neurips-submission-18302/dcs-submission-18302

该数据集是一个用于强化学习或机器人控制任务的视觉数据集，包含8个配置，每个配置对应不同的模拟环境任务：猎豹奔跑（cheetah_run）、跳跃者跳跃（hopper_op）、人形行走（humanoid_walk）和步行者奔跑（walker_run），每个任务都有基础版本和带有低干扰物（distractor_low）的变体。数据集由图像观察（observation）、状态向量（state）、掩码图像（mask）、动作向量（action）、奖励（reward）、终止标志（terminated）、截断标志（truncated）和预测掩码（pred_mask）等特征组成，用于训练和测试AI代理在复杂环境中的决策能力。每个配置分为训练集和测试集，训练集包含900万个样本，测试集包含100万个样本，总数据量从约56GB到81GB不等，支持大规模机器学习实验。This dataset is a visual dataset for reinforcement learning or robotics control tasks, comprising 8 configurations, each corresponding to different simulated environment tasks: cheetah_run, hopper_hop, humanoid_walk, and walker_run, with each task having a base version and a variant with low distractors (distractor_low). The dataset includes features such as image observations, state vectors, mask images, action vectors, rewards, termination flags, truncation flags, and prediction masks, designed to train and test AI agents for decision-making in complex environments. Each configuration is split into training and test sets, with 9 million samples in the training set and 1 million samples in the test set, totaling data sizes ranging from approximately 56GB to 81GB, supporting large-scale machine learning experiments.