dl2l-experiments
收藏资源简介:
DL2L Experiments Dataset是一个来自DL2L分布式人工生命模拟器的仿真轨迹数据集,专门用于训练JEPA世界模型。数据集按实验前缀(如p9)组织,每个实验包含Parquet格式的训练和验证文件,以及记录元数据的stats.json文件。以p9实验为例,数据来源于10个生物体在DL2L basic.conf世界配置下进行的10次试验模拟,其中试验1-8作为训练集(359,782个样本),试验9-10作为验证集(89,731个样本),按试验划分以防止跨试验污染。每个数据样本代表一个(s_t, a_t, emotion_target)元组,包含感知特征(距离、角度、方向)、对象类型独热编码(6种苹果和植物类型)、动作独热编码(9种行为动作)以及目标情感(9种情感在下一调节周期后的绝对唤醒水平)。双编码器版本还额外包含动作时刻的稳态状态列(饥饿、睡眠、疼痛、厌倦)。数据集适用于人工生命、强化学习和世界模型训练任务,尤其适合学习生物体在模拟环境中的感知-动作-情感动态。
The DL2L Experiments Dataset is a simulation trajectory dataset from the DL2L distributed artificial life simulator, specifically designed for training JEPA world models. It is organized by experiment prefixes (e.g., p9), with each experiment containing training and validation files in Parquet format, along with a stats.json file that records dataset metadata. Taking the p9 experiment as an example, the data is derived from 10 trials simulated by 10 organisms under the DL2L basic.conf world configuration, where trials 1-8 serve as the training set (359,782 samples) and trials 9-10 as the validation set (89,731 samples), with this trial-based partitioning aimed at preventing cross-trial contamination. Each data sample represents an (s_t, a_t, emotion_target) tuple, including the following column groups: perceptual features (distance, angle, direction), object type one-hot encodings (6 types of apples and plants), action one-hot encodings (9 behavioral actions), and target emotions (absolute arousal levels of 9 emotions after the next regulation cycle). Additionally, the dual-encoder version files include extra steady-state columns at the action moment (hunger, sleep, pain, boredom). The dataset is suitable for artificial life, reinforcement learning, and world model training tasks, particularly for learning perception-action-emotion dynamics of organisms in simulated environments.
数据集概述
数据集名称:DL2L Experiments Dataset
许可证:MIT
语言:英语
任务类别:其他
该数据集包含来自 DL2L 分布式人工生命模拟器的仿真轨迹数据,用于训练 JEPA 世界模型。训练好的模型见 felipedreis/dl2l-jepa。
数据集结构
数据按实验前缀组织,每个前缀包含用于模型训练的 Parquet 文件以及一个 stats.json 元数据文件。
p9/ train.parquet # 单编码器训练集(试验 1–8) val.parquet # 单编码器验证集(试验 9–10) train_dual.parquet # 双编码器训练集(添加 h_t 内稳态列) val_dual.parquet # 双编码器验证集 stats.json # 维度、特征顺序、归一化统计、划分大小
p9 实验详情
- 模拟配置:10 个智能体 × 10 轮试验,DL2L
basic.conf世界 - 数据划分:试验 1–8 → 训练集,试验 9–10 → 验证集(基于试验划分,防止跨试验污染)
- 数据量:训练集 359,782 条,验证集 89,731 条
样本格式
每行是一个 (s_t, a_t, emotion_target) 三元组:
| 列组 | 列名 | 描述 |
|---|---|---|
感知 s_t |
distance, angle, direction |
目标物体空间特征 |
| 物体类型 | type_GRAY_APPLE, type_GREEN_APPLE, type_RED_APPLE, type_ROTTEN_APPLE, type_CACTUS, type_ALOE |
独热编码的物体类型 |
动作 a_t |
a_APPROACH, a_AVOID, a_EAT, a_ESCAPE, a_PLAY, a_SLEEP, a_TOUCH, a_TURN, a_WANDER |
独热编码的选定动作 |
| 目标情绪 | final_hunger, final_sleep, final_apathy, final_stress, final_pain, final_tedium, final_fear, final_curiosity, final_fertility |
下一次调节后的绝对唤醒值 |
双编码器 Parquet 文件额外包含:
| 列名 | 描述 |
|---|---|
ht_hunger, ht_sleep, ht_pain, ht_tedium |
动作时的内稳态状态 (h_t) |
stats.json 关键信息
- 输入维度
input_dim:9 - 动作维度
action_dim:9 - 情绪维度
emotion_dim:9 - 潜在维度
latent_dim:64 - 内状态维度
internal_state_dim:4 - 内潜在维度
internal_latent_dim:16 - 实时情绪维度索引
live_emotion_dims:[0, 1, 4, 5] - 特征顺序:
["distance", "angle", "direction", ...] - 动作顺序:
["APPROACH", "AVOID", "EAT", "ESCAPE", "PLAY", "SLEEP", "TOUCH", "TURN", "WANDER"] - 情绪顺序:
["hunger", "sleep", "apathy", "stress", "pain", "tedium", "fear", "curiosity", "fertility"] - 训练集样本数
n_train:359,782 - 验证集样本数
n_val:89,731
数据提取
原始数据从 PostgreSQL 数据库中提取,命令如下:
bash python3 scripts/pg_extract.py --out /path/to/output --container <db-container>
提取内容涵盖轨迹、睡眠片段、印迹、唤醒历史、行为效率、感知覆盖、移动距离、合并批次统计等。详细列表见 scripts/pg_extract.py。
基于 CSV 输出组装 ML 训练数据集:
bash cd ml python3 -m scripts.prepare_dataset --wd /path/to/output --out data_p9 --dual
引用
DL2L — Distributed Live to Learn, Learn to Live Felipe Duarte dos Reis, CEFET-MG, 2017–2026 https://github.com/felipedreis/dl2l




