Popular games have become testing grounds for AI experimentation.
Recently, researchers in China published the results of AI experiments. Their project, built on the Qwen2 model, is called “Lumine”:
We present Lumine — an open blueprint for building generalist agents capable of executing multi‑hour, complex missions in real time within intricate 3D open‑world environments. Lumine adopts a human‑like interaction paradigm that seamlessly fuses perception, reasoning, and action, grounded in a vision‑language model.
They trained Lumine on a tailored dataset. From an initial collection exceeding 2,400 hours of recordings, filtering and post‑processing distilled the material into 200 high‑quality samples plus 15 hours of manually annotated reasoning traces that help cultivate the model’s hybrid reasoning capabilities.
The authors report that Lumine successfully completes the five‑hour Mondstadt campaign in Genshin Impact, and has shown successes in other games as well:
Without any fine‑tuning, it handles missions in Wuthering Waves and completes the full five‑hour opening chapter of Honkai: Star Rail. These encouraging results highlight Lumine’s effectiveness across varied worlds and interaction dynamics, marking a noteworthy advance toward universal agents for open gaming environments.
Source: iXBT.games