aifollow.news
返回 IT之家 科技新闻
IT之家 科技新闻· · 原发布时间 AI 评分54

米哈游公布游戏操作模型 Gamer,可看画面自主操作游戏

AI 辅助摘要

Gamer 以每秒 5 帧的游戏画面为输入,生成键盘和鼠标操作,无需读取游戏内部数据或调用专用 API。研究团队展示了它游玩《崩坏:星穹铁道》;强化学习版本在 40 项未接触过的小游戏任务中平均成功率为 49.7%。相关游戏数据及用其训练或微调的模型权重尚未公开。

正文 · AI 翻译

Thanks to IT Home reader 肖战割割 for the tip!

IT Home reported on November 17 that the AI research and development team behind miHoYo’s Honkai series has unveiled Gamer, a general-purpose vision-language-action foundation model. The model can play 3D and 2D games autonomously by simulating how a person operates them.

According to information published in the research paper and on the project page, Gamer does not need to read internal game data or rely on game-specific APIs. Instead, it takes game images captured at 5 frames per second (5 Hz) as input, then generates native keyboard key states and mouse movement commands at an equivalent rate of 30 Hz, simulating how a person plays games with a keyboard and mouse.

To address memory limits and inference latency during long-running tasks, the research team used more than 10,000 hours of keyboard and mouse interaction trajectories across games for continual pretraining, and further optimized the model with sparse thought instruction fine-tuning and temporal-aware reinforcement learning. The team also developed an inference engine called Streaming-vLLM. On NVIDIA H200 hardware, the engine cut end-to-end request latency from 1306 milliseconds to 65.12 milliseconds.

The research team also showed a live demonstration of Gamer carrying out main story missions, exploring environments, and completing puzzle challenges in Honkai: Star Rail. Test data showed that Gamer-RL, trained with reinforcement learning, achieved an average success rate of 49.7% across 40 minigame tasks it had not previously encountered, outperforming the general-purpose model without fine-tuning.

In response to concerns that the technology could be used to develop game cheats and automation scripts, the research team emphasized that Gamer is an academic research project intended to explore general-purpose intelligent agents in interactive environments. Potential applications include game testing, accessibility assistance, and human-computer interaction research. The team stated explicitly that it does not support using Gamer for game cheating, unauthorized access, exploiting security vulnerabilities, or any conduct that harms users’ interests or undermines fairness in games.

For data security reasons, miHoYo’s research team has so far released only general-purpose files used in the project’s early foundational pretraining research stage. It has not released any Honkai: Star Rail-related training data or model weights trained or fine-tuned using data from that game.

发现内容有误?提交纠错