spoc

Split Examples
  • Feature structure:
FeaturesDict({
    'episode_metadata': FeaturesDict({
        'file_path': string,
        'task_target_split': string,
        'task_type': string,
    }),
    'steps': Dataset({
        'action': Tensor(shape=(9,), dtype=float32),
        'discount': Scalar(shape=(), dtype=float32),
        'is_first': bool,
        'is_last': bool,
        'is_terminal': bool,
        'language_instruction': string,
        'observation': FeaturesDict({
            'an_object_is_in_hand': Scalar(shape=(), dtype=bool),
            'house_index': Scalar(shape=(), dtype=int64),
            'hypothetical_task_success': Scalar(shape=(), dtype=bool),
            'image': Image(shape=(224, 384, 3), dtype=uint8),
            'image_manipulation': Image(shape=(224, 384, 3), dtype=uint8),
            'last_action_is_random': Scalar(shape=(), dtype=bool),
            'last_action_str': string,
            'last_action_success': Scalar(shape=(), dtype=bool),
            'last_agent_location': Tensor(shape=(6,), dtype=float32),
            'manip_object_bbox': Tensor(shape=(10,), dtype=float32),
            'minimum_l2_target_distance': Scalar(shape=(), dtype=float32),
            'minimum_visible_target_alignment': Scalar(shape=(), dtype=float32),
            'nav_object_bbox': Tensor(shape=(10,), dtype=float32),
            'relative_arm_location_metadata': Tensor(shape=(4,), dtype=float32),
            'room_current_seen': Scalar(shape=(), dtype=bool),
            'rooms_seen': Scalar(shape=(), dtype=int64),
            'visible_target_4m_count': Scalar(shape=(), dtype=int64),
        }),
        'reward': Scalar(shape=(), dtype=float32),
    }),
})
  • Feature documentation:
Feature Class Shape Dtype Description
FeaturesDict
episode_metadata FeaturesDict
episode_metadata/file_path Tensor string
episode_metadata/task_target_split Tensor string
episode_metadata/task_type Tensor string
steps Dataset
steps/action Tensor (9,) float32
steps/discount Scalar float32
steps/is_first Tensor bool
steps/is_last Tensor bool
steps/is_terminal Tensor bool
steps/language_instruction Tensor string
steps/observation FeaturesDict
steps/observation/an_object_is_in_hand Scalar bool
steps/observation/house_index Scalar int64
steps/observation/hypothetical_task_success Scalar bool
steps/observation/image Image (224, 384, 3) uint8
steps/observation/image_manipulation Image (224, 384, 3) uint8
steps/observation/last_action_is_random Scalar bool
steps/observation/last_action_str Tensor string
steps/observation/last_action_success Scalar bool
steps/observation/last_agent_location Tensor (6,) float32
steps/observation/manip_object_bbox Tensor (10,) float32
steps/observation/minimum_l2_target_distance Scalar float32
steps/observation/minimum_visible_target_alignment Scalar float32
steps/observation/nav_object_bbox Tensor (10,) float32
steps/observation/relative_arm_location_metadata Tensor (4,) float32
steps/observation/room_current_seen Scalar bool
steps/observation/rooms_seen Scalar int64
steps/observation/visible_target_4m_count Scalar int64
steps/reward Scalar float32
@article{spoc2023,
    author    = {Kiana Ehsani, Tanmay Gupta, Rose Hendrix, Jordi Salvador, Luca Weihs, Kuo-Hao Zeng, Kunal Pratap Singh, Yejin Kim, Winson Han, Alvaro Herrasti, Ranjay Krishna, Dustin Schwenk, Eli VanderBilt, Aniruddha Kembhavi},
    title     = {Imitating Shortest Paths in Simulation Enables Effective Navigation and Manipulation in the Real World},
    journal   = {arXiv},
    year      = {2023},
    eprint    = {2312.02976},
}