jacta-manipulation

  

Jacta: A Versatile Planner for Learning Dexterous and Whole-body Manipulation

Robotic manipulation is challenging and data-driven approaches typically require large amounts of data or expert demonstrations. Therefore, we introduce a motion planner for dexterous and whole-body manipulation tasks. The planner’s demonstrations can directly be used by reinforcement learning. With this approach, we can efficiently learn policies for complex manipulation tasks, where traditional reinforcement learning alone only makes little progress.

Installation

Install cmake

sudo apt install cmake

Install mujoco_extensions

git clone https://github.com/bdaiinstitute/jacta-manipulation.git
cd jacta-manipulation
pip install src/mujoco_extensions -vv

Install jacta-manipulation

pip install -e .

Getting started

python examples/planner/example_notebook.py

Open the visualizer in your browser by clicking on the link in the terminal.

http://localhost:8080/

Citation

@inproceedings{brudigam2024jacta,
  author       = {Br{\"u}digam, Jan and Abbas, Ali-Adeeb and Sorokin, Maks and Fang, Kuan and Hung, Brandon and Guru, Maya and Sosnowski, Stefan and Wang, Jiuguang and Hirche, Sandra and Le Cleac'h, Simon},
  editor       = {Agrawal, Pulkit and Kroemer, Oliver and Burgard, Wolfram},
  title        = {Jacta: {A} Versatile Planner for Learning Dexterous and Whole-body Manipulation},
  booktitle    = {Conference on Robot Learning, 6-9 November 2024, Munich, Germany},
  series       = {Proceedings of Machine Learning Research},
  volume       = {270},
  pages        = {994--1020},
  publisher    = {{PMLR}},
  year         = {2024},
  url          = {https://proceedings.mlr.press/v270/bruedigam25a.html},
}

Run tests locally

In the virtual environment:

pip install -e .[dev]
python -m pytest

you might have to

unset PYTHONPATH