On Pre-Training for Visuo-Motor Control: Revisiting a Learning-from-Scratch Baseline

This is the repo of the paper: On Pre-Training for Visuo-Motor Control: Revisiting a Learning-from-Scratch Baseline

Setup

Our code is based on the isaacgym and the repo of Masked Visual Pre-training for Motor Control.

• Create a conda environment:

conda create -n mvp python=3.7
conda activate mvp

• Install Pytorch:

  • pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116

• Install the Preview 3 isaac gym version:

  • download issacgym from https://developer.nvidia.com/isaac-gym/download, and follow the instructions to install it.
  • For anaconda users, there are some installing tips in isaacgym/docs/install.html for handling potential issues (e.g., libpython3.7).

• Install MVP:

  • pip install -r requirements.txt
  • pip install -e .

LfS

If you want reproduce the results of LfS, just run the following commands:

bash script/train_w_DA.sh

The setting of LfS(w/o aug) is: is_image=True is_DA=False is_drac=False

The setting of LfS(+ aug) is: is_image=True is_DA=True is_drac=True

Pre-trained baseline

The pretrained baselines contains MVP, PVR (Moco) and R3M.

For R3M, we can just follow the setup from the official link: https://github.com/facebookresearch/r3m

For PVR, we use the moco_aug.pth.tar from the official link: https://github.com/sparisi/pvr_habitat/releases/download/models/moco_aug.pth.tar

For MVP, we choose vits-mae-hoi as the pretraiend model, and it can be downloaded from:

https://github.com/ir413/mvp

Besdies, you should change the line of 135 in actor_critic.py to your own model path.

Off-policy

The off-policy part is mainly based on the code of DrQ-v2. Our modifications have been indicated in the appendix, and implementation can be done very easily.

Citation

If you find our work useful in your research, please consider citing our work as follows:

@article{hansen2022pre,
  title={On Pre-Training for Visuo-Motor Control: Revisiting a Learning-from-Scratch Baseline},
  author={Hansen, Nicklas and Yuan, Zhecheng and Ze, Yanjie and Mu, Tongzhou and Rajeswaran, Aravind and Su, Hao and Xu, Huazhe and Wang, Xiaolong},
  journal={arXiv preprint arXiv:2212.05749},
  year={2022}
}