SAC-N-GMM

SAC-N-GMM: Robot Skill Refining and Sequencing for Long-Horizon Manipulation Tasks

Akshay L Chandra, Iman Nematollahi, Tim Welschehold

We present SAC-N-GMM, a single agent that learns to refine and sequence several robot skills to complete tasks.

Policy Rollouts in CALVIN

Below we show rollouts of SAC-N-GMM policy fine-tuned to refine and stitch four and seven skill tasks:

Four Skill Tasks

Seven Skill Tasks

Installation

To begin, clone this repository locally

git clone https://github.com/acl21/sac_n_gmm.git
export SACNGMM_ROOT=$(pwd)/sac_n_gmm

Install requirements:

cd SACNGMM_ROOT
conda create -n sacngmm_venv python=3.8
conda activate sacngmm_venv
sh install.sh

For Development:

pip install -r requirements-dev.txt
pre-commit install

Download

Download the CALVIN dataset and place it inside dataset/.

Robot Skill Repertoire

Step 1: Extract skill demos from the CALVIN dataset

Configure config/demos.yaml.

> python sac_n_gmm/extract_demos.py skill='open_drawer'

Step 2: Train and evaluate skill libraries (Dynamical Systems) with Riepybdlib

Configure config/gmm_train.yaml.

> python sac_n_gmm/scripts/gmm_train.py skill='open_drawer'

Configure config/gmm_eval.yaml.

> python sac_n_gmm/scripts/gmm_eval.py skill='open_drawer'

Train RL Agent

python sac_n_gmm/scripts/sac_n_gmm_train.py

Citation

If you find the code useful, please cite:

SAC-N-GMM

@inproceedings{chandra2024sacngmm,
    author  = {Akshay L Chandra and Iman Nematollahi and Tim Welschehold},
    title   = {SAC-N-GMM: Robot Skill Refining and Sequencing for Long-Horizon Manipulation Tasks},
    booktitle = {Master's Project},
    publisher = {Robot Learning Lab, Freiburg},
    year = 2024,
    url = {https://akshaychandra.com/assets/pdf/masterproject-report.pdf},
}

License

MIT License