• Description:

Franka cloth interaction tasks

Split Examples
'train' 800
'val' 200
  • Feature structure:
    'episode_metadata': FeaturesDict({
        'file_path': Text(shape=(), dtype=string),
    'steps': Dataset({
        'action': Tensor(shape=(4,), dtype=float32),
        'discount': Scalar(shape=(), dtype=float32),
        'is_first': bool,
        'is_last': bool,
        'is_terminal': bool,
        'language_embedding': Tensor(shape=(512,), dtype=float32),
        'language_instruction': Text(shape=(), dtype=string),
        'observation': FeaturesDict({
            'image': Image(shape=(32, 32, 3), dtype=uint8),
        'reward': Scalar(shape=(), dtype=float32),
  • Feature documentation:
Feature Class Shape Dtype Description
episode_metadata FeaturesDict
episode_metadata/file_path Text string Path to the original data file.
steps Dataset
steps/action Tensor (4,) float32 Robot action, consists of x,y,z goal and picker commandpicker<0.5 = open, picker>0.5 = close.
steps/discount Scalar float32 Discount if provided, default to 1.
steps/is_first Tensor bool
steps/is_last Tensor bool
steps/is_terminal Tensor bool
steps/language_embedding Tensor (512,) float32 Kona language embedding. See
steps/language_instruction Text string Language Instruction.
steps/observation FeaturesDict
steps/observation/image Image (32, 32, 3) uint8 Image observation of cloth.
steps/reward Scalar float32 Reward as a normalized performance metric in [0, 1].0 = no change from initial state. 1 = perfect fold.-ve performance means the cloth is worse off than initial state.
    author={Salhotra, Gautam and Liu, I-Chun Arthur and Dominguez-Kuhne, Marcus and Sukhatme, Gaurav S.},
    journal={IEEE Robotics and Automation Letters},
    title={Learning Deformable Object Manipulation From Expert Demonstrations},