Learning to Compose Hierarchical Object-Centric Controllers for Robotic Manipulation

This code is associated with our CoRL'20 paper.
[Paper] [Project Website]

Dependencies

Since our code was developed on a previous beta-version of Nvidia Isaac Gym, it may not run with the existing open source version. However, it maybe be possible to port over the environments to the new version. Our code further relies on an earlier version of isaacgym-utils, although for it most of the API changes should be consistent.

Training code

The training code is provided in the scripts folder, it contains examples to run the door env as well as the hex screw environment. Their corresponding config files can be found in the cfg folder. Most params in the config files should be correct but some may neeed a bit of tweaking.

An example command to run the code would be

python ./scripts/train_franka_door.py --seed 101 --cfg ./cfg/run_franka_door_open.yaml --logdir /tmp/franka_door --train 1

Code

• Environments: All the environments used in the paper are defined in object_axes_ctrls/envs.
• Task-Axes Controllers: The controllers and their composition is defined in object_axes_ctrlrs/controllers/projected_axes_attractors.py. Each controller is parameterized with respect to the environment and created in the coresponding environment class.
• Expanded-MDP: The expanded MDP formulation is implemented in envs/env_wrappers.py.

Assets

The assets folder contains the URDF assets (and meshes etc.) for creating and rendering environments.

Citation

In case you find our paper or code useful please consider citing:

@inproceedings{sharma2020learning,
  title={Learning to Compose Hierarchical Object-Centric Controllers for Robotic Manipulation},
  author={Sharma, Mohit and Liang, Jacky and Zhao, Jialiang and LaGrassa, Alex and Kroemer, Oliver},
  booktitle={arXiv preprint arXiv:2011.04627},
  year={2020}
}