miHoYo Publishes Gamer, an AI Agent That Plays Games From the Screen
miHoYo's Honkai AI R&D Team published a project page and paper on November 13 for Gamer, an AI model that reads the game screen and operates a keyboard and mouse the way a human player would.
4 sources 1d ago
Reporting from 4 sources: Denfaminicogamer, GameBusiness.jp, Inside, Game Spark.
Why it matters
The team names automated quest-progress checks and regression testing after updates as intended uses, which puts Gamer closer to a development tool than to a player-facing product.
Key Facts
- miHoYo's Honkai AI R&D Team published a project page and paper on Gamer on November 13.
- Gamer is built on Alibaba's Qwen3-VL (8B) model and trained on tens of thousands of hours of play records.
- Gamer reached an 80.0 percent intermediate goal achievement rate on the Honkai: Star Rail side quest Slow-Witted Thief, Part 1 and 83.6 percent on Fierce Animals.
- Gamer averaged a 49.7 percent success rate across seven Honkai: Star Rail mini-games, against 17.1 percent for GPT-6.1-Sol in the same evaluation.
- The Gamer paper reports results using data-center GPUs such as the H200, and the inference engine is slated for release soon.
The team built Gamer on Alibaba's publicly released Qwen3-VL (8B) model, trained it on tens of thousands of hours of play records, added annotations explaining why each operation was performed, then ran reinforcement learning so the agent could learn goal-directed actions by trial and error. During play Gamer uses no internal game data and works only from what appears on screen. Evaluation used Honkai: Star Rail plus ViZDoom, a research environment based on DOOM. In Star Rail, Gamer reached an 80.0 percent intermediate goal achievement rate on the side quest Slow-Witted Thief, Part 1 and 83.6 percent on Fierce Animals, and averaged a 49.7 percent success rate across seven mini-games it had not trained on.
Game Spark and Inside both describe the team as positioning Gamer as a "test player" for game development. The intended uses are checking that quests progress normally and regression testing existing content after updates.
The project page includes video of Gamer in Honkai: Star Rail's main story: walking toward destinations, following the character leading the way, and clearing battles with enemies and boss fights. Mini-game footage shows it clearing five types of mini-games, from ones requiring logical thinking to ones requiring timing and precision of operation.
On the same seven mini-games, GPT-6.1-Sol scored 17.1 percent against Gamer's 49.7 percent average. The paper reports results on high-performance data-center GPUs such as the H200, so personal use is considered difficult. A GitHub page has been published, with the inference engine slated for release soon.
The paper is titled "Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent" and was published on arXiv.
Written with AI by Yomimono from the 4 cited sources below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google
Sources
- Denfaminicogamer miHoYoが人間のようにキーボードとマウス入力でゲームを操作するAIモデル「Gamer」を開発中。『崩壊:スターレイル』をプレイする様子が公開
- GameBusiness.jp miHoYo、ゲームを遊ぶAIの研究成果を公開。AIがゲーム画面を見てキーボード&マウスを操作―『崩壊:スターレイル』などで検証。やっぱり『DOOM』系環境も使われる
- Inside AIがゲーム画面を見て操作!? miHoYo、汎用AIエージェント「Gamer」の研究成果を公開
- Game Spark miHoYo、ゲームを遊ぶAIの研究成果を公開。AIがゲーム画面を見てキーボード&マウスを操作―『崩壊:スターレイル』などで検証。やっぱり『DOOM』系環境も使われる