Learning Parameterized Task Structure for Generalization to Unseen Entities

Code for an implementation of Learning Parameterized Task Structure for Generalization to Unseen Entities published at AAAI 2022.

arxiv. project website. video intro.

Installation

conda create -n psgi python=3.8
cd <PSGI directory>
pip install -e .

Scripts

Run the following commands to run the experiments with one of the following environments: cooking, ETmining, or ai2thor.

Random

bash script/run_rl_baselines.sh --algorithm=random --env_id=cooking --graph_param=eval --seed 1

HRL

 bash script/run_hrl.sh --env_id=cooking --graph_param=eval --seed 1

MSGI plus

 bash script/run_msgi_plus.sh --env_id=cooking --graph_param=eval --seed 1

PSGI (no prior graph)

 bash script/run_np_psgi.sh --env_id=cooking --graph_param=eval --seed 1

PSGI

 bash script/meta_train_psgi.sh --env_id=cooking --graph_param=train --seed 1 --exp_id 1 # train and save psgi graphs
 bash script/meta_eval_psgi.sh --env_id=cooking --graph_param=eval --seed 1 --exp_id 1 --load_exp_id 1 # eval and load psgi graphs from train

Cite this work

@inproceedings{liu2022learning,
  title={Learning Parameterized Task Structure for Generalization to Unseen Entities},
  author={Liu, Anthony and Sohn, Sungryull and Qazwini, Mahdi and Lee, Honglak},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={36},
  number={7},
  pages={7534--7541},
  year={2022}
}