[
  {
    "seed": 0,
    "method": "bc",
    "expert_queries": 0,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 1.0,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.0,
      "mean_steps": 10.5,
      "failure_taxonomy": {
        "success": 48,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.0,
      "collision_cost_per_episode": 29.0,
      "intervention_rate": 0.8529411764705882,
      "expert_disagreement_rate": 0.9117647058823529,
      "mean_steps": 34.0,
      "failure_taxonomy": {
        "success": 0,
        "timeout": 0,
        "collision_loop": 48,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.25,
      "collision_cost_per_episode": 11.854166666666666,
      "intervention_rate": 0.5011990407673861,
      "expert_disagreement_rate": 0.6179056754596323,
      "mean_steps": 26.0625,
      "failure_taxonomy": {
        "success": 12,
        "timeout": 0,
        "collision_loop": 30,
        "wrong_target": 6
      }
    },
    "combined_ood": {
      "success_rate": 0.0,
      "collision_cost_per_episode": 19.979166666666668,
      "intervention_rate": 0.6691176470588235,
      "expert_disagreement_rate": 0.9473039215686274,
      "mean_steps": 34.0,
      "failure_taxonomy": {
        "success": 0,
        "timeout": 0,
        "collision_loop": 48,
        "wrong_target": 0
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.2708333333333333,
      "collision_cost_per_episode": 4.729166666666667,
      "intervention_rate": 0.42810457516339867,
      "expert_disagreement_rate": 0.8640522875816994,
      "mean_steps": 31.875,
      "failure_taxonomy": {
        "success": 13,
        "timeout": 18,
        "collision_loop": 13,
        "wrong_target": 4
      }
    }
  },
  {
    "seed": 0,
    "method": "dagger",
    "expert_queries": 867,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 1.0,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.0,
      "mean_steps": 10.5,
      "failure_taxonomy": {
        "success": 48,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "dynamics_ood": {
      "success_rate": 0.3541666666666667,
      "collision_cost_per_episode": 3.0416666666666665,
      "intervention_rate": 0.047619047619047616,
      "expert_disagreement_rate": 0.24816849816849818,
      "mean_steps": 22.75,
      "failure_taxonomy": {
        "success": 17,
        "timeout": 9,
        "collision_loop": 9,
        "wrong_target": 13
      }
    },
    "combined_ood": {
      "success_rate": 0.3958333333333333,
      "collision_cost_per_episode": 5.1875,
      "intervention_rate": 0.2613981762917933,
      "expert_disagreement_rate": 0.48632218844984804,
      "mean_steps": 20.5625,
      "failure_taxonomy": {
        "success": 19,
        "timeout": 2,
        "collision_loop": 13,
        "wrong_target": 14
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.6041666666666666,
      "collision_cost_per_episode": 1.7916666666666667,
      "intervention_rate": 0.14960629921259844,
      "expert_disagreement_rate": 0.4625984251968504,
      "mean_steps": 21.166666666666668,
      "failure_taxonomy": {
        "success": 29,
        "timeout": 12,
        "collision_loop": 1,
        "wrong_target": 6
      }
    }
  },
  {
    "seed": 0,
    "method": "bc_dr",
    "expert_queries": 0,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "dynamics_ood": {
      "success_rate": 0.2916666666666667,
      "collision_cost_per_episode": 8.708333333333334,
      "intervention_rate": 0.4421364985163205,
      "expert_disagreement_rate": 0.552917903066271,
      "mean_steps": 21.0625,
      "failure_taxonomy": {
        "success": 14,
        "timeout": 0,
        "collision_loop": 18,
        "wrong_target": 16
      }
    },
    "combined_ood": {
      "success_rate": 0.375,
      "collision_cost_per_episode": 6.291666666666667,
      "intervention_rate": 0.35978260869565215,
      "expert_disagreement_rate": 0.4967391304347826,
      "mean_steps": 19.166666666666668,
      "failure_taxonomy": {
        "success": 18,
        "timeout": 0,
        "collision_loop": 17,
        "wrong_target": 13
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.6875,
      "collision_cost_per_episode": 1.375,
      "intervention_rate": 0.16,
      "expert_disagreement_rate": 0.416969696969697,
      "mean_steps": 17.1875,
      "failure_taxonomy": {
        "success": 33,
        "timeout": 1,
        "collision_loop": 2,
        "wrong_target": 12
      }
    }
  },
  {
    "seed": 0,
    "method": "dagger_dr_nochunk",
    "expert_queries": 715,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 11.0,
      "intervention_rate": 0.5365853658536586,
      "expert_disagreement_rate": 0.6341463414634146,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 15.0,
      "intervention_rate": 0.7317073170731707,
      "expert_disagreement_rate": 0.7804878048780488,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.25,
      "collision_cost_per_episode": 14.104166666666666,
      "intervention_rate": 0.5600292825768668,
      "expert_disagreement_rate": 0.6961932650073206,
      "mean_steps": 28.458333333333332,
      "failure_taxonomy": {
        "success": 12,
        "timeout": 0,
        "collision_loop": 36,
        "wrong_target": 0
      }
    },
    "combined_ood": {
      "success_rate": 0.4375,
      "collision_cost_per_episode": 3.3958333333333335,
      "intervention_rate": 0.1198501872659176,
      "expert_disagreement_rate": 0.3850187265917603,
      "mean_steps": 27.8125,
      "failure_taxonomy": {
        "success": 21,
        "timeout": 15,
        "collision_loop": 12,
        "wrong_target": 0
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.3541666666666667,
      "collision_cost_per_episode": 2.0,
      "intervention_rate": 0.12005649717514125,
      "expert_disagreement_rate": 0.4477401129943503,
      "mean_steps": 29.5,
      "failure_taxonomy": {
        "success": 17,
        "timeout": 29,
        "collision_loop": 1,
        "wrong_target": 1
      }
    }
  },
  {
    "seed": 0,
    "method": "dagger_dr_chunked",
    "expert_queries": 715,
    "chunk_size": 2,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 14.5,
      "intervention_rate": 0.7073170731707317,
      "expert_disagreement_rate": 0.7560975609756098,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 15.0,
      "intervention_rate": 0.7317073170731707,
      "expert_disagreement_rate": 0.7804878048780488,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.25,
      "collision_cost_per_episode": 14.4375,
      "intervention_rate": 0.5761346998535871,
      "expert_disagreement_rate": 0.7137628111273792,
      "mean_steps": 28.458333333333332,
      "failure_taxonomy": {
        "success": 12,
        "timeout": 0,
        "collision_loop": 36,
        "wrong_target": 0
      }
    },
    "combined_ood": {
      "success_rate": 0.4375,
      "collision_cost_per_episode": 7.083333333333333,
      "intervention_rate": 0.2895148669796557,
      "expert_disagreement_rate": 0.5696400625978091,
      "mean_steps": 26.625,
      "failure_taxonomy": {
        "success": 21,
        "timeout": 1,
        "collision_loop": 25,
        "wrong_target": 1
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 2.1666666666666665,
      "intervention_rate": 0.23193916349809887,
      "expert_disagreement_rate": 0.5695817490494297,
      "mean_steps": 27.395833333333332,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 23,
        "collision_loop": 1,
        "wrong_target": 0
      }
    }
  },
  {
    "seed": 1,
    "method": "bc",
    "expert_queries": 0,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 1.0,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.0,
      "mean_steps": 10.5,
      "failure_taxonomy": {
        "success": 48,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "dynamics_ood": {
      "success_rate": 0.375,
      "collision_cost_per_episode": 10.125,
      "intervention_rate": 0.5022421524663677,
      "expert_disagreement_rate": 0.567713004484305,
      "mean_steps": 23.229166666666668,
      "failure_taxonomy": {
        "success": 18,
        "timeout": 1,
        "collision_loop": 21,
        "wrong_target": 8
      }
    },
    "combined_ood": {
      "success_rate": 0.375,
      "collision_cost_per_episode": 10.6875,
      "intervention_rate": 0.47467320261437906,
      "expert_disagreement_rate": 0.6331699346405228,
      "mean_steps": 25.5,
      "failure_taxonomy": {
        "success": 18,
        "timeout": 0,
        "collision_loop": 27,
        "wrong_target": 3
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.8541666666666666,
      "collision_cost_per_episode": 2.4791666666666665,
      "intervention_rate": 0.2215909090909091,
      "expert_disagreement_rate": 0.509469696969697,
      "mean_steps": 22.0,
      "failure_taxonomy": {
        "success": 41,
        "timeout": 3,
        "collision_loop": 3,
        "wrong_target": 1
      }
    }
  },
  {
    "seed": 1,
    "method": "dagger",
    "expert_queries": 585,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 13.0,
      "intervention_rate": 0.6341463414634146,
      "expert_disagreement_rate": 0.7317073170731707,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.4375,
      "collision_cost_per_episode": 4.791666666666667,
      "intervention_rate": 0.31725888324873097,
      "expert_disagreement_rate": 0.4517766497461929,
      "mean_steps": 16.416666666666668,
      "failure_taxonomy": {
        "success": 21,
        "timeout": 0,
        "collision_loop": 10,
        "wrong_target": 17
      }
    },
    "combined_ood": {
      "success_rate": 0.2916666666666667,
      "collision_cost_per_episode": 8.958333333333334,
      "intervention_rate": 0.4772277227722772,
      "expert_disagreement_rate": 0.6237623762376238,
      "mean_steps": 21.041666666666668,
      "failure_taxonomy": {
        "success": 14,
        "timeout": 0,
        "collision_loop": 21,
        "wrong_target": 13
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.7083333333333334,
      "collision_cost_per_episode": 1.8125,
      "intervention_rate": 0.21160042964554243,
      "expert_disagreement_rate": 0.514500537056928,
      "mean_steps": 19.395833333333332,
      "failure_taxonomy": {
        "success": 34,
        "timeout": 3,
        "collision_loop": 3,
        "wrong_target": 8
      }
    }
  },
  {
    "seed": 1,
    "method": "bc_dr",
    "expert_queries": 0,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "dynamics_ood": {
      "success_rate": 0.375,
      "collision_cost_per_episode": 5.6875,
      "intervention_rate": 0.39648682559598497,
      "expert_disagreement_rate": 0.493099121706399,
      "mean_steps": 16.604166666666668,
      "failure_taxonomy": {
        "success": 18,
        "timeout": 0,
        "collision_loop": 11,
        "wrong_target": 19
      }
    },
    "combined_ood": {
      "success_rate": 0.2708333333333333,
      "collision_cost_per_episode": 8.291666666666666,
      "intervention_rate": 0.4405737704918033,
      "expert_disagreement_rate": 0.5276639344262295,
      "mean_steps": 20.333333333333332,
      "failure_taxonomy": {
        "success": 13,
        "timeout": 0,
        "collision_loop": 19,
        "wrong_target": 16
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.6041666666666666,
      "collision_cost_per_episode": 1.5416666666666667,
      "intervention_rate": 0.22427745664739884,
      "expert_disagreement_rate": 0.45433526011560693,
      "mean_steps": 18.020833333333332,
      "failure_taxonomy": {
        "success": 29,
        "timeout": 2,
        "collision_loop": 4,
        "wrong_target": 13
      }
    }
  },
  {
    "seed": 1,
    "method": "dagger_dr_nochunk",
    "expert_queries": 851,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 14.5,
      "intervention_rate": 0.7073170731707317,
      "expert_disagreement_rate": 0.7560975609756098,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 15.0,
      "intervention_rate": 0.7317073170731707,
      "expert_disagreement_rate": 0.7804878048780488,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.375,
      "collision_cost_per_episode": 11.1875,
      "intervention_rate": 0.5020275750202757,
      "expert_disagreement_rate": 0.6504460665044607,
      "mean_steps": 25.6875,
      "failure_taxonomy": {
        "success": 18,
        "timeout": 0,
        "collision_loop": 30,
        "wrong_target": 0
      }
    },
    "combined_ood": {
      "success_rate": 0.4791666666666667,
      "collision_cost_per_episode": 3.2291666666666665,
      "intervention_rate": 0.09467455621301775,
      "expert_disagreement_rate": 0.1932938856015779,
      "mean_steps": 21.125,
      "failure_taxonomy": {
        "success": 23,
        "timeout": 7,
        "collision_loop": 11,
        "wrong_target": 7
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.5833333333333334,
      "collision_cost_per_episode": 2.4791666666666665,
      "intervention_rate": 0.02601156069364162,
      "expert_disagreement_rate": 0.09441233140655106,
      "mean_steps": 21.625,
      "failure_taxonomy": {
        "success": 28,
        "timeout": 8,
        "collision_loop": 9,
        "wrong_target": 3
      }
    }
  },
  {
    "seed": 1,
    "method": "dagger_dr_chunked",
    "expert_queries": 851,
    "chunk_size": 2,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 11.0,
      "intervention_rate": 0.5365853658536586,
      "expert_disagreement_rate": 0.6341463414634146,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 15.0,
      "intervention_rate": 0.7317073170731707,
      "expert_disagreement_rate": 0.7804878048780488,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.375,
      "collision_cost_per_episode": 11.145833333333334,
      "intervention_rate": 0.49716139497161393,
      "expert_disagreement_rate": 0.64882400648824,
      "mean_steps": 25.6875,
      "failure_taxonomy": {
        "success": 18,
        "timeout": 0,
        "collision_loop": 30,
        "wrong_target": 0
      }
    },
    "combined_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 4.041666666666667,
      "intervention_rate": 0.1683366733466934,
      "expert_disagreement_rate": 0.3927855711422846,
      "mean_steps": 20.791666666666668,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 4,
        "collision_loop": 13,
        "wrong_target": 7
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.6666666666666666,
      "collision_cost_per_episode": 2.0,
      "intervention_rate": 0.13418217433888344,
      "expert_disagreement_rate": 0.32615083251714005,
      "mean_steps": 21.270833333333332,
      "failure_taxonomy": {
        "success": 32,
        "timeout": 10,
        "collision_loop": 4,
        "wrong_target": 2
      }
    }
  },
  {
    "seed": 2,
    "method": "bc",
    "expert_queries": 0,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 1.0,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.0,
      "mean_steps": 10.5,
      "failure_taxonomy": {
        "success": 48,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "dynamics_ood": {
      "success_rate": 0.3333333333333333,
      "collision_cost_per_episode": 11.604166666666666,
      "intervention_rate": 0.5309661436829067,
      "expert_disagreement_rate": 0.6036333608587944,
      "mean_steps": 25.229166666666668,
      "failure_taxonomy": {
        "success": 16,
        "timeout": 0,
        "collision_loop": 26,
        "wrong_target": 6
      }
    },
    "combined_ood": {
      "success_rate": 0.3541666666666667,
      "collision_cost_per_episode": 6.5,
      "intervention_rate": 0.4073226544622426,
      "expert_disagreement_rate": 0.5091533180778032,
      "mean_steps": 18.208333333333332,
      "failure_taxonomy": {
        "success": 17,
        "timeout": 0,
        "collision_loop": 15,
        "wrong_target": 16
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.4375,
      "collision_cost_per_episode": 2.25,
      "intervention_rate": 0.26679841897233203,
      "expert_disagreement_rate": 0.4990118577075099,
      "mean_steps": 21.083333333333332,
      "failure_taxonomy": {
        "success": 21,
        "timeout": 12,
        "collision_loop": 5,
        "wrong_target": 10
      }
    }
  },
  {
    "seed": 2,
    "method": "dagger",
    "expert_queries": 573,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 1.0,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.0,
      "mean_steps": 10.5,
      "failure_taxonomy": {
        "success": 48,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 16.0,
      "intervention_rate": 0.7804878048780488,
      "expert_disagreement_rate": 0.8292682926829268,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.7291666666666666,
      "collision_cost_per_episode": 2.5208333333333335,
      "intervention_rate": 0.03622497616777884,
      "expert_disagreement_rate": 0.24118207816968543,
      "mean_steps": 21.854166666666668,
      "failure_taxonomy": {
        "success": 35,
        "timeout": 4,
        "collision_loop": 6,
        "wrong_target": 3
      }
    },
    "combined_ood": {
      "success_rate": 0.14583333333333334,
      "collision_cost_per_episode": 16.416666666666668,
      "intervention_rate": 0.5918918918918918,
      "expert_disagreement_rate": 0.8148648648648649,
      "mean_steps": 30.833333333333332,
      "failure_taxonomy": {
        "success": 7,
        "timeout": 0,
        "collision_loop": 40,
        "wrong_target": 1
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.3958333333333333,
      "collision_cost_per_episode": 4.8125,
      "intervention_rate": 0.3951219512195122,
      "expert_disagreement_rate": 0.689198606271777,
      "mean_steps": 29.895833333333332,
      "failure_taxonomy": {
        "success": 19,
        "timeout": 15,
        "collision_loop": 13,
        "wrong_target": 1
      }
    }
  },
  {
    "seed": 2,
    "method": "bc_dr",
    "expert_queries": 0,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 0.0,
      "intervention_rate": 0.0,
      "expert_disagreement_rate": 0.14285714285714285,
      "mean_steps": 7.0,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 0,
        "wrong_target": 24
      }
    },
    "dynamics_ood": {
      "success_rate": 0.3333333333333333,
      "collision_cost_per_episode": 8.395833333333334,
      "intervention_rate": 0.48163693599160545,
      "expert_disagreement_rate": 0.5613850996852047,
      "mean_steps": 19.854166666666668,
      "failure_taxonomy": {
        "success": 16,
        "timeout": 0,
        "collision_loop": 17,
        "wrong_target": 15
      }
    },
    "combined_ood": {
      "success_rate": 0.3958333333333333,
      "collision_cost_per_episode": 4.729166666666667,
      "intervention_rate": 0.30324909747292417,
      "expert_disagreement_rate": 0.4368231046931408,
      "mean_steps": 17.3125,
      "failure_taxonomy": {
        "success": 19,
        "timeout": 0,
        "collision_loop": 12,
        "wrong_target": 17
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.6041666666666666,
      "collision_cost_per_episode": 1.4375,
      "intervention_rate": 0.1796116504854369,
      "expert_disagreement_rate": 0.41140776699029125,
      "mean_steps": 17.166666666666668,
      "failure_taxonomy": {
        "success": 29,
        "timeout": 3,
        "collision_loop": 2,
        "wrong_target": 14
      }
    }
  },
  {
    "seed": 2,
    "method": "dagger_dr_nochunk",
    "expert_queries": 933,
    "chunk_size": 1,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 13.5,
      "intervention_rate": 0.6585365853658537,
      "expert_disagreement_rate": 0.7317073170731707,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 15.0,
      "intervention_rate": 0.7317073170731707,
      "expert_disagreement_rate": 0.7804878048780488,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.3333333333333333,
      "collision_cost_per_episode": 11.916666666666666,
      "intervention_rate": 0.5479338842975207,
      "expert_disagreement_rate": 0.6702479338842975,
      "mean_steps": 25.208333333333332,
      "failure_taxonomy": {
        "success": 16,
        "timeout": 0,
        "collision_loop": 30,
        "wrong_target": 2
      }
    },
    "combined_ood": {
      "success_rate": 0.4375,
      "collision_cost_per_episode": 2.9375,
      "intervention_rate": 0.02806122448979592,
      "expert_disagreement_rate": 0.18197278911564627,
      "mean_steps": 24.5,
      "failure_taxonomy": {
        "success": 21,
        "timeout": 15,
        "collision_loop": 11,
        "wrong_target": 1
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.3958333333333333,
      "collision_cost_per_episode": 2.5416666666666665,
      "intervention_rate": 0.01864406779661017,
      "expert_disagreement_rate": 0.16101694915254236,
      "mean_steps": 24.583333333333332,
      "failure_taxonomy": {
        "success": 19,
        "timeout": 17,
        "collision_loop": 8,
        "wrong_target": 4
      }
    }
  },
  {
    "seed": 2,
    "method": "dagger_dr_chunked",
    "expert_queries": 933,
    "chunk_size": 2,
    "in_domain": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 11.0,
      "intervention_rate": 0.5365853658536586,
      "expert_disagreement_rate": 0.6829268292682927,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "visual_ood": {
      "success_rate": 0.5,
      "collision_cost_per_episode": 15.0,
      "intervention_rate": 0.7317073170731707,
      "expert_disagreement_rate": 0.7804878048780488,
      "mean_steps": 20.5,
      "failure_taxonomy": {
        "success": 24,
        "timeout": 0,
        "collision_loop": 24,
        "wrong_target": 0
      }
    },
    "dynamics_ood": {
      "success_rate": 0.3333333333333333,
      "collision_cost_per_episode": 11.145833333333334,
      "intervention_rate": 0.5067300079176563,
      "expert_disagreement_rate": 0.6904196357878069,
      "mean_steps": 26.3125,
      "failure_taxonomy": {
        "success": 16,
        "timeout": 1,
        "collision_loop": 31,
        "wrong_target": 0
      }
    },
    "combined_ood": {
      "success_rate": 0.5833333333333334,
      "collision_cost_per_episode": 3.2708333333333335,
      "intervention_rate": 0.15281757402101243,
      "expert_disagreement_rate": 0.4364851957975167,
      "mean_steps": 21.8125,
      "failure_taxonomy": {
        "success": 28,
        "timeout": 6,
        "collision_loop": 13,
        "wrong_target": 1
      }
    },
    "combined_ood_with_safety_layer": {
      "success_rate": 0.5208333333333334,
      "collision_cost_per_episode": 1.8958333333333333,
      "intervention_rate": 0.13320079522862824,
      "expert_disagreement_rate": 0.37773359840954274,
      "mean_steps": 20.958333333333332,
      "failure_taxonomy": {
        "success": 25,
        "timeout": 14,
        "collision_loop": 5,
        "wrong_target": 4
      }
    }
  }
]
