This code has been developed to reproduce the PUBC methodology described in our paper titled Improving Behavioural Cloning with Positive Unlabeled Learning. We have included two of the most challenging real-world physical robotic manipulation tasks in this repository for demonstration.
pip install -r requirements.txt
Firstly, you need to visit THIS LINK to download the datasets to your local device.
If you want only run the PU learning(filter part) without policy learning:
python main.py --raw-dataset-path='<path to the mixed dataset>' --pos-seed-dataset-path='<path to the seed dataset>' --train-policy=False
If you want only run full PUBC:
python main.py --raw-dataset-path='<path to the mixed dataset>' --pos-seed-dataset-path='<path to the seed dataset>' --train-policy=True --policy='bc'
If you want to use the trained filter for traing BC:
python main.py --raw-dataset-path='<path to the mixed dataset>' --pos-seed-dataset-path='<path to the seed dataset>' --load-trained-filter=True --trained-filter-path=='<path to the trained models>' --ckpt-iterations=<iterations number for training the trained models> --train-policy=True --policy='bc'
To submit the trained policy model for evaluation on a real robot cluster, please follow the instructions provided on this page: https://webdav.tuebingen.mpg.de/trifinger-rl/docs/real_robot/submission_system.html