Model-Based Reinforcement Learning with Multi-Task Offline Pretraining (ECML 2024)

Model-Based Reinforcement Learning with Multi-Task Offline Pretraining [arXiv]

Minting Pan*, Yitao Zheng*, Yunbo Wang, Xiaokang Yang

Setting up

Create an environment

conda env create -f env.yaml

Experiments

Fine-tuning command on Meta-World:

python dreamer.py \
--logdir \
path/to/log \
--config \
defaults metaworld \
--task \
target_metaworld_task \
--video_dir \
path/to/offline/datasets \
--pretrain_model_dir \
path/to/teacher/model \
--source_tasks \
['list', 'of', 'source', 'metaworld', 'tasks']

Fine-tuning command on Deepmind Control Suite:

python dreamer.py \
--logdir \
path/to/log \
--config \
defaults dmc \
--task \
target_dmc_task \
--video_dir \
path/to/offline/datasets \
--pretrain_model_dir \
path/to/teacher/model \
--source_tasks \
['list', 'of', 'source', 'dmc', 'tasks']

Acknowledgement

We appreciate the following github repos where we borrow code from:

https://github.com/jsikyoon/dreamer-torch