo’s technical report identifies PUBG Ally as a design precedent, it does not list player conversations among its training data.
Kim sees initial opportunities in homes, offices, and factories, where robots must work alongside people rather than simply perform repetitive tasks. “Everyone seems to be working on building a foundation model, but I think that’s an enormous challenge,” he says. “If you’re actually going to develop a foundation model, it needs to be highly generalized. It has to be capable of handling all kinds of different situations, and I think achieving that level of generalization is going to take time.”
For now, Ludo intends to develop the intelligence rather than manufacture robots, working with hardware makers to integrate its software into their machines.
Krafton’s Investment Strategy for Ludo Robotics
Krafton reported second quarter revenue of 1.29 trillion won, about $890 million, up 95% from a year earlier. Its broader AI investments include $70 million for a computing cluster and a reported $33 million investment in HyperAccel, a South Korean startup developing AI chips. Ludo’s team includes Rob Nowak, chief scientist, and Yea-Seul Kim, human-robot interaction lead, both of whom bring academic experience from the University of Wisconsin-Madison.
Kim says Ludo is focusing on research this year before identifying a suitable environment for deployment in the real world 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 says. “That would allow us to continue developing our capabilities as we move toward building a foundation model.”
Krafton may eventually seek external funding. “It’s difficult to say what kind of commercialization can realistically happen with humanoid robots within a single year,” Kim says. “Krafton is investing heavily in AI overall, but robotics is a very long-term project. We may need to raise additional funds from outside the company to properly pursue this area.”
Ludo’s Push for Unified Robotic Intelligence
Kangwook Lee, Ludo’s chief technology officer and Krafton’s chief AI officer, says developing AI characters for video games was easier. “In a game environment, the appropriate programs and algorithms could make characters move and use their bodies naturally, allowing us to focus on improving their social intelligence rather than solving the underlying challenges of physical capability,” he explains.
Real-world robotics is more complicated because physical movement and social interaction must work together. Lee points to how people naturally gesture while speaking. “It’s almost impossible to think about physical movement and social interaction as two completely separate things,” he says.
Both Ludi 0.1 and 0.2 rely on multiple foundation models connected through an agent harness. One interprets the situation and decides what to do, while separate systems execute physical actions. That arrangement becomes less efficient when circumstances change midway through a task.
“The problem is that the other models have no visibility into the reasoning behind a decision. Instead, each receives a final instruction to execute without understanding the broader context, which becomes a significant limitation when circumstances change,” Lee explains. “The system must then return to the reasoning model to reassess the situation and determine the next course of action. This constant back-and-forth disrupts the connection between reasoning and physical execution, creating significant inefficiencies.”
Although Ludi 0.2 improves perception and social understanding, it retains that architecture. Ludo plans a more fundamental redesign with Ludi 1.0, scheduled for 2027. The goal is a unified foundation model that retains context and adapts to changing circumstances without repeatedly transferring control between systems.
“If something changes in the environment, the robot should be able to build on its previous reasoning and adjust its actions much faster,” Lee says. “Instead of passing the problem back to another model to reconstruct the context and determine what to do next, the system could use that shared understanding to respond directly to the new situation.”
That integration also presents computing challenges. Ludi 0.1 processed AI tasks locally on a workstation beside the robot rather than relying on the cloud. Ludo selected a smaller, 4-billion-parameter model to reduce delays. Kim sees hardware limitations as a major obstacle. “A foundation model needs to have generalized capabilities, but one of the biggest challenges is developing those capabilities within the specifications available on the robot itself,” he says.
Ludi’s AI Benchmarks Reveal Performance Gaps
Ludo measures progress through its Interaction Core28 benchmark, which tests AI across 28 simulated household scenarios, including delivering objects, interpreting ambiguous requests and responding when someone speaks while the robot is moving. Before additional training, the smaller Qwen model powering Ludi 0.1 completed 10 scenarios, improving to 17 after fine-tuning. An OpenAI frontier model, using a carefully optimized prompt, completed 26, revealing a substantial performance gap.
These results come entirely from simulations, making comparisons with physical robots difficult. Figure, for instance, reported a 56% task-completion rate across 420 trials in a home environment, though its evaluation did not include interrupted chores or situations involving people occupying the same space. Meanwhile, the consumer robotics company 1X markets its $20,000 NEO household humanoid with remote human assistance when needed.
For Ludi 0.2, Lee says the team retrained the visual reasoning model responsible for interpreting camera images, identifying people and determining how the robot should behave. “We needed to evaluate how robots use their bodies when interacting with people, rather than focusing solely on whether they can perform a particular action,” he explained. “It’s especially important to understand whether the robot behaves appropriately when people are nearby and how its physical movements affect those interactions.”
Krafton describes additional evaluations involving personal preferences, relationships between people in a room, and gestures made while carrying objects. However, the company has not published Ludi 0.2’s Interaction Core28 results or disclosed enough about its new evaluations for independent researchers to reproduce them.
Lee considers recognizing unclear instructions and asking for clarification, rather than making assumptions, an important measure of intelligence. “Our benchmark includes several scenarios in which people deliberately give the robot ambiguous instructions in a simulated environment. In those situations, the expected response is to ask clarifying questions rather than immediately attempting to carry out the task,” he says. “Once the person provides the missing information, the robot should have enough context to proceed.”
In a video game, players can laugh off an AI teammate’s mistakes and continue the match. In the physical world, misunderstandings could force people to repeat instructions or intervene in tasks they expected a robot to handle. That is the real test of Krafton’s gaming expertise: whether its experience making virtual characters responsive to players can translate into reliable physical assistance.
