{
  "count": 50,
  "by_task": {
    "arrange_largest_number": {
      "policy": {
        "successes": 0,
        "mean_score": 2.0
      },
      "hybrid": {
        "successes": 1,
        "mean_score": 37.0
      },
      "pi05_native": {
        "successes": 0,
        "mean_score": 3.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 0,
        "qwen_only": 0
      }
    },
    "build_tower": {
      "policy": {
        "successes": 4,
        "mean_score": 82.0
      },
      "hybrid": {
        "successes": 4,
        "mean_score": 86.0
      },
      "pi05_native": {
        "successes": 3,
        "mean_score": 68.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 1,
        "qwen_only": 2
      }
    },
    "classify_objects": {
      "policy": {
        "successes": 0,
        "mean_score": 6.0
      },
      "hybrid": {
        "successes": 0,
        "mean_score": 22.0
      },
      "pi05_native": {
        "successes": 1,
        "mean_score": 26.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 1,
        "qwen_only": 0
      }
    },
    "classify_objects_by_language": {
      "policy": {
        "successes": 0,
        "mean_score": 0.0
      },
      "hybrid": {
        "successes": 0,
        "mean_score": 0.0
      },
      "pi05_native": {
        "successes": 0,
        "mean_score": 0.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 0,
        "qwen_only": 0
      }
    },
    "fold_clothes": {
      "policy": {
        "successes": 1,
        "mean_score": 32.0
      },
      "hybrid": {
        "successes": 4,
        "mean_score": 80.0
      },
      "pi05_native": {
        "successes": 0,
        "mean_score": 16.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 0,
        "qwen_only": 1
      }
    },
    "imitate_sorting_sequence": {
      "policy": {
        "successes": 0,
        "mean_score": 1.0
      },
      "hybrid": {
        "successes": 1,
        "mean_score": 36.0
      },
      "pi05_native": {
        "successes": 0,
        "mean_score": 0.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 0,
        "qwen_only": 0
      }
    },
    "make_kong": {
      "policy": {
        "successes": 1,
        "mean_score": 20.0
      },
      "hybrid": {
        "successes": 3,
        "mean_score": 60.0
      },
      "pi05_native": {
        "successes": 3,
        "mean_score": 60.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 2,
        "qwen_only": 0
      }
    },
    "organize_table": {
      "policy": {
        "successes": 0,
        "mean_score": 10.0
      },
      "hybrid": {
        "successes": 0,
        "mean_score": 65.0
      },
      "pi05_native": {
        "successes": 0,
        "mean_score": 35.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 0,
        "qwen_only": 0
      }
    },
    "pack_objects_into_box": {
      "policy": {
        "successes": 0,
        "mean_score": 20.0
      },
      "hybrid": {
        "successes": 0,
        "mean_score": 2.0
      },
      "pi05_native": {
        "successes": 0,
        "mean_score": 34.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 0,
        "qwen_only": 0
      }
    },
    "put_bottles_into_dustbin": {
      "policy": {
        "successes": 3,
        "mean_score": 64.0
      },
      "hybrid": {
        "successes": 4,
        "mean_score": 82.0
      },
      "pi05_native": {
        "successes": 3,
        "mean_score": 76.0
      },
      "pi05_vs_qwen_success_discordance": {
        "pi05_only": 2,
        "qwen_only": 2
      }
    }
  },
  "overall": {
    "policy": {
      "successes": 9,
      "mean_score": 23.7
    },
    "hybrid": {
      "successes": 17,
      "mean_score": 47.0
    },
    "pi05_native": {
      "successes": 10,
      "mean_score": 31.8
    },
    "pi05_vs_qwen_success_discordance": {
      "pi05_only": 6,
      "qwen_only": 5
    }
  }
}
