tout

  • Description :

Franka ramassant et versant des tâches

Diviser Exemples
  • Structure des fonctionnalités :
FeaturesDict({
    'steps': Dataset({
        'action': FeaturesDict({
            'open_gripper': bool,
            'rotation_delta': Tensor(shape=(3,), dtype=float32),
            'terminate_episode': float32,
            'world_vector': Tensor(shape=(3,), dtype=float32),
        }),
        'is_first': bool,
        'is_last': bool,
        'is_terminal': bool,
        'observation': FeaturesDict({
            'image': Image(shape=(480, 640, 3), dtype=uint8),
            'natural_language_embedding': Tensor(shape=(512,), dtype=float32),
            'natural_language_instruction': string,
            'state': Tensor(shape=(7,), dtype=float32),
        }),
        'reward': Scalar(shape=(), dtype=float32),
    }),
})
  • Documentation des fonctionnalités :
Fonctionnalité Classe Forme Type D Description
FonctionnalitésDict
pas Base de données
étapes/actions FonctionnalitésDict
étapes/action/open_gripper Tenseur bouffon
étapes/action/rotation_delta Tenseur (3,) flotteur32
étapes/action/terminate_episode Tenseur flotteur32
étapes/action/world_vector Tenseur (3,) flotteur32
étapes/is_first Tenseur bouffon
étapes/est_dernier Tenseur bouffon
étapes/is_terminal Tenseur bouffon
étapes/observation FonctionnalitésDict
étapes/observation/image Image (480, 640, 3) uint8
étapes/observation/natural_lingual_embedding Tenseur (512,) flotteur32
étapes/observation/instruction_langue_naturelle Tenseur chaîne
étapes/observation/état Tenseur (7,) flotteur32 tableau numpy de forme (7,). Contient les états des articulations du robot (sous forme d'angles d'articulation absolus) à chaque pas de temps
étapes/récompense Scalaire flotteur32
@inproceedings{zhou2023train,
  author={Zhou, Gaoyue and Dean, Victoria and Srirama, Mohan Kumar and Rajeswaran, Aravind and Pari, Jyothish and Hatch, Kyle and Jain, Aryan and Yu, Tianhe and Abbeel, Pieter and Pinto, Lerrel and Finn, Chelsea and Gupta, Abhinav},
  booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
  title={Train Offline, Test Online: A Real Robot Learning Benchmark},
  year={2023},
 }