This is the codebase of our paper ''Planning for Sample Efficient Imitation Learning'' at the NeurIPS 2022.
Step 1: Preparing python packages. This project is dependent on the following packages ray, torch, dmc2gym, opencv-python, and kornia. We can use pip to install them.
Step 2: Compiling cpp source. After installing these dependencies, we need to compile the C++ source of the MCTS as follows.
cd mcts_tree_sample
bash make.sh
Step 3: Download the data. Finally, we need to download the demonstration data at GoogleDrive, and put them into the ./data folder.
We put the launch scripts at the ./scripts folder. For example, you can launch the training of walker by
bash ./scripts/walker_state.sh
Note that the training is carried out on a server with 4 NVIDIA 3090 GPUs, 128 CPU Cores, and 512GB RAM.
If you find this work useful and would like to cite it in your research:
@inproceedings{efficientimitate,
title={Planning for Sample Efficient Imitation Learning},
author={Yin, Zhao-Heng and Ye, Weirui and Chen, Qifeng and Gao, Yang},
booktitle={Neural Information Processing Systems},
year={2022}
}
MIT License