四个在谷歌Gemini 4 Argon上运行的AI代理在没有人类帮助的情况下击败了《我的世界》的最终Boss——末影龙。谷歌DeepMind的研究人员在arXiv上发表的一篇论文中写道，从他们首次运行程序开始，用时27小时14分钟。
Four AI agents running on Google’s Gemini 4 Argon have defeated the Ender Dragon, Minecraft’s final boss, without human help. It took them 27 hours and 14 minutes from their first program, Google DeepMind researchers wrote in a paper published on arXiv.
这些代理在Easy难度下的随机生成的生存世界中开始，作弊命令已关闭。他们的唯一目标是击败末影龙。没有人给他们分配角色、策略或特定于游戏的工具。
The agents started in a randomly generated survival world on Easy difficulty, with cheat commands turned off. Their only goal was to beat the dragon. No one gave them roles, a strategy or game-specific tools.
据我们所知，这是AI代理首次实现《我的世界》的端到端自主通关，且仅使用标准Mineflayer库，未使用人类预先勘探的种子或其他专门为游戏编写的工具。马克桑斯·福尔多、桑德尔·辛格和罗伯塔·雷莱阿努，谷歌DeepMind代码，而非键盘。这些代理从未看到屏幕或按下任何键。每个代理通过Mineflayer控制一个玩家，Mineflayer是一个用于《我的世界》机器人的开源库。
“To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans.”Maxence Faldor, Satinder Singh and Roberta Raileanu, Google DeepMind Code, not keyboards The agents never saw the screen or pressed a key. Each one controlled a player through Mineflayer, an open-source library for Minecraft bots.
为了行动，代理会编写一个简短的JavaScript程序，例如前往某地并挖掘附近铁矿的指令。随后，它会读取自身的位置、生命值和库存，或遇到的错误。
To act, an agent wrote a short JavaScript program, such as an order to walk somewhere and mine nearby iron. It then read back its position, health and inventory, or the error it hit.
在整个过程中，这四个代理共编写了2783个程序，交换了229条聊天消息，并死亡605次。其中，末影龙导致了69次死亡。
Over the run, the four wrote 2,783 programs, swapped 229 chat messages and died 605 times. The dragon accounted for 69 of those deaths.
- 人名和地名的转写应完全使用中文字符；品牌名称和产品名称可保持原样（即使用拉丁字母）。
Four agents, four jobs The agents split the work themselves. One, janesick, made tools, found the Nether fortress and located the stronghold that holds the portal to the End.
翻译内容：共有四名成员，各自负责不同的任务。其中一名成员名叫“Janesick”，他负责制作工具、寻找下界（Nether）的堡垒，并定位通往“末地”（End）的传送门所在的位置。他通过代码追踪了两个“Enderman”怪物的移动轨迹，计算出它们相遇的地点——该地点距离传送门房间约43米，距离传送门本身约1.3公里。
It threw two Eyes of Ender from different spots, tracked their flights in code and calculated where the paths crossed. The estimate landed 43 metres from the portal room, 1.3 km away.
另一名成员名叫“Gyldenia”，负责收集食物和物资；“Auratian”负责启动下界的传送门；“Bulicame”则协助团队建造用于躲避“Enderman”怪物的庇护所，并最终给予了这些怪物致命一击。整个团队会定期共享坐标信息及行动经验，在成员死亡后也会互相捡起对方丢失的物品。
Another, gyldenia, gathered food and supplies. auratian lit the Nether portal, and bulicame helped build the team’s Enderman shelters and later landed the final blow. The team kept shared notes of coordinates and lessons, and recovered each other’s items after deaths.
这场战斗持续了近14个小时。大约在战斗进行到21个小时时，团队发现那些位于柱子顶部的十颗水晶一直在不断地治疗那条龙；同时他们还发现可以通过堆积泥土来接近这些水晶。在接下来的4小时40分钟内，团队成功摧毁了所有十颗水晶。39分钟后，那条龙终于死亡了（此时它的生命值只剩下20点中的2.5点）。
The fight in the End took nearly 14 hours. About 21 hours in, auratian worked out that ten crystals on top of pillars kept healing the dragon. It also found the team could reach them by stacking dirt. The four destroyed all ten over the next four hours and 40 minutes. The dragon died 39 minutes later, with bulicame on 2.5 of 20 health points.
为了取得胜利，玩家们总共消耗了约13.5亿个输入令牌（input tokens）。根据Argon游戏在Gemini 4阶段的初始定价，这些令牌的价值约为520美元；按照标准价格计算，则为1040美元。整个实验过程（包括一名专门负责监控玩家是否作弊的第五名“代理”）的总成本约为810美元，或1620美元。
What it cost, and its limits Up to the win, the players processed about 1.35 billion input tokens. At Gemini 4 Argon’s introductory list price, that comes to about $520, or $1,040 at the standard price, the researchers estimate. The whole session, including a fifth agent that watched for cheating and found none, came to about $810, or $1,620.
论文中提到了该系统的局限性：Mineflayer会向代理们精确地提示他们周围存在的方块和生物的位置（即使这些生物隐藏在坚硬的岩石后面）；同时还有一个插件负责为代理们规划行动路线。尽管他们在“简单难度”模式下进行实验，但仍有22%的程序出现了错误。需要注意的是，这些实验结果仅基于一次实验得出的。虽然此前已有两个社区项目成功杀死了这条龙，但这些项目都是使用了专门为该游戏设计的工具或采用了更为简单的实验环境。
The paper lists the limits. Mineflayer tells the agents exactly which blocks and creatures are around them, even behind solid rock, and a plugin plans their routes. They played on Easy, 22% of their programs ended in an error, and the result comes from a single run. Two community projects have reported earlier dragon kills, but with game-specific tools or on a fixed, easier setup, the paper notes.
实验中的代理们是在一个受限制的环境（即“沙盒环境”）中运行的，他们的程序仅能控制自己所操控的角色。研究人员可以通过实时监控界面暂停或停止实验的进行，但实际上他们并没有这么做。Faldor在X平台上写道：Google DeepMind计划将这种技术应用于编程、科学研究以及机器人开发领域。上周，该公司的联合创始人Sergey Brin表示，Gemini系统正在为下一代 Gemini 系统的开发提供支持。
The agents ran in a sandboxed world, and their programs could control only their own player. The researchers could pause or stop the run from a live dashboard, but did not need to. Google DeepMind plans to use the same setup for coding, science and robotics work, Faldor wrote on X. Last week, cofounder Sergey Brin said Gemini now helps build the next Gemini.