arth-shukla/MS-HAB-SetTable

ManiSkill-HAB SetTable Dataset是一个用于ManiSkill-HAB SetTable的全身低级控制/操作演示数据集。每个子任务/对象组合（例如pick 013_apple）有1000个成功的剧集（200个样本/演示），这些剧集是通过RL策略收集的，并经过过滤以确保机器人行为安全，并使用基于规则的事件标签系统进行标注。SetTable包含Pick、Place、Open和Close子任务，相对于其他MS-HAB长期任务（如TidyHouse、PrepareGroceries），SetTable的难度较低，主要挑战在于技能链的结合而非单个子任务。数据集的结构包括每个子任务/对象组合的JSON和HDF5文件，分别包含剧集元数据和演示数据。数据集是纯合成的，使用ManiSkill的默认渲染进行数据生成。

The ManiSkill-HAB SetTable Dataset is a whole-body, low-level control/manipulation demonstration dataset for ManiSkill-HAB SetTable. Each subtask/object combination (e.g., pick 013_apple) has 1000 successful episodes (200 samples/demonstration) gathered using RL policies, filtered for safe robot behavior with a rule-based event labeling system. SetTable contains the Pick, Place, Open, and Close subtasks, which are relatively easy compared to other MS-HAB long-horizon tasks like TidyHouse and PrepareGroceries. The primary challenge of SetTable comes from skill chaining rather than individual subtasks. The dataset structure includes JSON and HDF5 files for each subtask/object combination, containing episode metadata and demonstration data, respectively. The dataset is purely synthetic, using ManiSkills default rendering for data generation.