How georgewuzy/cedex-grasp works
Project Type and Purpose
• This is a research-oriented, full-pipeline software project (not a web app) for large-scale dexterous-grasp synthesis.
• Core purpose: generate human-like contact representations, remap them to multiple robotic hands, create millions of grasps (CEDex dataset), validate them in simulation, and provide visualization tools.
File-Structure Analysis
• Top-level files:
• README.md, LICENSE, requirements.txt, .gitignore
• assets/ contains pipeline GIFs for documentation
• cedex/ stores precomputed grasp datasets per gripper (.pt files)
• data/ holds URDF descriptions and metadata for hands/objects
• contactgen/ is a self-contained Python package for contact-map generation
• pointnet_lib/ is a CUDA-accelerated extension for PointNet2 used by contactgen
• utils/ and utils_model/ supply helper functions, hand kinematics, optimizers
• validation/ contains scripts and modules for Isaac Gym–based grasp evaluation
• CLI scripts at root: inf_contactdb.py, inf_ycb.py, generate_data.py, eval_grasp.py, eval_grasp_filtered.py, vis_cedex.py, vis_generated_grasp.py
• Architectural patterns:
• Modular, layered pipeline: Data → Contact Generation → Grasp Generation → Validation → Visualization
• contactgen package uses PyTorch + pointnet_lib
• generate_data binds contact outputs with optimization routines in utils_model
• validation leverages NVIDIA Isaac Gym (external simulator)
Key README Insights
• Two-stage grasp creation:
Stage 1: generate human-like contact maps (contactgen/ + pointnet_lib)
Stage 2: optimize robot hand poses on contact maps (generate_data.py + utils_model)
Then validate (eval_grasp.py) and filter (eval_grasp_filtered.py) in simulation.
• External dependencies: PyTorch, PyTorch3D, Isaac Gym
• Dataset artifacts: cedex/*.pt, plus download links for real/synthetic objects