This is code for MoCoDA: Model-based Counterfactual Data Augmentation, by Silviu Pitis, Elliot Creager, Ajay Mandlekar, Animesh Garg, published in proceedings of NeurIPS 2022.
Toy Navigation Offline Data is auto generated by the script. Fetch Sweep Offline Data: https://drive.google.com/file/d/1beDmzn2w7yn0MHE7FlW_hJO02rWSJGWw/view?usp=sharing
First run the augmented_offline_X.py script, then the corresponding batchrl_X.py script. See run_toy_experiments.sh for commands for the navigation environment.
For the batchrl scripts, you need to have mrl on your PATH: https://github.com/spitis/mrl (commit 249d652d07bbb5ecc63c2fd37ebd6ec05f4c9607 should work).
Python 3.9 with torch 1.10.1 and numpy 1.20.3 works, but probably other versions do too.