datasets.robotics package

Submodules

datasets.robotics.MultimodalManipulationDataset module

Implements dataset for MultiModal Manipulation Task.

class datasets.robotics.MultimodalManipulationDataset.MultimodalManipulationDataset(filename_list, transform=None, episode_length=50, training_type='selfsupervised', n_time_steps=1, action_dim=4, pairing_tolerance=0.06, filedirprefix='')

Bases: Dataset

Multimodal Manipulation dataset.

__init__(filename_list, transform=None, episode_length=50, training_type='selfsupervised', n_time_steps=1, action_dim=4, pairing_tolerance=0.06, filedirprefix='')

Initialize dataset.

Parameters:

  • filename_list (str) – List of files to get data from

  • transform (fn, optional) – Optional function to transform data. Defaults to None.

  • episode_length (int, optional) – Length of each episode. Defaults to 50.

  • training_type (str, optional) – Type of training. Defaults to “selfsupervised”.

  • n_time_steps (int, optional) – Number of time steps. Defaults to 1.

  • action_dim (int, optional) – Action dimension. Defaults to 4.

  • pairing_tolerance (float, optional) – Pairing tolerance. Defaults to 0.06.

  • filedirprefix (str, optional) – File directory prefix (unused). Defaults to “”.

class datasets.robotics.MultimodalManipulationDataset.MultimodalManipulationDataset_robust(filename_list, transform=None, episode_length=50, training_type='selfsupervised', n_time_steps=1, action_dim=4, pairing_tolerance=0.06, filedirprefix='', noise_level=0, image_noise=False, force_noise=False, prop_noise=False)

Bases: Dataset

Multimodal Manipulation dataset.

__init__(filename_list, transform=None, episode_length=50, training_type='selfsupervised', n_time_steps=1, action_dim=4, pairing_tolerance=0.06, filedirprefix='', noise_level=0, image_noise=False, force_noise=False, prop_noise=False)
Parameters:

  • hdf5_file (handle) – h5py handle of the hdf5 file with annotations.

  • transform (callable, optional) – Optional transform to be applied on a sample.

datasets.robotics.MultimodalManipulationDataset_robust module

Implements MultimodalManipulationDataset with robustness transforms.

class datasets.robotics.MultimodalManipulationDataset_robust.MultimodalManipulationDataset_robust(filename_list, transform=None, episode_length=50, training_type='selfsupervised', n_time_steps=1, action_dim=4, pairing_tolerance=0.06, filedirprefix='', noise_level=0, image_noise=False, force_noise=False, prop_noise=False)

Bases: Dataset

Multimodal Manipulation dataset.

__init__(filename_list, transform=None, episode_length=50, training_type='selfsupervised', n_time_steps=1, action_dim=4, pairing_tolerance=0.06, filedirprefix='', noise_level=0, image_noise=False, force_noise=False, prop_noise=False)
Parameters:

  • hdf5_file (handle) – h5py handle of the hdf5 file with annotations.

  • transform (callable, optional) – Optional transform to be applied on a sample.

datasets.robotics.ProcessForce module

Implements processforce, which truncates force readings to a window size.

class datasets.robotics.ProcessForce.ProcessForce(window_size, key='force', tanh=False)

Bases: object

Truncate a time series of force readings with a window size. :param window_size: Length of the history window that is

used to truncate the force readings

__init__(window_size, key='force', tanh=False)

Initialize ProcessForce object.

Parameters:

  • window_size (int) – Windows size

  • key (str, optional) – Key where data is stored. Defaults to ‘force’.

  • tanh (bool, optional) – Whether to apply tanh to output or not. Defaults to False.

datasets.robotics.ToTensor module

Implements another utility for this dataset.

class datasets.robotics.ToTensor.ToTensor(device=None)

Bases: object

Convert ndarrays in sample to Tensors.

__init__(device=None)

Initialize ToTensor object.

datasets.robotics.get_data module

Implements dataloaders for robotics data.

datasets.robotics.get_data.combine_modalitiesbuilder(unimodal, output)

Create a function data combines modalities given the type of input.

Parameters:

  • unimodal (str) – Input type as a string. Can be ‘force’, ‘proprio’, ‘image’. Defaults to using all modalities otherwise

  • output (int) – Index of output modality.

datasets.robotics.get_data.get_data(device, configs, filedirprefix='', unimodal=None, output='contact_next')

Get dataloaders for robotics dataset.

Parameters:

  • device (torch.utils.device) – Device to load data to.

  • configs (dict) – Configuration dictionary

  • filedirprefix (str, optional) – File directory prefix path. Defaults to “”.

  • unimodal (str, optional) – Input modality as a string. Defaults to None. Can be ‘force’, ‘proprio’, ‘image’. Defaults to using all modalities otherwise.

  • output (str, optional) – Output format. Defaults to ‘contact_next’.

Returns:

_description_

Return type:

_type_

datasets.robotics.utils module

Extraneous methods for robotics dataset work.

datasets.robotics.utils.augment_val(val_filename_list, filename_list)

Augment lists of filenames so that they match the current directory.

Module contents

Package that implements dataloaders for robotics data on Multibench.