Geomatch-trained-model

GeoMatch Trained Model

This repository contains the final trained GeoMatch model checkpoint developed during my internship at the Human Interactive Robotics (HIRo) Laboratory, IISc Bangalore.

Overview

GeoMatch is a Graph Neural Network (GNN)-based framework for robotic grasp generation. Given an object and a robotic hand, the model predicts feasible grasp poses that can subsequently be evaluated using a physics simulator.

This repository contains the final trained model checkpoint used for inference.

---

Model File

grasp_gnn.pth

---

Usage

1. Clone the GeoMatch repository:

https://github.com/google-deepmind/geomatch

1. Download grasp_gnn.pth from this repository.

2. Replace the default checkpoint (or place it in the appropriate checkpoint directory).

3. Run the GeoMatch inference pipeline to generate grasp predictions.

---

Related Repository

The code, documentation, bridge implementation, and project handover are available in my GitHub fork:

https://github.com/Saanvi-Mehra/Geomatch

---

Internship Contributions

The work completed during this internship includes:

• Training the GeoMatch model
• Generating grasp predictions for unseen objects
• Developing a bridge between GeoMatch and GenDexGrasp
• Integrating GeoMatch with the GenDexGrasp evaluation framework in Isaac Gym
• Developing utilities for grasp visualization
• Documenting the complete workflow for future contributors

Note: GenDexGrasp was used only as a physics-based evaluation framework. No modifications were made to its core learning pipeline.

---

Acknowledgements

This work is based on the original GeoMatch framework developed by Google DeepMind.

Original repository:

https://github.com/google-deepmind/geomatch

Please cite the original GeoMatch work if you use this model.

---

License

This model is released under the Apache-2.0 License, consistent with the original GeoMatch repository.