GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation

Julien Mérand¹,    Boris Meden¹,    Mathieu Grossard¹,    Liming Chen²

¹Université Paris-Saclay, CEA-List    ²École Centrale de Lyon, CNRS, LIRIS, UMR5205, Institut Universitaire de France (IUF)

2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

Project Page   |   arXiv

GOAG is novel deep generative model that learns a compact latent representation of a specific gripper's contact surface distribution, enabling the efficient sampling of valid grasp configurations without relying on object-specific training data. By introducing object features only at inference time, GOAG can effectively retrieve admissible contact areas that are compatible with the gripper’s capabilities.

See https://github.com/CEA-LIST/GOAG for full code and instructions.

Your directory tree should strictly follow this structure:

# Model Checkpoints
GOAG                        # Main folder
├── ...
├── logs                    # ckpts folder
    ├── allegro_cvae
    ├── allegro_pointnet
    ├── barrett_cvae
    ├── barrett_pointnet
    ├── shadowhand_cvae
    └── shadowhand_pointnet
└── ...

# Dataset Directory
GOAG_DATA
├── handprints
├── pointclouds
    ├── dexgrab
    ├── dexgraspnet
    ├── multidex
    ├── realdex
    └── unidexgrasp
├── urdf
    ├── objects
    │   ├── dexgrab
    │   ├── dexgraspnet
    │   ├── multidex
    │   ├── realdex
    │   └── unidexgrasp
    └── robot
        ├── allegro
        ├── barrett
        └── shadowhand
└── workspaces