EfficientGrasp

Data, trained models, meshes and videos of EfficientGrasp: A Unified Data-Efficient Learning to Grasp Method for Multi-fingered Robot Hands (Kelin Li, Nicholas Baron, Xian Zhang and Nicolas Rojas, IEEE Robotics and Automation Letters 2022, presented at IROS 2022).

Code: github.com/KelinLi-ReCAST/EfficientGrasp

Every file sits at the same path as in the GitHub repository, so the dataset is downloaded into the root of a clone of the code.

Download

git clone https://github.com/KelinLi-ReCAST/EfficientGrasp
cd EfficientGrasp
pip install -U huggingface_hub
hf download KelinLiIC/EfficientGrasp --repo-type dataset --local-dir . --exclude README.md --exclude .gitattributes
tar -xf gripper_representation/data/WorkspaceArrays.tar -C gripper_representation/data   # autoencoder training set, 44 751 files

Contents

Path	Contents
gripper_representation/data/WorkspaceArrays.tar	Fingertip workspaces of parametric grippers, the autoencoder training set (WorkspaceArrays/ws<N>.npy, packed because a folder on the Hub holds at most 10 000 files)
gripper_representation/feature_extraction/	RUTH and coupled-rotation workspaces, trained autoencoder (saved_models/), TensorBoard logs
gripper_representation/ruth_workspace_sampling/contact_points.npy	RUTH fingertip workspace sampled in PyBullet
contact_point_selection/UniGrasp/data/	Object point clouds, PSSN contact points, gripper workspace features, UniGrasp labels
contact_point_selection/UniGrasp/saved_models/	PSSN checkpoints (point_set_selection/220model.ckpt is EfficientGrasp) and the released UniGrasp models
contact_point_selection/UniGrasp/point_set_selection/logs/	PSSN training logs
contact_point_selection/UniGrasp/gripper_urdf/, rl_inverse_kinematics/*/gym/, rl_inverse_kinematics/*/meshes/	Meshes referenced by the URDFs
grasp_quality/data/	DeepSDF weights and the MANO pose distribution
rl_inverse_kinematics/<gripper>/log/kelin-v0/	Trained SAC policies loaded by RL_gripper.py
rl_inverse_kinematics/<gripper>/{contact_points,contact_list,ws1}.npy, obj_pc/	RL training targets and object point clouds
rl_inverse_kinematics/early_prototype/log/, revision_lr_ablation/log/	First SAC version and the learning-rate study
rl_inverse_kinematics/her_baseline_attempt/results and plots/	DDPG and DQN weights of the HER baseline attempt
third_party/	pybullet-object-models (YCB objects) in three variants; the meshes of superquadric_objects/, which EfficientGrasp does not use, are not included
results/rl_ik_error/, results/pssn_accuracy/	Evaluation logs
results/videos/efficientgrasp/, results/videos/unigrasp_baseline/	PyBullet recordings of the grasp trials
evaluation/real_world/	RealSense point clouds (object_pointclouds/), contact-point renders (object_photos/), occlusion analysis (occlusion/)

Not included: the MANO models, which are licence-restricted (mano.is.tue.mpg.de), and the UniGrasp training data, available from the UniGrasp authors.

Third-party content keeps its own licence: the UniGrasp models and labels, the BarrettHand and gripper meshes, and the YCB object models in third_party/ (see the LICENSE files in those folders).

Citation

@article{li2022efficientgrasp,
  title   = {EfficientGrasp: A Unified Data-Efficient Learning to Grasp Method for Multi-Fingered Robot Hands},
  author  = {Li, Kelin and Baron, Nicholas and Zhang, Xian and Rojas, Nicolas},
  journal = {IEEE Robotics and Automation Letters},
  volume  = {7},
  number  = {4},
  pages   = {8619--8626},
  year    = {2022},
  doi     = {10.1109/LRA.2022.3187875},
  note    = {arXiv:2206.15159}
}