Learning Rational Subgoals from Demonstrations and Instructions

This repository is the implementation of "Learning Rational Subgoals from Demonstrations and Instructions" published at AAAI 2023.

Requirements

Environment has been specified by rsg.yml

Training

To train the model(s) in the paper, run this command:

./scripts/train_all.sh

Evaluation

To evaluate the models (including classification, and planning with instructions/final goal), run:

./scripts/eval_all.sh

Datasets

The datasets will be automatically generated when training models. To force re-generate datasets, add argument --force_regen in scripts/train_all.py or scripts/eval_all.py when calling projects/rsg/scripts/learn_classifier.py.