PUBG maker Krafton shows Ludi 0.2, robot AI built to read people

Ludo lists five areas it improved: reading intent, feedback and social cues; memory and reasoning across changing tasks; coordinating navigation, manipulation and interaction; recognising gestures; and smoother movement in supported environments[1]. The announcement gives no benchmark figures or hardware details for the new version[1].

"Robots are becoming increasingly capable of performing physical tasks, but understanding and responding to people remains a major challenge," said Kangwook Lee, Krafton's chief AI officer and Ludo's CTO[1].

The clearest picture comes from the Ludi 0.1 paper Ludo published in August. A Qwen3.5 vision-language model, fine-tuned on 14,432 synthetic multi-turn interaction traces, looks at the robot's cameras and the conversation and picks one action per turn: speak, wait, look up the map, navigate, check reachability or run a manipulation skill[2]. A harness validates and executes each call, and a user can interrupt mid-action, as in the paper's demo where a Unitree G1 humanoid drops a reach for a Coke when told the recipient prefers Pepsi[2].

Manipulation runs through vision-language-action policies built on NVIDIA's GR00T N1.7 and fine-tuned on teleoperated demonstrations, and the whole stack can run on local hardware without cloud models[2]. In Ludo's 28-scenario simulation benchmark, the fine-tuned 9B model completed 20 scenarios against 11 before fine-tuning, while GPT-5.6 with an optimised prompt completed 26. Ludo deployed the smaller 4B model because it responds faster[2].

Krafton pitches the work as an extension of the AI it builds for games, including the Co-Playable Characters it develops with NVIDIA[1]. Ludo is headquartered in Palo Alto with an R&D lab in Seoul, and trains on a mix of simulation, AI-generated social interactions and real human behaviour data[1][5].

Next on its roadmap is Ludi 1.0, a foundation model meant to fold perception, dialogue and physical action into one system rather than a set of tools stitched together by a harness[1][2].