miHoYo releases 'Gamer' VLA game-playing research
miHoYo's Honkai AI R&D team has published research on "Gamer," a Vision-Language-Action (VLA) foundation model capable of playing complex 3D games like Honkai: Star Rail. Unlike traditional bots, Gamer interacts directly with the game interface by observing the screen and controlling a keyboard and mouse, bypassing the need for internal game APIs. The research demonstrates a complete pipeline for training, annotation, and real-time inference (using a custom Streaming-vLLM engine), with potential applications in automated game QA, accessibility features, and generalist AI agent development.
* Hot take: This is a massive leap in VLA performance, proving that complex, real-time gaming environments can be mastered without "cheating" through game state access, setting a new benchmark for human-AI interaction.
* It showcases a highly sophisticated training and inference pipeline, particularly the custom Streaming-vLLM engine, which is critical for real-time actuation.
* The focus on human-like interaction (vision-in, keyboard/mouse-out) makes this significantly more transferable to real-world physical robotics tasks than state-based agent approaches.
* While technically impressive, it inevitably raises questions regarding the future of bot-detection in live-service titles.
DISCOVERED
9h ago
2026-11-17
PUBLISHED
9h ago
2026-11-17
RELEVANCE
AUTHOR
gamespark