Cluttered Pushing

This repository contains the accompanying code for the paper Learning Goal-Oriented Non-Prehensile Pushing in Cluttered Scenes by N. Dengler, D. Großklaus and M. Bennewitz submitted for IROS, 2022. you can find the paper at http://arxiv.org/abs/2203.02389

Setup

Step 1: Clone the repository

cd 
git clone https://github.com/NilsDengler/cluttered-pushing.git

Step 2: Create a virtual environment

cd cluttered-pushing
conda env create -f environment.yml -n <env_name>

conda activate <env_name>

Step 3: Install the package

Dependencies:

cd cluttered-pushing/push_gym/push_gym/utils/Lazy_Theta_with_optimization_any_angle_pathfinding

mkdir build && cd build

cmake ..

make

Package:

cd cluttered-pushing/push_gym

pip install -e .

Usage

Change directory to cluttered-pushing/Networks/RL/scripts

cd cluttered-pushing/RL

Training

• To train an RL-agent, customize the parameters given in scripts/parametes.yaml.

• Set the train: True in scripts/parametes.yaml.

• Check or change the network's hyperparameter in scripts/train_agent_script.

• A VAE model is required for training, please refere to VAE Readme to train a VAE model or download already trained models.

• To start training:

  python run_agent.py

Testing

• To evaluate a trained agent, set train: False in scripts/parametes.yaml.
• per default the testing uses log_dir_name: "../Logs/example_agent/model_test/" as specified in scripts/parametes.yaml. Please note, that this is an example agent and not the agent used to reproduce the results of the paper.
• Testing:

  python scripts/run_agent.py

Baseline

• To run the Baseline by Krivic and Piater, set train: False and test_baseline: True in scripts/parametes.yaml.
• Testing:

  python scripts/run_agent.py

For more information, please refer the README in cluttered-pushing/RL.