ArticuRiddle

The dataset of Learning to Explore Hidden Kinematics for Articulated Object Manipulation

🌐 Website | 📖 Paper | 💻 Code

⭐ Introduction

Appearance is an unreliable cue for articulation: a cabinet front can look like a drawer and swing like a door, and a handle can sit where it suggests the wrong hinge. ArticuRiddle is a set of 100 articulated objects built to test manipulation under such visually ambiguous kinematics. Each object keeps the appearance of an existing asset while its articulation contradicts what the appearance implies, so a policy that relies on appearance alone is misled and must interact with the object to discover how it moves.

📊 Dataset Overview

Every object is derived from a training asset of PartManip (objects from GAPartNet / PartNet-Mobility) by one of two edits:

edit	count	what changes
handle (cfH)	50	the handle is moved or rotated so that it suggests the wrong motion; the joint is unchanged
joint (cfS)	50	the joint type is swapped (a drawer front that swings, a door that slides); the appearance is unchanged

📁 Structure

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 refer to the edited (true) joint type, not to the appearance: the joint-swapped objects in door/ look like drawers. Each asset folder follows the PartManip format (mobility_new.urdf, textured_objs/, bbox_info.json, part_pregrasp_dof_state.npy) and can be used as an evaluation split of the PartManip environment.

🔍 Provenance (metadata.csv)

column	meaning
asset	folder name; the source name with -cfH / -cfS inserted
category	door (revolute) or drawer (prismatic) after the edit
edit_family	handle or joint, see the table above; edit gives the specific edit (handle_translate, handle_rotate, joint_swap)
object_class, partnet_mobility_id	PartNet-Mobility class and object id
source_category, source_split, source_asset	the PartManip asset the object is derived from
source_joint_type, joint_type	target-joint type before and after the edit

📄 License

The objects are derived from PartNet-Mobility via GAPartNet / PartManip; please follow the terms of use of those datasets.

📝 Citation

@misc{liu2026learningexplorehiddenkinematics,
      title={Learning to Explore Hidden Kinematics for Articulated Object Manipulation},
      author={Ruiyao Liu and Boshu Lei and Zhuoyang Pan and Kostas Daniilidis},
      year={2026},
      eprint={2609.36553},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2609.36553},
}