FFHNet (ICRA 2022 Paper,video) is an ML model which can generate a wide variety of high-quality multi-fingered grasps for unseen objects from a single view.

Generating and evaluating grasps with FFHNet takes only 30ms on a commodity GPU. To the best of our knowledge, FFHNet is the first ML-based real-time system for multi-fingered grasping with the ability to perform grasp inference at 30 frames per second (FPS).

For training, we synthetically generate 180k grasp samples for 129 objects. We are able to achieve 91% grasping success for unknown objects in simulation and we demonstrate the model's capabilities of synthesizing high-quality grasps also for real unseen objects.

Create a new conda environment with cudatoolkit 10.1

conda create -n myenv python==3.8
conda install -c anaconda cudatoolkit=10.1

Install all dependencies.

pip install -r requirements.txt
pip install git+https://github.com/otaheri/chamfer_distance
pip install git+https://github.com/otaheri/bps_torch

Now there is some problem with LFS. If you have trouble clone the repo, please run

GIT_LFS_SKIP_SMUDGE=1 git clone repo_link

The model is located at here link. After download, extract to repo root path.

You can download here link

python eval.py

Data distribution from FFHGenerator	Filter grasps with 0.5 thresh	Filter grasps with 0.75 thresh

Filter grasps with 0.9 thresh	Best grasp

If you found FFHNet useful in your research, please consider citing:

@INPROCEEDINGS{2022ffhnet,
  author={Mayer, Vincent and Feng, Qian and Deng, Jun and Shi, Yunlei and Chen, Zhaopeng and Knoll, Alois},
  booktitle={2022 International Conference on Robotics and Automation (ICRA)},
  title={FFHNet: Generating Multi-Fingered Robotic Grasps for Unknown Objects in Real-time},
  year={2022},
  volume={},
  number={},
  pages={762-769},
  doi={10.1109/ICRA46639.2022.9811666}}

bps_torch