miHoYo AI research has produced Gamer, a model capable of playing Honkai: Star Rail through ordinary screen controls. Instead of receiving direct instructions from the game’s internal systems, Gamer watches screenshots and operates a keyboard and mouse. It can explore, fight enemies, navigate menus and attempt quests without being a playable feature for users. However, its test results reveal significant differences between tasks it handles successfully and puzzles it struggles to complete.
miHoYo AI Gamer plays Honkai: Star Rail through screenshots
Developed by miHoYo’s Honkai AI research team, Gamer was presented in a paper published on November 13, 2026. The team belongs to the company behind Genshin Impact and Honkai: Star Rail. According to the official Gamer project page, the model studies visual information before deciding which controls to use. Its approach resembles how a person interacts with a game using a monitor and input devices.
Gamer processes five screenshots every second, allowing it to make decisions while gameplay continues. It determines mouse movements and keyboard inputs based on what appears on the screen. Additionally, a task memory records completed objectives, unsuccessful attempts, important locations and available resources. That memory also tracks unfinished work, helping the model maintain context across longer activities.
How Gamer learned to explore, fight and complete quests
Training involved more than 10,000 hours of recorded gameplay, including approximately 1,000 curated hours from Honkai: Star Rail. Those selected recordings included explanations of important decisions and summaries of remembered information. Researchers subsequently improved the model through reinforcement learning, which uses feedback to refine its behavior. The process combined recorded examples with additional training rather than relying entirely on prerecorded actions.
Gamer can control character movement, adjust the camera and interact with objects in its surroundings. It also navigates menus, participates in battles and explores areas while completing quests. Demonstration videos show the model fighting enemies, progressing through quests and defeating a boss. However, successful demonstrations do not establish that it can complete every activity in the game.
miHoYo AI test results show strengths and weaknesses
The research includes measurable results from quests and several types of puzzles. As Dexerto reported on November 17, Gamer completed most objectives in two tested side quests. Its puzzle performance was more uneven, with considerable differences between individual minigames. These results provide a clearer picture than the demonstration footage alone.
- Opportunity Makes the Thief (Part I): 80% of objectives completed.
- Bestial Ferocity: 83.6% of objectives completed.
- Overall puzzle success: 49.7% across 40 puzzles from seven minigame types.
- Without reinforcement learning: 39.2% average puzzle success.
- Best compared general-purpose AI: 17.1% average puzzle success.
- Navigation Compass: 95.8% success rate.
- Abacus Circuitry: 22.2% success rate.
- Response time: Reduced from 1,306 milliseconds to approximately 65 milliseconds.
The comparison suggests that reinforcement learning improved Gamer’s average performance across the tested puzzles. Nevertheless, its 49.7% overall success rate means many attempts remained unsuccessful. Navigation Compass produced particularly strong results, while Abacus Circuitry exposed a substantial weakness. Faster responses also represent a separate improvement from solving tasks correctly.
Could miHoYo AI help developers test games?
The miHoYo AI project explores agents capable of interacting with games through familiar controls. Its stated research areas include quality testing, accessibility and human-AI interaction. Dexerto also identifies possible development applications, including checking quests and reproducing bugs before content launches. Reducing development costs is another potential use discussed in its coverage.
These possibilities concern research applications rather than confirmed production tools already deployed by miHoYo. The company also develops Honkai: Star Rail content, including version 4.7, although Gamer remains a separate research project. More broadly, research into how game developers use generative AI provides another perspective on industry technology. Gamer specifically focuses on performing gameplay actions through visual information and controls.
What has miHoYo released from the Gamer project?
The researchers have released code and several miHoYo AI models associated with Gamer. However, they have withheld the Honkai: Star Rail training data and models trained on that material. Data security is the stated reason for those restrictions. Consequently, the released materials should not be confused with access to the complete demonstrated Honkai: Star Rail system.
Gamer is not an official gameplay feature, and ordinary players cannot use it within Honkai: Star Rail. Its authors also explicitly reject using the research to cheat or exploit games. The project instead investigates how an AI system can operate within an interactive environment. That distinction matters when interpreting its capabilities and the materials made publicly available.
What the research demonstrates so far
The miHoYo AI research shows that a trained model can perform several gameplay tasks using screenshots and conventional inputs. Its memory system supports longer activities, while additional training improves some measured results. However, the puzzle tests reveal limitations that remain important when assessing its reliability. For now, Gamer is a miHoYo research project, not a new Honkai: Star Rail feature.









