MISCGrasp

MISCGrasp: Leveraging Multiple Integrated Scales and Contrastive Learning for Enhanced Volumetric Grasping
Paper: https://arxiv.org/abs/2507.02672
Project page: https://miscgrasp.github.io/

This repository provides the research codebase for MISCGrasp, a volumetric 6-DoF grasp detection framework that integrates multi-scale feature learning with contrastive feature enhancement for improved generalization across object geometries and scales.

Note: This project has been developed over a long period of time and the repository may still be roughly organized in places. If it provides you with useful insights, please consider starring the repo.

---

Links

• Project page: https://miscgrasp.github.io/
• Paper (arXiv:2507.02672): https://arxiv.org/abs/2507.02672
• Video: available via the project page

---

Method Overview

Robotic grasping must generalize across objects of diverse shapes and scales. MISCGrasp addresses this by:

• Integrating multi-scale geometric features to balance fine-grained local graspability and global structure.
• Using multi-scale contrastive learning to enhance feature consistency among positive grasp samples across scales.
• Demonstrating performance in both simulation and real-world tabletop decluttering experiments (see project page/paper for details).

---

Repository Structure

High-level layout:

• src/
  Core implementation (models, training/inference utilities, etc.).
• scripts/
  Experiment / evaluation scripts and helpers.
• data_generator/
  Utilities for data generation / preprocessing (e.g., synthetic scene generation, dataset preparation).
• train.sh
  Shell entrypoint for training.
• run_single.sh
  Shell entrypoint for single-case inference / demo.

---

Environment Setup

This codebase is Python-based. Prepare a Python environment suitable for deep learning and 3D processing (e.g., PyTorch + common scientific stack). A typical setup pattern:

1. Create an isolated environment (Conda or venv).
2. Install PyTorch compatible with your CUDA/toolchain.
3. Install remaining dependencies required by src/ and scripts/.

Tip: If you want to make this repo easier for others to run, consider adding a requirements.txt or environment.yml.

---

Quick Start

Training

Run training via:

bash train.sh

Edit train.sh (and the Python entrypoints it calls) to set dataset paths, checkpoints, GPUs, and other options.

Single-run Inference / Demo

Run a single example via:

bash run_single.sh

Adjust paths/options inside run_single.sh as needed.

---

Data

The project introduces and evaluates a dataset containing both power and pinch grasps (see the project page for details). If you distribute dataset assets separately, consider documenting:

• download instructions
• expected directory layout
• preprocessing steps (and optional checksums)

---

Results and Evaluation

MISCGrasp is evaluated in simulated settings and physical tabletop decluttering experiments. For reproducibility, consider adding:

• exact commands for each benchmark/split
• checkpoint links
• scripts to reproduce paper tables/figures

---

Citation

If you use this work in academic research, please cite:

@INPROCEEDINGS{fan2025miscgrasp,
  author={Fan, Qingyu and Cai, Yinghao and Li, Chao and Jiao, Chunting and Zheng, Xudong and Lu, Tao and Liang, Bin and Wang, Shuo},
  booktitle={2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={MISCGrasp: Leveraging Multiple Integrated Scales and Contrastive Learning for Enhanced Volumetric Grasping}, 
  year={2025},
  volume={},
  number={},
  pages={11335-11342},
  keywords={Shape;Focusing;Grasping;Contrastive learning;Transformers;Feature extraction;Intelligent robots;Faces},
  doi={10.1109/IROS60139.2025.11246166}
}

---

Acknowledgements

• VGN and related volumetric grasping literature/tools.
• Any third-party libraries, datasets, and upstream repos used in this project (please follow their licenses and citation requirements).

---

Contact

For questions, suggestions, or bug reports, please open an Issue on GitHub.