causal-gpt-rl-unity-datasets
收藏资源简介:
Causal GPT-RL Unity ML-Agents轨迹数据集是一个包含多个Unity ML-Agents环境离线轨迹的集合,以Minari标准格式封装。该数据集专为测试Causal GPT-RL算法而设计,同时涵盖连续和离散动作空间。数据集包含8个不同的Unity环境:Crawler(爬行者)、PushBlock(推箱子)、SoccerTwos(双人足球)、DungeonEscape(地牢逃脱)、3DBallHard(三维球体困难版)、Pyramids(金字塔迷宫)、Worm(蠕虫)和Walker(步行者)。每个环境提供三个校准的技能等级:expert(专家级,技能水平1.0)、medium(中级,约0.80)和simple(初级,约0.60),这些等级通过向原始策略注入不同程度的动作噪声生成。对于单智能体场景,噪声水平在每回合内从校准范围内独立采样;对于合作场景(如SoccerTwos和DungeonEscape),噪声在每场比赛或每组内共享。数据集总计包含21个子集,每个子集约100万次状态转移,观察空间和动作空间格式因环境而异,包括Box、Discrete、MultiDiscrete等类型,部分多智能体环境使用ego-centric的Dict包装结构。所有数据均基于ML-Agents release_23版本生成,使用Minari 0.5.3格式存储,适用于离线强化学习、行为克隆、策略评估等研究任务。
The Causal GPT-RL Unity ML-Agents Trajectory Dataset is a collection of offline trajectories from multiple Unity ML-Agents environments, packaged in the Minari standard format. It is specifically designed for testing Causal GPT-RL algorithms and covers both continuous and discrete action spaces. The dataset includes eight distinct Unity environments: Crawler, PushBlock, SoccerTwos, DungeonEscape, 3DBallHard, Pyramids, Worm, and Walker. Each environment offers three calibrated skill levels: expert (skill level 1.0), medium (approximately 0.80), and simple (approximately 0.60), generated by injecting varying degrees of action noise into the original policies. For single-agent scenarios, noise levels are independently sampled from a calibrated range within each episode; for cooperative scenarios (such as SoccerTwos and DungeonEscape), noise is shared per match or group. The dataset comprises a total of 21 subsets, each containing approximately 1 million state transitions. Observation and action space formats vary by environment, including types such as Box, Discrete, and MultiDiscrete, with some multi-agent environments using ego-centric Dict wrapper structures. All data is generated based on ML-Agents release_23 and stored in Minari 0.5.3 format, making it suitable for research tasks like offline reinforcement learning, behavioral cloning, and policy evaluation.
数据集概览:Causal GPT-RL — Unity ML-Agents trajectories
该数据集收录了通过 Unity ML-Agents 录制的轨迹,并以 Minari 数据集格式打包,包含连续和离散动作空间,用于测试 Causal GPT-RL。数据集提供多个环境的“质量阶梯”:expert(专家级)、medium(中级)和 simple(初级)三个等级,其中低级技能等级通过在原始策略中注入动作噪声生成,技能等级以随机策略(0.0)和专家策略(1.0)为锚点进行校准。
数据集内容总表
| 数据集ID | 片段数 | 转换数 | 观测空间 | 动作空间 |
|---|---|---|---|---|
| unity/crawler/expert-v0 | 1,048 | 1,000,002 | Tuple(Box(126), Box(32)) |
Box(20, [-1, 1]) |
| unity/crawler/medium-v0 | 1,129 | 1,000,000 | Tuple(Box(126), Box(32)) |
Box(20, [-1, 1]) |
| unity/crawler/simple-v0 | 1,344 | 1,000,000 | Tuple(Box(126), Box(32)) |
Box(20, [-1, 1]) |
| unity/pushblock/expert-v0 | 52,801 | 1,000,000 | Tuple(Box(105), Box(105)) |
Discrete(7) |
| unity/pushblock/medium-v0 | 2,450 | 1,000,000 | Tuple(Box(105), Box(105)) |
Discrete(7) |
| unity/pushblock/simple-v0 | 1,783 | 1,000,000 | Tuple(Box(105), Box(105)) |
Discrete(7) |
| unity/soccer-twos/expert-v0 | 14,588 | 1,000,816 | ego Dict 包裹 Tuple(Box(264), Box(72)) |
ego Dict 包裹 MultiDiscrete([3, 3, 3]) |
| unity/soccer-twos/medium-v0 | 13,040 | 1,001,776 | ego Dict 包裹 Tuple(Box(264), Box(72)) |
ego Dict 包裹 MultiDiscrete([3, 3, 3]) |
| unity/soccer-twos/simple-v0 | 12,860 | 1,003,048 | ego Dict 包裹 Tuple(Box(264), Box(72)) |
ego Dict 包裹 MultiDiscrete([3, 3, 3]) |
| unity/dungeon-escape/expert-v0 | 35,787 | 1,003,523 | ego Dict 包裹 Tuple(Box(10), Box(360), Box(1)) |
ego Dict 包裹 Discrete(7) |
| unity/dungeon-escape/medium-v0 | 19,023 | 1,003,025 | ego Dict 包裹 Tuple(Box(10), Box(360), Box(1)) |
ego Dict 包裹 Discrete(7) |
| unity/dungeon-escape/simple-v0 | 11,367 | 1,003,316 | ego Dict 包裹 Tuple(Box(10), Box(360), Box(1)) |
ego Dict 包裹 Discrete(7) |
| unity/3dball-hard/expert-v0 | 1,009 | 1,000,008 | Tuple(Box(27), Box(18)) |
Box(2) |
| unity/3dball-hard/medium-v0 | 1,250 | 1,000,008 | Tuple(Box(27), Box(18)) |
Box(2) |
| unity/3dball-hard/simple-v0 | 1,655 | 1,000,008 | Tuple(Box(27), Box(18)) |
Box(2) |
| unity/pyramids/expert-v0 | 5,348 | 1,000,000 | Tuple(Box(56), Box(56), Box(56), Box(4)) |
Discrete(5) |
| unity/pyramids/medium-v0 | 1,858 | 1,000,000 | Tuple(Box(56), Box(56), Box(56), Box(4)) |
Discrete(5) |
| unity/pyramids/simple-v0 | 1,399 | 1,000,000 | Tuple(Box(56), Box(56), Box(56), Box(4)) |
Discrete(5) |
| unity/worm/expert-v0 | 1,000 | 1,000,000 | Box(64) |
Box(9, [-1, 1]) |
| unity/worm/medium-v0 | 1,000 | 1,000,000 | Box(64) |
Box(9, [-1, 1]) |
| unity/worm/simple-v0 | 1,000 | 1,000,000 | Box(64) |
Box(9, [-1, 1]) |
| unity/walker/expert-v0 | 1,458 | 1,000,010 | Box(243) |
Box(39, [-1, 1]) |
| unity/walker/medium-v0 | 1,712 | 1,000,000 | Box(243) |
Box(39, [-1, 1]) |
| unity/walker/simple-v0 | 2,247 | 1,000,000 | Box(243) |
Box(39, [-1, 1]) |
数据集技术细节
- ML-Agents 版本:所有环境、轨迹数据集和策略均基于 ML-Agents
release_23。 - 存储格式:每个数据集存储在
<name>/<tier>/data/main_data.hdf5,附带metadata.json,使用minari_version0.5.3。 - PushBlock:使用股票
PushBlock.onnx策略在模型移除版 Unity 构建中外部运行,为原始离散 PushBlock 环境。 - SoccerTwos 和 DungeonEscape:采用显式自我智能体模式,每个物理智能体存储为一个自我中心片段,使用
observations["agents"]["agent_0"]和actions["agents"]["agent_0"]。 - SoccerTwos:股票策略自我对弈数据集,包含 3,647 场完整四玩家比赛,每场比赛表示为四个链接的去中心化智能体片段,所有 14,588 个片段均正常终止,无截断。
质量阶梯详情
DungeonEscape 质量阶梯
| 等级 | 组成功率 | epsilon 池 | 平均片段长度 | 近似归一化技能 |
|---|---|---|---|---|
| expert-v0 | 0.9532 | 无噪声 | 28.04 | 1.00 |
| medium-v0 | 0.7851 | {0.42, 0.44, 0.46, 0.48} | 52.73 | ~0.78 |
| simple-v0 | 0.6247 | {0.60, 0.65, 0.70, 0.75} | 88.27 | ~0.56 |
- unity/dungeon-escape/medium-v0:包含 6,341 个完整组,19,023 个终止片段,1,003,025 次转换,4,978 次组成功,1,363 次失败(成功率 0.7851)。
- unity/dungeon-escape/simple-v0:包含 3,789 个完整组,11,367 个终止片段,1,003,316 次转换,2,367 次组成功,1,422 次失败(成功率 0.6247)。
Pyramids 质量阶梯
| 等级 | 近似归一化回报 | epsilon 范围 | 平均回报 | 到达率 | 平均片段长度 |
|---|---|---|---|---|---|
| expert-v0 | 1.00 | -- | 1.795309 | 0.9906 | 186.99 |
| medium-v0 | ~0.80 | 0.50–0.66 | 1.228100 | 0.8827 | 538.21 |
| simple-v0 | ~0.58 | 0.62–0.74 | 0.618400 | 0.6662 | 714.80 |
单智能体连续控制质量阶梯
Crawler(动作空间 Box(20);锚点:专家 2576.4899,随机 -0.88)
| 等级 | 近似归一化回报 | noise_std 范围 | 平均回报 | 终止率 | 平均片段长度 |
|---|---|---|---|---|---|
| expert-v0 | 1.00 | -- | 2576.4899 | 0.077 | 954.20 |
| medium-v0 | ~0.82 | 0.15–0.23 | 2113.0960 | 0.209 | 885.74 |
| simple-v0 | ~0.60 | 0.21–0.32 | 1557.9818 | 0.460 | 744.05 |
Worm(动作空间 Box(9);锚点:专家 1044.1063,随机 0.80)
| 等级 | 近似归一化回报 | noise_std 范围 | 平均回报 | 终止率 | 平均片段长度 |
|---|---|---|---|---|---|
| expert-v0 | 1.00 | -- | 1044.1063 | 0.000 | 1000.00 |
| medium-v0 | ~0.80 | 0.09–0.15 | 839.4910 | 0.000 | 1000.00 |
| simple-v0 | ~0.60 | 0.15–0.22 | 632.4776 | 0.000 | 1000.00 |
Walker(动作空间 Box(39);锚点:专家 1363.6454,随机 -0.49)
| 等级 | 近似归一化回报 | noise_std 范围 | 平均回报 | 终止率 | 平均片段长度 |
|---|---|---|---|---|---|
| expert-v0 | 1.00 | -- | 1363.6454 | 0.512 | 685.88 |
| medium-v0 | ~0.82 | 0.05–0.11 | 1122.0456 | 0.652 | 584.11 |
| simple-v0 | ~0.61 | 0.09–0.15 | 835.2383 | 0.822 | 445.04 |
3DBallHard(动作空间 Box(2);锚点:专家 99.1077,随机):1,000,008 次转换,1,009 个片段,平均回报 99.107732,平均片段长度 991.0882。
相关仓库
- Unity 构建和 ONNX 策略:https://huggingface.co/datasets/ccnets/causal-gpt-rl-unity-envs
- Causal GPT-RL 策略检查点:https://huggingface.co/ccnets/causal-gpt-rl-unity
许可证
该数据集使用 CC-BY-4.0 许可证。




