causal-gpt-rl-unity-envs
收藏资源简介:
该数据集(仓库)提供了Causal GPT-RL项目所使用的八个Unity ML-Agents环境及其相关策略文件。每个环境均以模型移除(model-removed)的Unity构建形式提供,即引擎二进制文件不含内置策略,策略需在运行时外部提供。配套提供了每个环境的库存内置ONNX策略文件(stock built-in policy),以及四个离散动作环境的logits暴露变体(存放于tier_policies/目录下)。八个环境包括:Crawler(10个智能体,158维观察,连续动作Box(20))、3DBallHard(12个智能体,45维观察,连续动作Box(2))、PushBlock(32个智能体,210维观察,离散动作Discrete(7))、DungeonEscape(36个智能体,371维观察,离散动作Discrete(7))、SoccerTwos(32个智能体,336维观察,多离散动作MultiDiscrete([3,3,3]))、Pyramids(16个智能体,172维观察,离散动作Discrete(5))、Worm(10个智能体,64维观察,连续动作Box(9))和Walker(10个智能体,243维观察,连续动作Box(39))。所有构建均基于ML-Agents release-23定制项目,移除了内置模型,并可在无头模式下运行。该数据集主要用于运行Causal GPT-RL策略、记录轨迹以及进行强化学习评估与实验。
This dataset (repository) provides eight Unity ML-Agents environments used in the Causal GPT-RL project, along with their associated policy files. Each environment is provided as a model-removed Unity build, meaning the engine binary does not contain a built-in policy; the policy must be supplied externally at runtime. It includes a stock built-in ONNX policy for each environment, as well as logits-exposed variants for four discrete action environments (located in the tier_policies/ directory). The eight environments are: Crawler (10 agents, 158-dimensional observation, continuous action Box(20)), 3DBallHard (12 agents, 45-dimensional observation, continuous action Box(2)), PushBlock (32 agents, 210-dimensional observation, discrete action Discrete(7)), DungeonEscape (36 agents, 371-dimensional observation, discrete action Discrete(7)), SoccerTwos (32 agents, 336-dimensional observation, multi-discrete action MultiDiscrete([3,3,3])), Pyramids (16 agents, 172-dimensional observation, discrete action Discrete(5)), Worm (10 agents, 64-dimensional observation, continuous action Box(9)), and Walker (10 agents, 243-dimensional observation, continuous action Box(39)). All builds are based on a customized ML-Agents release-23 project with built-in models removed and can run in headless mode. This dataset is primarily used for running Causal GPT-RL policies, recording trajectories, and conducting reinforcement learning evaluation and experiments.
Causal GPT-RL — Unity ML-Agents 环境数据集
数据集概述
该数据集提供用于在 Unity ML-Agents 中运行 Causal GPT-RL 策略的公开材料。每个环境都是一个移除模型后的 Unity 构建——即不包含内置策略的引擎二进制文件,策略在运行时提供。所有八个环境都配有内置的原始策略,其中四个离散动作环境额外提供暴露 logits 的变体(位于 tier_policies/ 目录下)。
环境详情
| 环境 | 构建目录 | 原始策略文件 | 智能体数量 | 观测维度 | 动作空间 |
|---|---|---|---|---|---|
| Crawler | Crawler/ |
Crawler.onnx |
10 | 158(126 + 32) | Box(20) 连续 |
| 3DBallHard | 3DBallHard/ |
3DBallHard.onnx |
12 | 45(27 + 18) | Box(2) 连续 |
| PushBlock | PushBlock/ |
PushBlock.onnx |
32 | 210(105 + 105) | Discrete(7) |
| DungeonEscape | DungeonEscape/ |
DungeonEscape.onnx |
36 | 371 | Discrete(7) |
| SoccerTwos | SoccerTwos/ |
SoccerTwos.onnx |
32 | 336(264 + 72) | MultiDiscrete([3,3,3]) |
| Pyramids | Pyramids/ |
Pyramids.onnx |
16 | 172(56 + 56 + 56 + 4) | Discrete(5) |
| Worm | Worm/ |
Worm.onnx |
10 | 64 | Box(9) 连续 |
| Walker | Walker/ |
Walker.onnx |
10 | 243 | Box(39) 连续 |
目录结构
数据集包含以下主要组成部分:
- 8 个环境构建目录:每个目录包含可执行文件(如
.exe)、UnityPlayer.dll、UnityCrashHandler64.exe、*_Data/及MonoBleedingEdge/等运行时组件 - 8 个原始策略 ONNX 文件:与各环境对应,作为收集驱动策略
tier_policies/目录:包含 4 个离散策略的 logits 暴露变体(PushBlock、Pyramids、DungeonEscape、SoccerTwos),附带README.md说明各层级的校准温度带- 许可证文件:
LICENSE(Apache-2.0)和NOTICE(归属及修改声明)
关键特征
- 模型移除构建:所有构建均未内置策略,需在运行时提供策略(可以是原始策略或 Causal GPT-RL 策略),所有构建均支持无头启动
- 构建来源:基于定制的 ML-Agents release-23 项目生成,其中 Crawler、PushBlock、Pyramids、Worm、Walker、DungeonEscape 和 SoccerTwos 源自
ml-agents-release_23 - 不可从原始 ML-Agents 源码重新构建:项目修改存在于编译后的二进制文件中,构建按原样提供
- 已排除内容:各构建均排除了
*_BurstDebugInformation_DoNotShip/文件夹(Burst 调试符号) - Windows SmartScreen 提示:构建未进行代码签名,首次启动时 Windows SmartScreen 可能显示警告
环境特定说明
- Crawler:行为标识为
Crawler?team=0;配套数据集中观测存储为Tuple(Box(126), Box(32)) - 3DBallHard:ONNX 将两个传感器通道作为单独输入(
obs_0 [27]和obs_1 [18]),而非拼接的[45]向量;返回continuous_actions(采样)和deterministic_continuous_actions(贪心),均为float32[B, 2] - PushBlock:动作 0 表示无操作,1–6 表示移动/转向;配套数据集中观测存储为
Tuple(Box(105), Box(105)) - DungeonEscape:36 个智能体按 12 个竞技场 × 3 个协作智能体 排列;同一竞技场内的三个智能体共享协作目标,但每次策略调用接收各自的自我观测
- SoccerTwos:按 8 个 2v2 场地 排列,行为分别为
SoccerTwos?team=0和SoccerTwos?team=1,各 16 个智能体;原始图与 release-23 源码策略计算等价,仅移除了引用无图值的悬空输出声明以便当前 ONNX Runtime 版本加载
配套资源
- 策略模型仓库:https://huggingface.co/ccnets/causal-gpt-rl-unity (ONNX 策略)
- 轨迹数据与复现方案:https://huggingface.co/datasets/ccnets/causal-gpt-rl-unity-datasets (记录轨迹及复现配方)
- 复现方法:https://github.com/ccnets-team/causal-gpt-rl/tree/main/examples/unity_collection (轨迹收集与测量配方)
- 评估教程:https://github.com/ccnets-team/causal-gpt-rl/tree/main/examples/unity (包含 DungeonEscape 评估及 SoccerTwos 因果对比教程)
许可证
Unity 构建、原始策略(全部 8 个 .onnx 文件)以及由此衍生的 tier_policies/ 导出均采用 Apache-2.0 许可证(Unity ML-Agents)。构建为修改后的衍生作品,修改内容在 NOTICE 中描述。Unity 运行时组件(UnityPlayer.dll、UnityCrashHandler64.exe、MonoBleedingEdge/)作为已构建播放器的一部分重新分发,受 Unity Technologies 软件条款约束。




