vsislab/SMG-multimodal-grasping

Hybrid Robotic Grasping with a Soft Multimodal Gripper and a Deep Multistage Learning Scheme

Hybrid Robotic Grasping with a Soft Multimodal Gripper and a Deep Multistage Learning Scheme

IEEE Transactions on Robotics (TRO) 2023 [Paper]

Fukang Liu^1,2,Fuchun Sun²,Bin Fang²,Xiang Li²,Songyu Sun³,Huaping Liu²

¹Carnegie Mellon University
²Tsinghua University
³ University of California, Los Angeles

Contact

If you have any questions, please let me know: Fukang Liu fukangl[at]andrew[dot]cmu[dot]edu

Installation

The implementation requires the following dependencies:

• Python, PyTorch, NumPy, OpenCV-Python, SciPy, Matplotlib, CoppeliaSim
  

Instructions

1. The soft multimodal gripper was developed based on our previous work, please find more details: "Multimode Grasping Soft Gripper Achieved by Layer Jamming Structure and Tendon-Driven Mechanism".

2. Checkout this repository and download the datasets. UnZip it and put the unzipped file to the /code directory.

3. Run CoppeliaSim (navigate to your CoppeliaSim directory and run ./coppeliaSim.sh). From the main menu, select File > Open scene..., and open the file code/simulation/simulation-lc.ttt(lightly-cluttered) or simulation-hc.ttt(highly-cluttered) from this repository. Choose the Vortex physics engine for simulation (you can also choose other physics engines that CoppeliaSim supports (e.g., Bullet, ODE), but Vortex works best for the simulation model of SMG).

4. In another terminal window, run the following example:

python main.py --is_sim --method 'reinforcement' --is_ets --is_pe --is_oo --explore_rate_decay

Training

To train an Reactive Enveloping and Sucking Policy (E+S Reactive) in simulation with lightly cluttered environment, run the following:

python main.py --is_sim --method 'reactive' --is_pe --is_oo --explore_rate_decay

To train an Reactive Enveloping, Sucking and Enveloping_then_Sucking Policy (E+S+ES Reactive) in simulation with lightly cluttered environment, run the following:

python main.py --is_sim --method 'reactive' --is_ets --is_pe --is_oo --explore_rate_decay

To train a DRL Enveloping and Sucking Policy (E+S DRL) in simulation with lightly cluttered environment, run the following:

python main.py --is_sim --method 'reinforcement' --is_pe --is_oo --explore_rate_decay

To train a DRL multimodal grasping policy (E+S+ES DRL(PE+OO)) in simulation with lightly cluttered environment, run the following:

python main.py --is_sim --method 'reinforcement' --is_ets --is_pe --is_oo --explore_rate_decay

Tranining policies in highly cluttered environment, add --is_cluttered. For example, For training a DRL multimodal grasping policy (E+S+ES DRL(PE+OO)) in simulation with highly cluttered environment, run the following:

python main.py --is_sim --is_cluttered --method 'reinforcement' --is_ets --is_pe --is_oo --explore_rate_decay

Evaluation

To test your own pre-trained model, simply change the location of --snapshot_file. For example, for testing the pre-trained E+S+ES DRL(PE+OO) model in simulation with lightly cluttered environment, run the following:

python main.py --is_sim --method 'reinforcement' --is_ets --is_pe --is_oo --explore_rate_decay \
--is_testing \
--load_snapshot --snapshot_file 'YOUR-SNAPSHOT-FILE-HERE'

To test the three ablation baselines, remove --is_pe or --is_oo. For testing a DRL multimodal grasping policy that executes actions without either preenveloping or orientation optimization (E+S+ES DRL) in simulation with lightly cluttered environment, remove both --is_pe and --is_oo:

python main.py --is_sim --method 'reinforcement' --is_ets --explore_rate_decay \
--is_testing \
--load_snapshot --snapshot_file 'YOUR-SNAPSHOT-FILE-HERE'

For testing a DRL multimodal grasping policy that executes actions with only preenveloping (E+S+ES DRL(PE)) in simulation with lightly cluttered environment, remove --is_oo:

python main.py --is_sim --method 'reinforcement' --is_ets --is_pe --explore_rate_decay \
--is_testing \
--load_snapshot --snapshot_file 'YOUR-SNAPSHOT-FILE-HERE'

To test a DRL multimodal grasping policy that executes actions with only orientation optimization (E+S+ES DRL(OO)) in simulation with lightly cluttered environment, remove --is_pe:

python main.py --is_sim --method 'reinforcement' --is_ets --is_oo --explore_rate_decay \
--is_testing \
--load_snapshot --snapshot_file 'YOUR-SNAPSHOT-FILE-HERE'

Bibtex

If you find the code or gripper design useful, please cite:

@article{liu2023hybrid,
    title={Hybrid Robotic Grasping with a Soft Multimodal Gripper and a Deep Multistage Learning Scheme},
    author={Liu, Fukang and Fang, Bin and Sun, Fuchun and 
    Li, Xiang and Sun, Songyu and Liu, Huaping},
    journal={IEEE Transactions on Robotics},
    year={2023},
    publisher={IEEE}
}

and

@article{fang2022multimode,
  title={Multimode grasping soft gripper achieved by layer jamming structure and tendon-driven mechanism},
  author={Fang, Bin and Sun, Fuchun and Wu, Linyuan and Liu, Fukang and 
  Wang, Xiangxiang and Huang, Haiming and Huang, Wenbing and Liu, Huaping and Wen, Li},
  journal={Soft Robotics},
  volume={9},
  number={2},
  pages={233--249},
  year={2022}
}

Acknowledgements

This code was developed using visual-pushing-grasping.