NewsPublished:
miHoYo's Gamer AI plays Honkai: Star Rail with a keyboard and mouse, much as a person would
miHoYo's Honkai AI R&D Team published a project page and report for Gamer, an agent that turns screenshots into keyboard and mouse input in Star Rail.

If you have ever wondered what it would look like for an AI to grind through Honkai: Star Rail for you, miHoYo's research team has just put a version of that on public display. The company's Honkai AI R&D Team has released a project page and technical report for Gamer, a game-playing agent that was shown clearing Star Rail quests and puzzles.
How it plays
Gamer does not read the game's internal data while playing. It looks at screenshots, sampled five times a second, and outputs what a person's hands would do: key presses, mouse movement and scrolling. Each 200 ms window carries six short key-state blocks, which the team says adds up to roughly 30 Hz control.
How it was trained
- Continued pre-training on more than 10,000 hours of keyboard-and-mouse gameplay.
- Instruction tuning with about 1,000 hours of curated Star Rail footage, including annotated reasoning and a running memory of what it has already done.
- Online reinforcement learning, with the game environment checking whether tasks were actually completed.
What it showed
The page features long main-story walkthroughs and five types of puzzle in Star Rail. In seven puzzle types with 40 held-out tasks, the reinforcement-learned model averaged a 49.7% success rate. The strongest general-purpose baseline the team tested reached 17.1%. The team also reports cutting inference latency from about 1,306 ms to 65 ms on one setup.
What it is not
This is research, not a feature. The page does not announce anything for players. The team says the likely upside is on the development side, such as automated testing and faster iteration, and it admits the evaluation covers a selected set of games and tasks.
Still, it is a vivid look at how a major anime-style game studio is thinking about agents. Project page: mihoyo-gamer.github.io.
Source: mihoyo-gamer.github.io