The dataset is organised first according to the number and type of hooks included in the workspace, thus the folders named "2_hooks_C", "3_hooks_C+", etc.
For each number and type of hook, the folder contains 5 subfolders of data.
Each subfolder contains the point cloud scene latent features, the mojuco simulation states and the grasp poses generated by GraspNet1Billion.
Each .npy file is indexed and contains 10k datapoints.
The datapoints are ordered within each .npy file, which means that the i-th datapoint in j-th good_grasp_features_j.npy file should correspond to the i-th datapoint in the j-th good_grasp_mjc_state_j.npy file.
The .npy files can be read via numpy, while the .mjc files can be read by pickle.
An example Python script is provided for reading the dataset. Please refer to https://github.com/IanYangChina/GAM-paper-codes for the official codes of the paper for more information of using the dataset.