ArticuRiddle
收藏资源简介:
ArticuRiddle是一个用于机器人操作中关节物体运动学探索的数据集,专门设计用于评估策略在面对视觉模糊的关节类型时的表现。该数据集包含100个经过特殊编辑的关节物体,每个物体在外观上保持原有资产特征,但其关节运动方式与外观暗示的矛盾:要么通过移动或旋转手柄使其暗示错误的运动方向(handle编辑,50个),要么直接交换关节类型(如将抽屉面改为摆动门,joint交换,50个)。物体按真实关节类型分为door(旋转关节)和drawer(平移关节)两类,每类各50个。每个物体以PartManip环境的标准格式存储,包含URDF模型、纹理、包围盒信息和预抓取状态文件。此外,还提供了一个元数据CSV文件,记录了每个物体的来源资产、编辑类型、原始和修改后的关节类型等详细信息。该数据集适用于机器人操作、关节物体交互、运动学推断等研究任务。
ArticuRiddle is a dataset for kinematic exploration of articulated objects in robotic manipulation, specifically designed to evaluate policy performance when faced with visually ambiguous joint types. It contains 100 specially edited articulated objects, each retaining the original asset appearance while its joint movement contradicts the visual cue: either by moving or rotating the handle to imply a wrong movement direction (handle edit, 50 objects) or by directly swapping the joint type (e.g., changing a drawer face to a swinging door, joint swap, 50 objects). Objects are categorized by their true joint type into door (revolute joint) and drawer (prismatic joint), 50 each. Each object is stored in the standard PartManip environment format, including URDF model, textures, bounding box information, and pre-grasp state files. Additionally, a metadata CSV file records detailed information such as source asset, edit type, original and modified joint types. The dataset is suitable for research tasks such as robotic manipulation, articulated object interaction, and kinematic inference.
ArticuRiddle 数据集概述
基本信息
- 数据集名称:ArticuRiddle
- 数据集规模:n<1K
- 任务类别:robotics
- 标签:robotics、articulated-objects、manipulation、isaac-gym
- 关联论文:Learning to Explore Hidden Kinematics for Articulated Object Manipulation
- 相关资源:
- 网站:https://hiddenkinematics.github.io/
- 论文:https://arxiv.org/abs/2609.36553
- 代码:https://github.com/ariel6905/HiddenKinematics
简介
ArticuRiddle 包含 100 个铰接物体,用于测试在外观视觉歧义运动学条件下的操作任务。每个物体保留现有资产的外观,但其铰接方式与外观所暗示的运动方式相矛盾。因此,仅依赖外观的策略会被误导,必须通过与物体交互来发现其运动方式。
数据集概览
每个物体均派生自 PartManip 的一个训练资产(物体来自 GAPartNet / PartNet-Mobility),通过以下两种编辑之一生成:
| 编辑类型 | 数量 | 变化内容 |
|---|---|---|
handle (cfH) |
50 | 手柄被移动或旋转,使其暗示错误的运动方式;关节保持不变 |
joint (cfS) |
50 | 关节类型被替换(抽屉前板变为摆动、门变为滑动);外观保持不变 |
目录结构
door/<asset>/ 50 objects whose target part is revolute (25 handle edits, 25 joint swaps) drawer/<asset>/ 50 objects whose target part is prismatic (25 handle edits, 25 joint swaps) metadata.csv one row per object: edit and source asset
door/drawer指编辑后的(真实)关节类型,而非外观:door/中的关节替换物体外观看起来像抽屉。- 每个资产文件夹遵循 PartManip 格式(
mobility_new.urdf、textured_objs/、bbox_info.json、part_pregrasp_dof_state.npy),可作为 PartManip 环境的评估划分使用。
来源信息(metadata.csv)
| 列名 | 含义 |
|---|---|
asset |
文件夹名称;源名称中插入了 -cfH / -cfS |
category |
编辑后的类别:door(旋转关节)或 drawer(移动关节) |
edit_family |
handle 或 joint,见上表;edit 给出具体编辑(handle_translate、handle_rotate、joint_swap) |
object_class、partnet_mobility_id |
PartNet-Mobility 类别与物体 id |
source_category、source_split、source_asset |
物体所派生的 PartManip 资产 |
source_joint_type、joint_type |
编辑前与编辑后的目标关节类型 |
许可证
物体派生自 PartNet-Mobility,经由 GAPartNet / PartManip;请遵循这些数据集的使用条款。
引用
bibtex @misc{liu2026learningexplorehiddenkinematics, title={Learning to Explore Hidden Kinematics for Articulated Object Manipulation}, author={Ruiyao Liu and Boshu Lei and Zhuoyang Pan and Kostas Daniellidis}, year={2026}, eprint={2609.36553}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2609.36553}, }





