Mapping State Space using Landmarks for Universal Goal Reaching

We combine low-level RL policy with search algorithm to solve the goal-reaching problem. We build a graph-based environment map from past experience which couples sampling and end-to-end training.

Paper

This paper is accepted by NeurIPS 2019. For more details, please refer to our paper.

Run

We provide the scripts for 2DPlane, PointMaze and AntMaze.

./scripts/train_2dplane.sh
./scripts/train_pointmaze.sh
./scripts/train_antmaze.sh

You can also customize your own environments and find suitable parameters.

The goal reaching environments can be found in ./goal_env.

Cite Our Paper

If you find it useful, please consider to cite our paper.

@inproceedings{huang2019mapping,
  title={Mapping state space using landmarks for universal goal reaching},
  author={Huang, Zhiao and Liu, Fangchen and Su, Hao},
  booktitle={Advances in Neural Information Processing Systems},
  pages={1940--1950},
  year={2019}
}

Acknowledgement

• Openai Baselines
• Pytorch Baseline implemented by TianhongDai

Pretrained Model

• The pretrained model for PointMaze with the default architecture can be founded here. This is trained with 0.4M interaction steps.
• The pretrained model for AntMaze: will upload soon.