JJho1314/AC-DiT-MSHab-Dataset
收藏资源简介：
AC-DiT MSHab再现数据集是一个用于移动操作任务的数据集，专为AC-DiT（自适应协调扩散变换器用于移动操作，NeurIPS 2025）基准测试设计，在ManiSkill-HAB（MSHab）Set-Table场景中评估。该数据集包含7个Set-Table子任务（如拾取苹果、拾取碗、放置苹果、放置碗、打开冰箱、打开厨房柜台抽屉、关闭厨房柜台抽屉）的轨迹演示，以HDF5格式存储。数据集包括训练分割（每个任务1000条轨迹，共7000条）和验证分割（每个任务100条轨迹，共700条），所有轨迹均为成功演示。数据内容涵盖机器人动作（13个自由度，包括关节位置、夹持器、基座速度等）、观测数据（如关节位置、基座速度、目标位置、物体姿态、传感器数据如RGB图像、深度图像、分割掩码）以及预计算的自然语言任务描述嵌入（SigLIP文本嵌入）。数据集用于第二阶段模仿学习训练和保留评估，支持AC-DiT模型的再现，并提供了数据使用、下载和点云重建的详细指南。数据集基于MSHab基准的官方演示收集流程生成，使用RL专家策略（SAC用于拾取/放置任务，PPO用于打开/关闭任务），并在Sapien模拟器（通过ManiSkill 3.0）和ReplicaCAD场景资产上运行。
AC-DiT MSHab Reproduction Dataset is a dataset for mobile manipulation tasks, designed for the AC-DiT (Adaptive Coordination Diffusion Transformer for Mobile Manipulation, NeurIPS 2025) benchmark and evaluated on the ManiSkill-HAB (MSHab) Set-Table scenario. It contains trajectory demonstrations (in HDF5 format) for 7 Set-Table subtasks (e.g., pick apple, pick bowl, place apple, place bowl, open fridge, open kitchen counter drawer, close kitchen counter drawer). The dataset includes a training split (1000 trajectories per task, totaling 7000) and a validation split (100 trajectories per task, totaling 700), with all trajectories being successful demonstrations. Data encompasses robot actions (13 degrees of freedom, including joint positions, gripper, base velocity, etc.), observations (e.g., joint positions, base velocity, goal positions, object poses, sensor data such as RGB images, depth images, segmentation masks), and precomputed natural-language task description embeddings (SigLIP text embeddings). It is used for stage-2 imitation learning training and held-out evaluation, supporting the reproduction of the AC-DiT model, and provides detailed guidelines for data usage, download, and point cloud reconstruction. The dataset is generated from scratch using the official demonstration-collection pipeline of the MSHab benchmark, employing RL expert policies (SAC for pick/place tasks, PPO for open/close tasks) and running on the Sapien simulator (via ManiSkill 3.0) with ReplicaCAD scene assets.
