[NeurIPS 2023 FMDM Workshop] Skill Reinforcement Learning and Planning for Open-World Long-Horizon Tasks
[NeurIPS 2023 FMDM Workshop] Skill Reinforcement Learning and Planning for Open-World Long-Horizon Tasks
Plan4MC is a multi-task agent in Minecraft, solving long-term tasks via planning over basic skills. It acquire three types of fine-grained basic skills through reinforcement learning without demonstrations. With a skill graph pre-generated by the Large Language Model, the skill search algorithm generates skill plans and interactively selects policies to solve complicated tasks. Plan4MC accomplishes 24 diverse hard tasks in Minecraft.
Install MineDojo environment following the official document. It requires python >= 3.9. We install jdk 1.8.0_171.
Upgrade the MineDojo package:
pip uninstall minedojo.python setup.py install.validate_install.py here.Install python packages in requirements.txt. Note that we require PyTorch>=1.8.1 and x-transformers==0.27.1.
Download the pretrained MineCLIP model named attn.pth. Move the file to mineclip_official/.
envs/hard_task_conf.yaml. To create an environment for a task:…python plan_all_tasks.py.Pre-trained models for skills are released in skills/models/.
To test Plan4MC on the task 'get_furnace_nearby_with_logs' as an example, run python test.py --task get_furnace_nearby_with_logs. Arguments:
--task: available tasks are listed in envs/hard_task_conf.yaml.
--save-path: output directory for testing results.
--test-episode: number of testing episodes.
--save-gif: set to 1 to save testing videos.
@article{yuan2023plan4mc,
title={{Plan4MC}: Skill Reinforcement Learning and Planning for Open-World {Minecraft} Tasks},
author={Yuan, Haoqi and Zhang, Chi and Wang, Hongcheng and Xie, Feiyang and Cai, Penglin and Dong, Hao and Lu, Zongqing},
journal={arXiv preprint arXiv:2303.16563},
year={2023},
}No open issues yet, or sync has not completed.