GAM: General Affordance-based Manipulation for Robotic Object Disentangling Tasks

• Paper

Installation guide

• Create conda env with the provided environment.yml file.
• Install graspnet-baseline.
• Install mujoco and mujoco-py (the experiments use 2.1.0 version, others may work).
• Make sure gym version <=0.20.0.
• From repository root, run python -m pip install ..

Data

• Follow https://github.com/IanYangChina/GAM-paper-data to download the data.
• Place the SG_data folder in the root directory of this repo.

Pretrained agents

• There are some pre-trained checkpoints that can be downloaded onedrive.
• Place the result folders in the root directory (e.g., GAM-paper-codes/results_3).
• Namings:
  • C/C+/S: hook shapes
  • gf: grasp filter checkpoints
  • hm: hemisphere action checkpoints
  • 3/30: num of episode timesteps

Scripts

• With the data and checkpoints, you may play around with evaluations.
  • E.g., python script/evaluate_dqn --render --nenv 1 --hs C --nh 3
• Other scripts for training, generating dataset are also available.

Citation

@article{yang2024gam,
  title={GAM: General affordance-based manipulation for contact-rich object disentangling tasks},
  author={Yang, Xintong and Wu, Jing and Lai, Yu-Kun and Ji, Ze},
  journal={Neurocomputing},
  pages={127386},
  year={2024},
  publisher={Elsevier}
}