Geometry-aware RL for Manipulation of Varying Shapes and Deformable Objects

⭐ Accepted at ICLR 2025 (Oral) ⭐

🚀 Getting Started

To get started with GeometryRL, follow the steps below:

0. 🛠️ Preliminaries:
   The default workspace structure looks like this:

   ├── ...
   ├── repos                  # Root workspace
   │   ├── geometry_rl        # geometry_rl repo
   │   ├── geometry_orbit     # geometry_orbit repo
   │   ├── others             # other repos
   └── ...
   

   Note: You can change your workspace structure by modifying docker/.env. Then, docker-compose will build an image with the context pointing to the Root workspace.

1. 📂 Clone the required repos:
   Clone geometry_orbit and ITPAL:

   git clone git@github.com:thobotics/geometry_orbit.git
   

   git clone git@github.com:ALRhub/ITPAL.git
   

   Place them into the correct workspace structure.

2. 🐳 Setup Docker:
   Start the Docker container:

   ./docker/container.sh start
   

3. 🏗️ Inside the container (enter via ./docker/container.sh enter):
   Run the following commands:

   orbit -p examples/torchrl/train.py -cn rigid_insertion_multi_hepi_trpl_cfg simulator.headless=True

   orbit -p examples/torchrl/play.py -cn rigid_insertion_multi_hepi_trpl_cfg --checkpoint_name model_checkpoint_best.pth

📖 Citation

@inproceedings{
     hoang2025geometryaware,
     title={Geometry-aware {RL} for Manipulation of Varying Shapes and Deformable Objects},
     author={Tai Hoang and Huy Le and Philipp Becker and Vien Anh Ngo and Gerhard Neumann},
     booktitle={The Thirteenth International Conference on Learning Representations},
     year={2025},
     url={https://openreview.net/forum?id=7BLXhmWvwF}
}