UniDexGrasp++: Improving Dexterous Grasping Policy Learning via Geometry-aware Curriculum and Iterative Generalist-Specialist Learning

This is the official repository of UniDexGrasp++: Improving Dexterous Grasping Policy Learning via Geometry-aware Curriculum and Iterative Generalist-Specialist Learning.

For more information, please visit our project page.

Update

[2023.10.29] Release code and state-based checkpoints and results summary.

Overview

In this work, we present a novel dexterous grasping policy learning pipeline, UniDexGrasp++. Same to UniDexGrasp, UniDexGrasp++ is trained on 3000+ different object instances with random object poses under a table-top setting. It significantly outperforms the previous SOTA and achieves 85.4% and 78.2% success rates on the train and test set.

Pipeline

We propose a novel, object-agnostic method for learning a universal policy for dexterous object grasping from realistic point cloud observations and proprioceptive information under a table-top setting, namely UniDexGrasp++. To address the challenge of learning the vision-based policy across thousands of object instances, we propose Geometry-aware Curriculum Learning (GeoCurriculum) and Geometry-aware iterative Generalist-Specialist Learning (GiGSL) which leverage the geometry feature of the task and significantly improve the generalizability. With our proposed techniques, our final policy shows universal dexterous grasping on thousands of object instances with 85.4% and 78.2% success rate on the train set and test set which outperforms the state-of-the-art baseline UniDexGrasp by 11.7% and 11.3%, respectively.

Installation

Details regarding installation of IsaacGym can be found here. We test with Preview Release 3/4 and Preview Release 4/4 version of IsaacGym and use the Preview Release 3/4 in our paper experiment.

Please follow the steps below to perform the installation：

1. Create virtual environment

conda create -n dexgrasp python==3.8
conda activate dexgrasp

2. Install isaacgym

Once you have downloaded IsaacGym:

cd <PATH_TO_ISAACGYM_INSTALL_DIR>/python
pip install -e .

Ensure that Isaac Gym works on your system by running one of the examples from the python/examples directory, like joint_monkey.py. Please follow troubleshooting steps described in the Isaac Gym Preview Release 3/4 install instructions if you have any trouble running the samples.

3. Install dexgrasp

Once Isaac Gym is installed and samples work within your current python environment, install this repo from source code:

cd <PATH_TO_DEXGRASP_POLICY_DIR>
pip install -e .

cd DexGrasp-test
pip install -e .

4. Install pointnet2_ops

pip install "git+https://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"

Dataset

We use the UniDexGrasp dataset. Addtionaly please download datasetv4.1_posedata.npy under assets. Please unpack them put them under the directory dexgrasp_policy/assets. The full objects and their corresponding scales informations we used are in dexgrasp/cfg/train_set.yaml, dexgrasp/cfg/test_set_seen_cat.yaml, dexgrasp/cfg/test_set_unseen_cat.yaml.

Training/Evaluation

We provide two tasks: for the state-based policy task, please see dexgrasp/tasks/shadow_hand_grasp.py; Please modify object_code_dict in cfg/shadow_hand_grasp.yaml in order to change the training objects; for the vision-based policy tasks, in order to train on more objects within a certain GPU memory limit, we randomly load objects from the dataset in the beginning of each episode during training. please see dexgrasp/tasks/shadow_hand_random_load_vision.py.

Run the following lines in dexgrasp folder.

training state-based policy training using ppo:

bash script/run_train_ppo_state.sh 

training state-based policy distillation using DAgger (Please modify expert in dexgrasp/cfg/dagger_value in order to assign training objects with different teacher policy):

bash script/run_train_dagger_state.sh 

training state to vision policy distillation using DAgger:

bash script/run_train_dagger_state_to_vision.sh

training vision-based policy training using ppo:

bash script/run_train_ppo_vision.sh

Add --test in the training scripts for evaluation. For more provided args (e.g., backbone type, test mode), please check these scripts and utils/config.py.

Results and Checkpoints

We provide our trained state-based policy checkpoint at dexgrasp/state_based_model and its detailed evaluation results on all training and test objects at results/state_based/train_set_results.yaml, results/state_based/test_set_seen_cat_results.yaml, results/state_based/test_set_unseen_cat_results.yaml.

Acknowledgement

The code base used in this project is sourced from these repository:

NVIDIA-Omniverse/IsaacGymEnvs

PKU-MARL/DexterousHands

PKU-EPIC/UniDexGrasp

Citation

If you find our papers helpful, please consider cite:

@article{xu2023unidexgrasp,
  title={UniDexGrasp: Universal Robotic Dexterous Grasping via Learning Diverse Proposal Generation and Goal-Conditioned Policy},
  author={Xu, Yinzhen and Wan, Weikang and Zhang, Jialiang and Liu, Haoran and Shan, Zikang and Shen, Hao and Wang, Ruicheng and Geng, Haoran and Weng, Yijia and Chen, Jiayi and others},
  journal={arXiv preprint arXiv:2303.00938},
  year={2023}
}
@article{wan2023unidexgrasp++,
  title={UniDexGrasp++: Improving Dexterous Grasping Policy Learning via Geometry-aware Curriculum and Iterative Generalist-Specialist Learning},
  author={Wan, Weikang and Geng, Haoran and Liu, Yun and Shan, Zikang and Yang, Yaodong and Yi, Li and Wang, He},
  journal={arXiv preprint arXiv:2304.00464},
  year={2023}
}

License

This work and the dataset are licensed under CC BY-NC 4.0.