This is a PyTorch implementation for our paper "Generating Adjacency-Constrained Subgoals in Hierarchical Reinforcement Learning" (NeurIPS 2020 spotlight).

• Python 3.6
• PyTorch 1.3
• OpenAI Gym
• MuJoCo

Also, to run the MuJoCo experiments, a license is required (see here).

Update: implementation for discrete control tasks is in the discrete/ folder; please refer to the usage therein.

• Ant Gather

python main.py --env_name AntGather

• Ant Maze

python main.py --env_name AntMaze

• Ant Maze Sparse

python main.py --env_name AntMazeSparse

• Ant Gather

python eval.py --env_name AntGather --model_dir [MODEL_DIR]

• Ant Maze

python eval.py --env_name AntMaze --model_dir [MODEL_DIR]

• Ant Maze Sparse

python eval.py --env_name AntMazeSparse --model_dir [MODEL_DIR]

Default model_dir is pretrained_models/.

See pretrained_models/ for pre-trained models on all tasks. The expected performances of the pre-trained models are as follows (averaged over 100 evaluation episodes):

If you find this work useful in your research, please cite:

@inproceedings{zhang2020generating,
  title={Generating adjacency-constrained subgoals in hierarchical reinforcement learning},
  author={Zhang, Tianren and Guo, Shangqi and Tan, Tian and Hu, Xiaolin and Chen, Feng},
  booktitle={NeurIPS},
  year={2020}
}