README.md

Goal-Conditioned Predictive Coding for Offline Reinforcement Learning

[Project page] [arXiv] [Setup] [Usage] [BibTex]

Setup

Install MuJoCo if it is not already the case:

1. Download MuJoCo binary here (mujoco-py requires MuJoCo 2.1.0)
2. Unzip the downloaded archive into ~/.mujoco/
3. Append the MuJoCo subdirectory bin path into the env variable LD_LIBRARY_PATH

Install dependencies:

conda create -n gcpc python=3.7
codna activate gcpc
pip install -r requirements.txt
pip install -e .

To install Antmaze-Ultra, please refer to link

Usage

Data preprocessing:

python -m gcpc.data.parse_d4rl

First Stage: Trajectory Representation Learning

Antmaze

python -m gcpc.train env_name=antmaze-large-play-v2 model=trajnet exp=antmaze_trl model.model_config.mask_type=mae_rc

Kitchen

python -m gcpc.train env_name=kitchen-mixed-v0 model=trajnet exp=kitchen_trl model.model_config.mask_type=mae_rc

Gym

python -m gcpc.train env_name=halfcheetah-medium-expert-v2 model=trajnet exp=gym_trl model.model_config.mask_type=mae_rc 

Second Stage: Policy Learning

Antmaze

python -m gcpc.train env_name=antmaze-large-play-v2 model=policynet exp=antmaze_pl model.model_config.tjn_ckpt_path=$tjn_ckpt_path

Kitchen

python -m gcpc.train env_name=kitchen-partial-v0 model=policynet exp=kitchen_pl model.model_config.tjn_ckpt_path=$tjn_ckpt_path

Gym

python -m gcpc.train env_name=halfcheetah-medium-expert-v2 model=policynet exp=gym_pl model.model_config.tjn_ckpt_path=$tjn_ckpt_path

To launch policy learning with multiple datasets and seeds, replace the TrajNet checkpoint path in scripts/launch_pl_<env>.sh and run

sh scripts/launch_pl_<env>.sh

Citation

If you find this repository useful for your research, please consider citing our work:

@inproceedings{zeng2023gcpc,
  title={Goal-Conditioned Predictive Coding for Offline Reinforcement Learning},
  author={Zeng, Zilai and Zhang, Ce and Wang, Shijie and Sun, Chen},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023}
}

Acknowledgements

This repo contains code adapted from rvs, trajectory-transformer, MaskDP_Public and decision-diffuser. We thank the authors and contributors for open-sourcing their code.