mt_opt

  • विवरण :

एमटी-ऑप्ट पेपर के लिए डेटासेट।

@misc{kalashnikov2021mtopt,
      title={MT-Opt: Continuous Multi-Task Robotic Reinforcement Learning at Scale},
      author={Dmitry Kalashnikov and Jacob Varley and Yevgen Chebotar and Benjamin Swanson and Rico Jonschkowski and Chelsea Finn and Sergey Levine and Karol Hausman},
      year={2021},
      eprint={2104.08212},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}

mt_opt/rlds (डिफ़ॉल्ट कॉन्फ़िगरेशन)

  • विन्यास विवरण : इस डेटासेट में वास्तविक रोबोटों के बेड़े में एकत्र किए गए कार्य एपिसोड शामिल हैं। यह चरणों और प्रकरणों का प्रतिनिधित्व करने के लिए RLDS प्रारूप का अनुसरण करता है।

  • डेटासेट का आकार : 4.38 TiB

  • विभाजन :

विभाजित करना उदाहरण
'train' 920,165
  • फ़ीचर संरचना :
FeaturesDict({
    'episode_id': string,
    'skill': uint8,
    'steps': Dataset({
        'action': FeaturesDict({
            'close_gripper': bool,
            'open_gripper': bool,
            'target_pose': Tensor(shape=(7,), dtype=float32),
            'terminate': bool,
        }),
        'is_first': bool,
        'is_last': bool,
        'is_terminal': bool,
        'observation': FeaturesDict({
            'gripper_closed': bool,
            'height_to_bottom': float32,
            'image': Image(shape=(512, 640, 3), dtype=uint8),
            'state_dense': Tensor(shape=(7,), dtype=float32),
        }),
    }),
    'task_code': string,
})
  • फ़ीचर दस्तावेज़ीकरण :
विशेषता कक्षा आकार डीटाइप विवरण
विशेषताएं डिक्ट
प्रकरण_आईडी टेन्सर डोरी
कौशल टेन्सर uint8
कदम डेटासेट
कदम / कार्रवाई विशेषताएं डिक्ट
कदम/कार्रवाई/close_gripper टेन्सर बूल
कदम/कार्रवाई/open_gripper टेन्सर बूल
कदम/कार्रवाई/target_pose टेन्सर (7,) फ्लोट32
कदम/कार्रवाई/समाप्ति टेन्सर बूल
चरण/is_first टेन्सर बूल
चरण/is_last टेन्सर बूल
कदम/is_terminal टेन्सर बूल
चरण/अवलोकन विशेषताएं डिक्ट
कदम / अवलोकन / ग्रिपर_क्लोज्ड टेन्सर बूल
चरण/अवलोकन/ऊंचाई_to_bottom टेन्सर फ्लोट32
चरण/अवलोकन/छवि छवि (512, 640, 3) uint8
चरण/अवलोकन/state_dense टेन्सर (7,) फ्लोट32
कार्य_कोड टेन्सर डोरी

एमटी_ऑप्ट/एसडी

  • Config विवरण : सफलता डिटेक्टर डेटासेट जिसमें कार्यों को पूरा करने की मानव क्यूरेटेड परिभाषाएं शामिल हैं।

  • डेटासेट का आकार : 548.56 GiB

  • विभाजन :

विभाजित करना उदाहरण
'test' 94,636
'train' 380,234
  • फ़ीचर संरचना :
FeaturesDict({
    'image_0': Image(shape=(512, 640, 3), dtype=uint8),
    'image_1': Image(shape=(480, 640, 3), dtype=uint8),
    'image_2': Image(shape=(480, 640, 3), dtype=uint8),
    'success': bool,
    'task_code': string,
})
  • फ़ीचर दस्तावेज़ीकरण :
विशेषता कक्षा आकार डीटाइप विवरण
विशेषताएं डिक्ट
इमेज_0 छवि (512, 640, 3) uint8
छवि_1 छवि (480, 640, 3) uint8
छवि_2 छवि (480, 640, 3) uint8
सफलता टेन्सर बूल
कार्य_कोड टेन्सर डोरी