JiayiChenPKU/BODex

[ICRA 2025] BODex: Scalable and Efficient Robotic Dexterous Grasp Synthesis Using Bilevel Optimization.

Project page | Paper | Grasp synthesis code | Grasp validation code | Object pre-processing code

Dataset structure:

object_assets
|- DGN_obj_processed.zip  # Our pre-proposed mesh and info files. Used for grasp synthesis and validation.
|- DGN_obj_split.zip      # JSON files containing all valid object id. Used for grasp synthesis, validation, and training.
|- DGN_obj_raw.zip        # The raw object meshes before pre-processing. Please see https://github.com/JYChen18/MeshProcess for our pre-processing methods.
|_ DGN_obj_vision.zip     # Our rendered single-view point clouds. Used for training networks. It is also generated by https://github.com/JYChen18/MeshProcess.

synthesized_grasps        # Successful grasps validated in MuJoCo using https://github.com/JYChen18/DexGraspBench. 
|- allegro.zip            # The hand and object are both floating. No table.
|- leap.zip
|- shadow.zip 
|_ ur10e_shadow.zip       # Each grasp also includes a collision-free approaching trajectory with the table from a fixed initial state. The Shadow Hand is mounted on a UR10e arm.