ScissorBot: Learning Generalizable Scissor Skill for Paper Cutting via Simulation, Imitation, and Sim2Real

CoRL 2024

Project Page | Video | arXiv

This repository contains code for both PaperCutting-Sim environments and training scripts for ScissorBot.

Installation

conda create -n scissorbot python==3.8.16
conda activate scissorbot
python setup.py
bash setup.sh
source ~/.bashrc

Simulation

A Simple Example

python paper_cutting_game.py

Generating expert demos

Generate demos with heuristic oracle policy:

python policy/generate_demos_high_level.py \
    setup.cuda=1 demos.curve_type=two_peak  \
    demos.angle_slot=5.73 demos.max_num=1000

You'll get source demos through the last command. Then process demos to render them with (point cloud and action pairs)

python dagger/prepare_offline.py -d SRC_DATA_PATH -o OUTPUT_PATH -g $GPU -p 1 

Training

Train our imitation learning policy

python dagger/bc_delta.py -b 48 \
-tp TRAIN_FOLDER_NAME -ep VAL_FOLDER_NAME \
-lvb 0.07 -s 140000 -y ./config/rl/bc_8d_no_pose.yaml \
--cuda -g $GPU \
-en EXP_NAME \
-m "{'seq_len':4}"

BibTeX

If you find our work useful in your research, please consider citing:

@inproceedings{lyuscissorbot,
  title={ScissorBot: Learning Generalizable Scissor Skill for Paper Cutting via Simulation, Imitation, and Sim2Real},
  author={Lyu, Jiangran and Chen, Yuxing and Du, Tao and Zhu, Feng and Liu, Huiquan and Wang, Yizhou and Wang, He},
  booktitle={8th Annual Conference on Robot Learning}
}

License

This work and the dataset are licensed under CC BY-NC 4.0.