xz992 / graspgpt_public

code implementation of GraspGPT and FoundationGrasp

An Official Code Implementation of GraspGPT & FoundationGrasp

GraspGPT: Leveraging Semantic Knowledge from a Large Language Model for Task-Oriented Grasping (RA-L 2023)

[paper] [website] [video]

Chao Tang, Dehao Huang, Wenqi Ge, Weiyu Liu, Hong Zhang
Southern University of Science and Technology, Georgia Institute of Technology

This is a Pytorch implementation of GraspGPT. If you find this work useful, please cite:

@article{tang2023graspgpt,
  title={GraspGPT: Leveraging Semantic Knowledge from a Large Language Model for Task-Oriented Grasping},
  author={Tang, Chao and Huang, Dehao and Ge, Wenqi and Liu, Weiyu and Zhang, Hong},
  journal={IEEE Robotics and Automation Letters},
  year={2023},
  publisher={IEEE}
}

We have borrowed tons of code from GCNGrasp. Please also consider citing their work:

@inproceedings{murali2020taskgrasp,
  title={Same Object, Different Grasps: Data and Semantic Knowledge for Task-Oriented Grasping},
  author={Murali, Adithyavairavan and Liu, Weiyu and Marino, Kenneth and Chernova, Sonia and Gupta, Abhinav},
  booktitle={Conference on Robot Learning},
  year={2020}
}

Note that we provide a light-weight version of GraspGPT in this repo. It is with faster inference speed and easier to train. Also, the model achieves comparable performance to the one described in our paper.

Installation

Our code is tested on Ubuntu 20.04, python 3.7, cuda 11.3, pytorch 1.11.0, and pytorch lightning 0.7.1.

1. Create a virtual env or conda environment with python 3.7:

conda create --name graspgpt python=3.7
conda activate graspgpt

2. Git clone the repo to your machine:

git clone https://github.com/mkt1412/GraspGPT_public.git

3. Install dependencies:

cd GraspGPT_public
conda install pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Dataset

The Language Augmented TaskGrasp (LA-TaskGrasp) dataset is developed based on TaskGrasp dataset. To perform training and evaluation on the LA-TaskGrasp dataset, download the dataset here and place it in the root folder as data:

unzip ~/Downloads/data.zip -d ./
rm ~/Downloads/data.zip

To visualize the collected point clouds and labeled task-oriented grasps, please refer to the github repo of TaskGrasp dataset under Usage section.

To run any of the demo scripts below, download the pre-trained models here and put them in the checkpoints folder.

Demo

We provide two types of demos:

• Database version, where GraspGPT uses the pre-generated natural language descriptions for task-oriented grasp pose prediction.
• Interactive version, where GraspGPT interacts with an LLM via prompts to generate language descriptions for task-oriented grasping.

Pre-trained models for each class and task split are provided:

gcngrasp_split_mode_o_split_idx_0_.yml
gcngrasp_split_mode_o_split_idx_1_.yml
gcngrasp_split_mode_o_split_idx_2_.yml
gcngrasp_split_mode_o_split_idx_3_.yml

gcngrasp_split_mode_t_split_idx_0_.yml
gcngrasp_split_mode_t_split_idx_1_.yml
gcngrasp_split_mode_t_split_idx_2_.yml
gcngrasp_split_mode_t_split_idx_3_.yml

Database version

Objetc class: pan, task: pour

python gcngrasp/demo_db.py --cfg_file cfg/eval/gcngrasp/gcngrasp_split_mode_o_split_idx_0_.yml --obj_name pan --obj_class saucepan --task pour

Objetc class: spatula, task: scoop

python gcngrasp/demo_db.py --cfg_file cfg/eval/gcngrasp/gcngrasp_split_mode_o_split_idx_0_.yml --obj_name spatula --obj_class spatula --task scoop

Objetc class: mug, task: drink

python gcngrasp/demo_db.py --cfg_file cfg/eval/gcngrasp/gcngrasp_split_mode_o_split_idx_0_.yml --obj_name mug --obj_class mug --task drink

Interactive version

To run interactive version on your machine, you need to set your OpenAI API key in gcngrasp/data_specificatgion.py line 116.

OPENAI_API_KEY = "xx-xxxxxx"   

Objetc class: pan, task: pour

python gcngrasp/demo_llm.py --cfg_file cfg/eval/gcngrasp/gcngrasp_split_mode_o_split_idx_0_.yml --obj_name pan --obj_class saucepan --task pour

Objetc class: spatula, task: scoop

python gcngrasp/demo_llm.py --cfg_file cfg/eval/gcngrasp/gcngrasp_split_mode_o_split_idx_0_.yml --obj_name spatula --obj_class spatula --task scoop

Objetc class: mug, task: drink

python gcngrasp/demo_llm.py --cfg_file cfg/eval/gcngrasp/gcngrasp_split_mode_o_split_idx_0_.yml --obj_name mug --obj_class mug --task drink

Training

To train GraspGPT from scratch:

python gcngrasp/train.py --cfg_file cfg/train/gcngrasp/gcngrasp_split_mode_t_split_idx_3_.yml

You can replace t_split_idx_3 with other split numbers (i.e., 0, 1, 2) or object class split (i.e., o_split_x).

Evaluation

To evaluate a pre-trained model:

python gcngrasp/eval.py cfg/eval/gcngrasp/gcngrasp_split_mode_t_split_idx_3_.yml --save

Similarly, feel free to try out different split numbers and object class split.

FoundationGrasp: Generalizable Task-Oriented Grasping with Foundation Models

[preprint] [website] [video]

Chao Tang, Dehao Huang, Wenlong Dong, Ruinian Xu, Hong Zhang

Since FoundationGrasp is an extension of GraspGPT, we do not provide a separate GitHub repo for the code implementation and data of FoundationGrasp. If you find FoundationGrasp useful in your research, please cite:

@article{tang2024foundationgrasp,
  title={FoundationGrasp: Generalizable Task-Oriented Grasping with Foundation Models},
  author={Tang, Chao and Huang, Dehao and Dong, Wenlong and Xu, Ruinian and Zhang, Hong},
  journal={arXiv preprint arXiv:2404.10399},
  year={2024}
}

Data

Step 1. Download the original LA-TaskGrasp dataset here and place it in the root folder as data.
Step 2. Download the updated language description data of the LaViA-TaskGrasp dataset here and replace the old one in the LA-TaskGrasp dataset (i.e., obj_gpt_vx, task_gpt_vx).
Step 3. For language instruction data, please refer to data/task_ins_v2.
Step 4. For image data, please refer to data/rgb_images.
Step 5. Modify the dataloader in gcngrasp/data/GCNLoader.py so that the model can read the updated language description and image data.

Code Implementation

Step 1. The semantic branch of FoundationGrasp is similar to GraspGPT. Feel free to use to the code implementation of GraspGPT provided above.
Step 2. The geometric branch of FoundationGrasp is largely inspired by XMFNet. Please refer to XMFnet/model.py to reproduce the geometric branch.