mt_opt

  • תיאור :

מערכי נתונים עבור נייר MT-Opt .

@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 (תצורת ברירת המחדל)

  • תיאור תצורה : מערך נתונים זה מכיל פרקי משימות שנאספו על פני שפע רובוטים אמיתיים. זה עוקב אחר פורמט RLDS כדי לייצג שלבים ופרקים.

  • גודל מערך נתונים : 4.38 TiB

  • פיצולים :

לְפַצֵל דוגמאות
'train' 920,165
  • מבנה תכונה :
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,
})
  • תיעוד תכונה :
תכונה מעמד צוּרָה Dtype תיאור
FeaturesDict
פרק_מזהה מוֹתֵחַ חוּט
מְיוּמָנוּת מוֹתֵחַ uint8
צעדים מערך נתונים
צעדים/פעולה FeaturesDict
steps/action/close_gripper מוֹתֵחַ bool
steps/action/open_gripper מוֹתֵחַ bool
steps/action/target_pose מוֹתֵחַ (7,) לצוף32
שלבים/פעולה/סיום מוֹתֵחַ bool
צעדים/הוא_ראשון מוֹתֵחַ bool
צעדים/הוא_אחרון מוֹתֵחַ bool
steps/is_terminal מוֹתֵחַ bool
צעדים/תצפית FeaturesDict
מדרגות/תצפית/גריפר_סגור מוֹתֵחַ bool
צעדים/תצפית/גובה_למטה מוֹתֵחַ לצוף32
צעדים/תצפית/תמונה תמונה (512, 640, 3) uint8
צעדים/תצפית/מצב_צפוף מוֹתֵחַ (7,) לצוף32
קוד_משימה מוֹתֵחַ חוּט

mt_opt/sd

  • תיאור תצורה : מערך הנתונים של גלאי ההצלחה שמכיל הגדרות שנאספו על ידי אנוש של השלמת משימות.

  • גודל מערך נתונים : 548.56 GiB

  • פיצולים :

לְפַצֵל דוגמאות
'test' 94,636
'train' 380,234
  • מבנה תכונה :
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,
})
  • תיעוד תכונה :
תכונה מעמד צוּרָה Dtype תיאור
FeaturesDict
תמונה_0 תמונה (512, 640, 3) uint8
תמונה_1 תמונה (480, 640, 3) uint8
תמונה_2 תמונה (480, 640, 3) uint8
הַצלָחָה מוֹתֵחַ bool
קוד_משימה מוֹתֵחַ חוּט