Aura Autopilot 开源:为 Unreal/Unity/Godot 拆分规划与反射的游戏测试智能体
观察帖介绍 Aura 将 Autopilot 开源至 Unreal、Unity 与 Godot:编码智能体读代码生成游戏专用控制器,决策模型选动作,快速反射可走代码路径;持续学习实为项目级记忆而非改权重。
有技术观察帖介绍 Aura 的游戏测试智能体 Autopilot 已对 Unreal、Unity、Godot 开源。其核心是把规划与反射拆开:编码智能体阅读游戏代码并生成游戏专用控制器;决策模型在游玩过程中选择动作;像躲避攻击这类快速反应可以作为代码路径执行,而不必等待又一次模型响应。
帖文强调,这一模式也适用于游戏之外:用模型去构建控制环,而不是让模型处理每一个时间敏感事件。所谓「持续学习」是项目级记忆(笔记、可复用控制器代码与历史运行),并非训练新权重;系统会标记变更过的源文件,以便后续运行重新检查假设。实用示例:改动后只验证「捡起石头是否更新背包」一类机制并审阅报告。发布基准是在一款 Unreal 游戏上的七项任务,每次尝试前清空记忆——作者称这是架构的早期证据,不能证明它可替代跨游戏 QA。
对品质团队,Autopilot 更像「可生成的专用测试司机」,适合回归单点机制;跨项目泛化、闪烁帧与非确定性物理仍需人工设计用例。引入前应评估开源许可、引擎版本矩阵,以及控制器代码被提交进仓库后的代码审查流程。
Aura Autopilot Open Source Game Testing Agent
Aura's new game-testing agent separates planning from reflexes.
Autopilot is now open source for Unreal, Unity and Godot. A coding agent reads the game code and builds a game-specific controller; a decision model chooses actions during play. Fast reactions, like dodging an attack, can run as code without waiting for another model response.
That is a useful pattern beyond games: use the model to build the control loop, not to handle every time-sensitive event.
The "continual learning" here is per-project memory: notes, reusable controller code and previous runs. It isn't training new model weights. The system flags changed source files so later runs can revisit those assumptions.
A practical use: ask it to verify one mechanic after a change, such as whether collecting a stone updates inventory, then inspect the report.
The launch benchmark is seven tasks in one Unreal game, with memory cleared before each attempt. Interesting early evidence for the architecture—not proof that it replaces QA across games.