Awesome

🧾 Introduction

H-InDex is a visual reinforcement learning framework that leverages hand-informed representations to learn dexterous manipulation skills with great efficiency. H-InDex consistes of three stages: pre-training, offline adaptation, and reinforcement learning. In this repo, all the stages are provided, together with the pre-trained checkpoint and the adapted checkpoints.

We also encourage the user to use our pre-trained representations directly for their own downstream tasks.

To benchmark our method, we also provide several strong baselines in this repo, including VC-1, MVP, R3M, and RRL.

Enjoy Dexterity!

💻 Installation

See INSTALL.md.

We also provide some error catching solutions in INSTALL.md.

Feel free to post an issue if you have any questions.

🛠️ Usage

We use wandb to log the training process. Remember to set your wandb account before training by wandb login. You could also disable wandb by use_wandb=0 in our script.

Given a task name task_name, you could run the following pipeline.

• Stage 1: Human Hand Pretraining.
  • Simply download the pre-trained hand representation from FrankMocap by this command

    wget https://dl.fbaipublicfiles.com/eft/fairmocap_data/hand_module/checkpoints_best/pose_shape_best.pth -O archive/frankmocap_hand.pth --no-check-certificate`
    

• Stage 2: Offline Adaptation.
  • First, download the initial model weights in Stage 2 from here and put it under stage2_adapt/.
  • Second, generate image dataset for offline adaptation. See scripts/adroit/gen_img_dataset.sh or scripts/dexmv/gen_img_dataset.sh for details. An example:

    bash scripts/adroit/gen_img_dataset.sh hammer
    

  • Third, adapt affine transformation in pretrained model. See scripts/train_stage2.sh for details. An example:

    bash scripts/train_stage2.sh hammer-v0
    

  • For the users' convenience, we also provide the adapted checkpoints for all the tasks. You can download them from here and put them under archive/ folder.
• Stage 3: Reinforcement Learning.
  • Train RL agents with the pre-trained representations. See scripts/adroit/train.sh or scripts/dexmv/train.sh for details. An example:

    bash scripts/adroit/train.sh hammer hindex test 0 0
    

    Arguments are task name, representation name, experiment name, seed, and GPU id respectively.

🦉 Tasks

We provide 12 dexterous manipulation Tasks in total:

• Adroit (3): pen, door, hammer
• DexMV (9): pour, place_inside, relocate-mug, relocate-foam_brick, relocate-large_clamp, relocate-mustard_bottle, relocate-potted_meat_can, relocate-sugar_box, relocate-tomato_soup_can

🙏 Acknowledgement

Our work is based on many open-source projects. The algorithms are mainly built upon RRL and TTP. The simulation environments are from DAPG and DexMV. The pre-trained hand representation is from FrankMocap. Baselines are from RRL, MVP, R3M and VC-1. We thank all these authors for their nicely open sourced code and their great contributions to the community.

🏷️ License

H-InDex is licensed under the MIT license. See the LICENSE file for details.

📝 Citation

If you find our work useful, please consider citing:

@article{Ze2023HInDex,
  title={H-InDex: Visual Reinforcement Learning with Hand-Informed Representations for Dexterous Manipulation},
  author={Yanjie Ze and Yuyao Liu and Ruizhe Shi and Jiaxin Qin and Zhecheng Yuan and Jiashun Wang and Xiaolong Wang and Huazhe Xu},
  journal={NeurIPS}, 
  year={2023},
}