{
  "provenance": {
    "real": "supplementary_v8/real_data.json",
    "real_task_metrics": {
      "file": "website/task_metrics.csv",
      "source_sha256": {
        "vla": "8986d2d1edb38728e94e099e066b5e6e71c0052bb39cfe99a8ec982accfbcd13",
        "dp": "c2029272c1ff690ec68ed0588b02c06043b339354afc2192ce27dc0dfd5112d6"
      }
    },
    "simulation": "tables/01_dexmimicgen.tex"
  },
  "notes": {
    "real": "Source checkpoint measurements, per-task metrics, and task-averaged point estimates; no confidence intervals are provided. Baseline scores are already negated so higher is better. Scatter coordinates are normalized within each task.",
    "simulation": "Task-specific RACE settings. Eight tasks, ten training seeds per task, twelve checkpoints per run; ten successful reference trajectories per source. Means and 95% hierarchical bootstrap intervals from the paper table."
  },
  "real": {
    "vla": {
      "raw": [
        {
          "task": "fruit",
          "step": 400,
          "success_rate": 0.7,
          "Validation Loss": -0.03145711123943329,
          "Off-Manifold Norm": -0.849388445933305,
          "Valid MSE": -0.013763005167414587,
          "Ours": 0.30216761812563353
        },
        {
          "task": "fruit",
          "step": 600,
          "success_rate": 0.9,
          "Validation Loss": -0.019576378166675568,
          "Off-Manifold Norm": -0.8055763415522423,
          "Valid MSE": -0.013375992934948842,
          "Ours": 0.3876026871394025
        },
        {
          "task": "fruit",
          "step": 800,
          "success_rate": 0.9329999999999999,
          "Validation Loss": -0.0186528991907835,
          "Off-Manifold Norm": -0.7777713457205331,
          "Valid MSE": -0.013568150408796992,
          "Ours": 0.3484724269117447
        },
        {
          "task": "fruit",
          "step": 1000,
          "success_rate": 0.833,
          "Validation Loss": -0.018624763935804367,
          "Off-Manifold Norm": -0.7695595242729536,
          "Valid MSE": -0.013970970718664521,
          "Ours": 0.344695868847409
        },
        {
          "task": "fruit",
          "step": 1200,
          "success_rate": 0.778,
          "Validation Loss": -0.018863920122385025,
          "Off-Manifold Norm": -0.7556088976316473,
          "Valid MSE": -0.0141139874314483,
          "Ours": 0.3436626520642078
        },
        {
          "task": "fruit",
          "step": 1400,
          "success_rate": 0.778,
          "Validation Loss": -0.01926344819366932,
          "Off-Manifold Norm": -0.7589406872811837,
          "Valid MSE": -0.014213233757399277,
          "Ours": 0.33321650533739616
        },
        {
          "task": "fruit",
          "step": 1600,
          "success_rate": 0.8440000000000001,
          "Validation Loss": -0.01992371864616871,
          "Off-Manifold Norm": -0.750891426653615,
          "Valid MSE": -0.014275959092864443,
          "Ours": 0.3162676544549373
        },
        {
          "task": "fruit",
          "step": 1800,
          "success_rate": 0.711,
          "Validation Loss": -0.02062116377055645,
          "Off-Manifold Norm": -0.7437618818218221,
          "Valid MSE": -0.014569021151830413,
          "Ours": 0.32575078336216784
        },
        {
          "task": "fruit",
          "step": 1999,
          "success_rate": 0.7559999999999999,
          "Validation Loss": -0.021131418645381927,
          "Off-Manifold Norm": -0.7374340006021682,
          "Valid MSE": -0.014657992179988363,
          "Ours": 0.31404210367275387
        },
        {
          "task": "stack_cup",
          "step": 400,
          "success_rate": 0.65,
          "Validation Loss": -0.02716926671564579,
          "Off-Manifold Norm": -0.7906451182771915,
          "Valid MSE": -0.011666157639718545,
          "Ours": 0.22518065080604438
        },
        {
          "task": "stack_cup",
          "step": 600,
          "success_rate": 0.9,
          "Validation Loss": -0.01696336269378662,
          "Off-Manifold Norm": -0.761608881460283,
          "Valid MSE": -0.011875048243030482,
          "Ours": 0.27264078255133606
        },
        {
          "task": "stack_cup",
          "step": 800,
          "success_rate": 0.9670000000000001,
          "Validation Loss": -0.016206564381718636,
          "Off-Manifold Norm": -0.7469830358188775,
          "Valid MSE": -0.011660181143431739,
          "Ours": 0.2742133808408556
        },
        {
          "task": "stack_cup",
          "step": 1000,
          "success_rate": 0.9670000000000001,
          "Validation Loss": -0.01633152738213539,
          "Off-Manifold Norm": -0.7494767205904512,
          "Valid MSE": -0.011794899275379144,
          "Ours": 0.2744220791416658
        },
        {
          "task": "stack_cup",
          "step": 1200,
          "success_rate": 0.95,
          "Validation Loss": -0.016858812421560287,
          "Off-Manifold Norm": -0.7663835027851922,
          "Valid MSE": -0.012411684188133348,
          "Ours": 0.2596076096874775
        },
        {
          "task": "stack_cup",
          "step": 1400,
          "success_rate": 0.85,
          "Validation Loss": -0.01754317246377468,
          "Off-Manifold Norm": -0.75903956020892,
          "Valid MSE": -0.0124679204103923,
          "Ours": 0.25513594004956897
        },
        {
          "task": "stack_cup",
          "step": 1600,
          "success_rate": 0.883,
          "Validation Loss": -0.018064280971884727,
          "Off-Manifold Norm": -0.7545573115055791,
          "Valid MSE": -0.012625381791928916,
          "Ours": 0.26850281664390135
        },
        {
          "task": "stack_cup",
          "step": 1800,
          "success_rate": 0.867,
          "Validation Loss": -0.01830761879682541,
          "Off-Manifold Norm": -0.7525040214740226,
          "Valid MSE": -0.012576707546304631,
          "Ours": 0.27104510236997564
        },
        {
          "task": "stack_cup",
          "step": 1999,
          "success_rate": 0.9329999999999999,
          "Validation Loss": -0.01896173134446144,
          "Off-Manifold Norm": -0.7453701406097821,
          "Valid MSE": -0.012594179715718083,
          "Ours": 0.2511498045325992
        },
        {
          "task": "two_cubes",
          "step": 400,
          "success_rate": 0.467,
          "Validation Loss": -0.029780499637126923,
          "Off-Manifold Norm": -0.7951902307573434,
          "Valid MSE": -0.013634035122780722,
          "Ours": 0.0031926521763309774
        },
        {
          "task": "two_cubes",
          "step": 600,
          "success_rate": 0.7,
          "Validation Loss": -0.01917542889714241,
          "Off-Manifold Norm": -0.7850233720591908,
          "Valid MSE": -0.013835715684859277,
          "Ours": 0.003479826110937466
        },
        {
          "task": "two_cubes",
          "step": 800,
          "success_rate": 0.867,
          "Validation Loss": -0.018522106111049652,
          "Off-Manifold Norm": -0.7572253878803933,
          "Valid MSE": -0.014081986238581362,
          "Ours": 0.004460115390535956
        },
        {
          "task": "two_cubes",
          "step": 1000,
          "success_rate": 0.7170000000000001,
          "Validation Loss": -0.018556931987404823,
          "Off-Manifold Norm": -0.7455926148744845,
          "Valid MSE": -0.014353379677362907,
          "Ours": 0.003911323661954446
        },
        {
          "task": "two_cubes",
          "step": 1200,
          "success_rate": 0.833,
          "Validation Loss": -0.019167261198163033,
          "Off-Manifold Norm": -0.7706976973557812,
          "Valid MSE": -0.014787799846380486,
          "Ours": 0.003913120955425934
        },
        {
          "task": "two_cubes",
          "step": 1400,
          "success_rate": 0.833,
          "Validation Loss": -0.01976574771106243,
          "Off-Manifold Norm": -0.7421383494669315,
          "Valid MSE": -0.01462943082947866,
          "Ours": 0.004036573301100974
        },
        {
          "task": "two_cubes",
          "step": 1600,
          "success_rate": 0.8170000000000001,
          "Validation Loss": -0.02095158025622368,
          "Off-Manifold Norm": -0.7523062639154384,
          "Valid MSE": -0.01509461563445941,
          "Ours": 0.004017233530728814
        },
        {
          "task": "two_cubes",
          "step": 1800,
          "success_rate": 0.833,
          "Validation Loss": -0.02188330888748169,
          "Off-Manifold Norm": -0.739468011526368,
          "Valid MSE": -0.015176821977596094,
          "Ours": 0.004211353665987673
        },
        {
          "task": "two_cubes",
          "step": 1999,
          "success_rate": 0.633,
          "Validation Loss": -0.022230282425880432,
          "Off-Manifold Norm": -0.7496402823880868,
          "Valid MSE": -0.0152954450274174,
          "Ours": 0.0038885283504318533
        }
      ],
      "zscore": [
        {
          "task": "fruit",
          "step": 400,
          "success_rate": -1.2930834040542667,
          "Validation Loss": -2.6051058989716,
          "Off-Manifold Norm": -2.1824020455696456,
          "Valid MSE": 0.6787606857050125,
          "Ours": -1.3075142391772057
        },
        {
          "task": "fruit",
          "step": 600,
          "success_rate": 1.2016112661468885,
          "Validation Loss": 0.3270784896109432,
          "Off-Manifold Norm": -0.9452171112009371,
          "Valid MSE": 1.5738609494609008,
          "Ours": 2.0847623618403133
        },
        {
          "task": "fruit",
          "step": 800,
          "success_rate": 1.613235886730078,
          "Validation Loss": 0.5549946139537978,
          "Off-Manifold Norm": -0.16004775151949574,
          "Valid MSE": 1.1294300217140159,
          "Ours": 0.5310600274907363
        },
        {
          "task": "fruit",
          "step": 1000,
          "success_rate": 0.3658885516295011,
          "Validation Loss": 0.561938440981422,
          "Off-Manifold Norm": 0.07184117329240024,
          "Valid MSE": 0.1977680860914313,
          "Ours": 0.3811083776364442
        },
        {
          "task": "fruit",
          "step": 1200,
          "success_rate": -0.32015248267581553,
          "Validation Loss": 0.5029143022830541,
          "Off-Manifold Norm": 0.46578493265875365,
          "Valid MSE": -0.1330077566219938,
          "Ours": 0.34008357180563176
        },
        {
          "task": "fruit",
          "step": 1400,
          "success_rate": -0.32015248267581553,
          "Validation Loss": 0.4043101196752903,
          "Off-Manifold Norm": 0.37170043210520926,
          "Valid MSE": -0.3625493707708057,
          "Ours": -0.07469012149410588
        },
        {
          "task": "fruit",
          "step": 1600,
          "success_rate": 0.5030967584905661,
          "Validation Loss": 0.24135429016504487,
          "Off-Manifold Norm": 0.5989988949825065,
          "Valid MSE": -0.5076235043746209,
          "Ours": -0.7476595581247963
        },
        {
          "task": "fruit",
          "step": 1800,
          "success_rate": -1.155875197193203,
          "Validation Loss": 0.06922369072316581,
          "Off-Manifold Norm": 0.8003260300153765,
          "Valid MSE": -1.185431346793403,
          "Ours": -0.37112335738448426
        },
        {
          "task": "fruit",
          "step": 1999,
          "success_rate": -0.5945688963979442,
          "Validation Loss": -0.0567080484211192,
          "Off-Manifold Norm": 0.979015445235851,
          "Valid MSE": -1.3912077644105452,
          "Ours": -0.836027062592551
        },
        {
          "task": "stack_cup",
          "step": 400,
          "success_rate": -2.397925258802999,
          "Validation Loss": -2.5660919630151544,
          "Off-Manifold Norm": -2.3137867931139415,
          "Valid MSE": 1.2251795950194682,
          "Ours": -2.2506660310127597
        },
        {
          "task": "stack_cup",
          "step": 600,
          "success_rate": 0.15064905971695863,
          "Validation Loss": 0.45122057676193467,
          "Off-Manifold Norm": -0.22328198895338147,
          "Valid MSE": 0.7326664440649103,
          "Ours": 0.7048642414027789
        },
        {
          "task": "stack_cup",
          "step": 800,
          "success_rate": 0.8336669770803079,
          "Validation Loss": 0.6749633272516548,
          "Off-Manifold Norm": 0.8297263871342648,
          "Valid MSE": 1.239270717220344,
          "Ours": 0.802796161945436
        },
        {
          "task": "stack_cup",
          "step": 1000,
          "success_rate": 0.8336669770803079,
          "Validation Loss": 0.6380187873698305,
          "Off-Manifold Norm": 0.6501900364732434,
          "Valid MSE": 0.9216381832329387,
          "Ours": 0.8157926312590154
        },
        {
          "task": "stack_cup",
          "step": 1200,
          "success_rate": 0.6603639234209495,
          "Validation Loss": 0.4821302195285103,
          "Off-Manifold Norm": -0.5670375823764803,
          "Valid MSE": -0.5325903748041025,
          "Ours": -0.10676302401416018
        },
        {
          "task": "stack_cup",
          "step": 1400,
          "success_rate": -0.35906580398703336,
          "Validation Loss": 0.27980339647852165,
          "Off-Manifold Norm": -0.03830008329960164,
          "Valid MSE": -0.6651816861575252,
          "Ours": -0.3852315932547242
        },
        {
          "task": "stack_cup",
          "step": 1600,
          "success_rate": -0.02265399394239864,
          "Validation Loss": 0.12574088200285194,
          "Off-Manifold Norm": 0.2844057310544177,
          "Valid MSE": -1.0364372617778814,
          "Ours": 0.44717672440889744
        },
        {
          "task": "stack_cup",
          "step": 1800,
          "success_rate": -0.18576275032767608,
          "Validation Loss": 0.05379955578392781,
          "Off-Manifold Norm": 0.432235241550913,
          "Valid MSE": -0.9216752490177658,
          "Ours": 0.605494920086639
        },
        {
          "task": "stack_cup",
          "step": 1999,
          "success_rate": 0.48706086976159224,
          "Validation Loss": -0.13958478216208042,
          "Off-Manifold Norm": 0.9458490515305662,
          "Valid MSE": -0.9628703677803866,
          "Ours": -0.633464030821095
        },
        {
          "task": "two_cubes",
          "step": 400,
          "success_rate": -2.128416529751334,
          "Validation Loss": -2.4560143368003247,
          "Off-Manifold Norm": -1.802851492707323,
          "Valid MSE": 1.5063489758707445,
          "Ours": -1.8933371998987845
        },
        {
          "task": "two_cubes",
          "step": 600,
          "success_rate": -0.34095579171026624,
          "Validation Loss": 0.549650434433517,
          "Off-Manifold Norm": -1.2864187959750353,
          "Valid MSE": 1.1722122862574302,
          "Ours": -1.1259601970782893
        },
        {
          "task": "two_cubes",
          "step": 800,
          "success_rate": 0.9401855956410576,
          "Validation Loss": 0.7348136766738242,
          "Off-Manifold Norm": 0.12559924808280634,
          "Valid MSE": 0.7642005933016948,
          "Ours": 1.4935374613918426
        },
        {
          "task": "two_cubes",
          "step": 1000,
          "success_rate": -0.21054020138108862,
          "Validation Loss": 0.7249434068515376,
          "Off-Manifold Norm": 0.7164940871157982,
          "Valid MSE": 0.3145662592704467,
          "Ours": 0.02707375733517686
        },
        {
          "task": "two_cubes",
          "step": 1200,
          "success_rate": 0.6793544149827041,
          "Validation Loss": 0.5519653049031499,
          "Off-Manifold Norm": -0.5587361322103,
          "Valid MSE": -0.4051645649636309,
          "Ours": 0.03187642746340143
        },
        {
          "task": "two_cubes",
          "step": 1400,
          "success_rate": 0.6793544149827041,
          "Validation Loss": 0.3823436331590783,
          "Off-Manifold Norm": 0.8919559138819951,
          "Valid MSE": -0.14278479717877202,
          "Ours": 0.361761836962999
        },
        {
          "task": "two_cubes",
          "step": 1600,
          "success_rate": 0.5566103299670091,
          "Validation Loss": 0.04625769908458957,
          "Off-Manifold Norm": 0.37546958957222587,
          "Valid MSE": -0.9134853009493884,
          "Ours": 0.31008272115549607
        },
        {
          "task": "two_cubes",
          "step": 1800,
          "success_rate": 0.6793544149827041,
          "Validation Loss": -0.2178106878205052,
          "Off-Manifold Norm": 1.0275975968024362,
          "Valid MSE": -1.0496816456928584,
          "Ours": 0.828804337677468
        },
        {
          "task": "two_cubes",
          "step": 1999,
          "success_rate": -0.8549466477134915,
          "Validation Loss": -0.31614913048486565,
          "Off-Manifold Norm": 0.5108899854373569,
          "Valid MSE": -1.246211805915675,
          "Ours": -0.03383914500930772
        }
      ],
      "metrics": {
        "Validation Loss": {
          "pearson": 0.7692187506440756,
          "spearman": 0.6887789481645624,
          "mmrv": 0.06611111111111108,
          "nregret": 0.14306151645146037,
          "regret_pp": 3.3333333333333326,
          "balanced_loss": 0.17316726471672572,
          "pooled_spearman": 0.5841737864870744,
          "pooled_pearson": 0.7692187506440755
        },
        "Off-Manifold Norm": {
          "pearson": 0.4867419676465752,
          "spearman": 0.22480261118630396,
          "mmrv": 0.11162962962962963,
          "nregret": 0.3173040576204788,
          "regret_pp": 8.166666666666673,
          "balanced_loss": 0.3624661490481103,
          "pooled_spearman": 0.20103888677222537,
          "pooled_pearson": 0.4867419676465751
        },
        "Valid MSE": {
          "pearson": 0.004726098734282276,
          "spearman": 0.20283952494073798,
          "mmrv": 0.21325925925925926,
          "nregret": 0.3805436337614818,
          "regret_pp": 14.43333333333333,
          "balanced_loss": 0.4709902593895821,
          "pooled_spearman": 0.21081585391008434,
          "pooled_pearson": 0.004726098734282175
        },
        "Ours": {
          "pearson": 0.791261390083532,
          "spearman": 0.7905601535128192,
          "mmrv": 0.04203703703703702,
          "nregret": 0.04721030042898181,
          "regret_pp": 1.0999999999999972,
          "balanced_loss": 0.10080591952223171,
          "pooled_spearman": 0.7693251016602788,
          "pooled_pearson": 0.7912613900835316
        }
      },
      "checkpoint": "step",
      "steps": [
        400,
        600,
        800,
        1000,
        1200,
        1400,
        1600,
        1800,
        1999
      ],
      "legend": [
        "stack_cup",
        "two_cubes",
        "fruit"
      ],
      "tasks": {
        "fruit": {
          "name": "PnP Three Fruit",
          "color": "#4285F4",
          "marker": "square"
        },
        "stack_cup": {
          "name": "Nest Cups",
          "color": "#EA4335",
          "marker": "triangle"
        },
        "two_cubes": {
          "name": "Put Two Cubes",
          "color": "#34A853",
          "marker": "diamond"
        }
      },
      "axes": {
        "xlim": [
          -3.5,
          3.0
        ],
        "ylim": [
          -3.0,
          5.5
        ]
      },
      "method_keys": {
        "race": "Ours",
        "loss": "Validation Loss",
        "mse": "Valid MSE",
        "omn": "Off-Manifold Norm"
      },
      "task_metrics": {
        "fruit": {
          "Validation Loss": {
            "spearman": 0.6778302010921357,
            "mmrv": 0.08022222222222221,
            "nregret": 0.4291845493543811
          },
          "Off-Manifold Norm": {
            "spearman": -0.25941649671427414,
            "mmrv": 0.1527777777777778,
            "nregret": 0.7596566523572549
          },
          "Valid MSE": {
            "spearman": 0.510464719340991,
            "mmrv": 0.08266666666666668,
            "nregret": 0.14163090128694544
          },
          "Ours": {
            "spearman": 0.7782494901428224,
            "mmrv": 0.0577777777777778,
            "nregret": 0.14163090128694544
          }
        },
        "stack_cup": {
          "Validation Loss": {
            "spearman": 0.7782494901428224,
            "mmrv": 0.03499999999999997,
            "nregret": 0.0
          },
          "Off-Manifold Norm": {
            "spearman": 0.54393781569122,
            "mmrv": 0.055777777777777815,
            "nregret": 0.10725552050439395
          },
          "Valid MSE": {
            "spearman": 0.2845213189769458,
            "mmrv": 0.23522222222222222,
            "nregret": 0.0
          },
          "Ours": {
            "spearman": 0.6610936529170212,
            "mmrv": 0.04988888888888886,
            "nregret": 0.0
          }
        },
        "two_cubes": {
          "Validation Loss": {
            "spearman": 0.6102571532587293,
            "mmrv": 0.0831111111111111,
            "nregret": 0.0
          },
          "Off-Manifold Norm": {
            "spearman": 0.38988651458196594,
            "mmrv": 0.12633333333333333,
            "nregret": 0.08499999999978759
          },
          "Valid MSE": {
            "spearman": -0.18646746349572285,
            "mmrv": 0.3218888888888889,
            "nregret": 0.9999999999975001
          },
          "Ours": {
            "spearman": 0.9323373174786143,
            "mmrv": 0.018444444444444413,
            "nregret": 0.0
          }
        }
      }
    },
    "dp": {
      "raw": [
        {
          "task": "apple",
          "epoch": 5,
          "successes": 11,
          "trials": 30,
          "success_rate": 0.36666666666666664,
          "Loss": -0.021055004801601173,
          "Off_Manifold_Norm_PolicyReference": -0.2076067233322499,
          "MSE_mean_pred": -0.023078521713614464,
          "SurVAL": 0.005777882819087753
        },
        {
          "task": "apple",
          "epoch": 10,
          "successes": 9,
          "trials": 30,
          "success_rate": 0.3,
          "Loss": -0.026134621612727643,
          "Off_Manifold_Norm_PolicyReference": -0.1718211664952454,
          "MSE_mean_pred": -0.027358384802937508,
          "SurVAL": 0.013108038368560285
        },
        {
          "task": "apple",
          "epoch": 15,
          "successes": 14,
          "trials": 30,
          "success_rate": 0.4666666666666667,
          "Loss": -0.03630741186439991,
          "Off_Manifold_Norm_PolicyReference": -0.17018492837936353,
          "MSE_mean_pred": -0.029620902612805367,
          "SurVAL": 0.016614779491305916
        },
        {
          "task": "apple",
          "epoch": 20,
          "successes": 14,
          "trials": 30,
          "success_rate": 0.4666666666666667,
          "Loss": -0.044215056244283915,
          "Off_Manifold_Norm_PolicyReference": -0.17065818908835995,
          "MSE_mean_pred": -0.030683642253279686,
          "SurVAL": 0.018616006726114152
        },
        {
          "task": "apple",
          "epoch": 25,
          "successes": 17,
          "trials": 30,
          "success_rate": 0.5666666666666667,
          "Loss": -0.04551219753921032,
          "Off_Manifold_Norm_PolicyReference": -0.16765697999232454,
          "MSE_mean_pred": -0.03164514899253845,
          "SurVAL": 0.020038886478269666
        },
        {
          "task": "apple",
          "epoch": 30,
          "successes": 20,
          "trials": 30,
          "success_rate": 0.6666666666666666,
          "Loss": -0.05258262291550636,
          "Off_Manifold_Norm_PolicyReference": -0.1681531897776291,
          "MSE_mean_pred": -0.031992293894290924,
          "SurVAL": 0.021954941515385844
        },
        {
          "task": "apple",
          "epoch": 35,
          "successes": 20,
          "trials": 30,
          "success_rate": 0.6666666666666666,
          "Loss": -0.052979411631822584,
          "Off_Manifold_Norm_PolicyReference": -0.16847168438247634,
          "MSE_mean_pred": -0.03226938843727112,
          "SurVAL": 0.023775728847683968
        },
        {
          "task": "apple",
          "epoch": 40,
          "successes": 21,
          "trials": 30,
          "success_rate": 0.7,
          "Loss": -0.05979545276612044,
          "Off_Manifold_Norm_PolicyReference": -0.168091521880727,
          "MSE_mean_pred": -0.03267678990960121,
          "SurVAL": 0.024815965695788103
        },
        {
          "task": "apple",
          "epoch": 45,
          "successes": 19,
          "trials": 30,
          "success_rate": 0.6333333333333333,
          "Loss": -0.06097684571519494,
          "Off_Manifold_Norm_PolicyReference": -0.16881644521807568,
          "MSE_mean_pred": -0.03281441330909729,
          "SurVAL": 0.024803114233616943
        },
        {
          "task": "apple",
          "epoch": 50,
          "successes": 21,
          "trials": 30,
          "success_rate": 0.7,
          "Loss": -0.06487621514126658,
          "Off_Manifold_Norm_PolicyReference": -0.17016112807608705,
          "MSE_mean_pred": -0.033486317843198776,
          "SurVAL": 0.024827208816892384
        },
        {
          "task": "pan",
          "epoch": 5,
          "successes": 6,
          "trials": 30,
          "success_rate": 0.2,
          "Loss": -0.024748352449387313,
          "Off_Manifold_Norm_PolicyReference": -0.32063024336580703,
          "MSE_mean_pred": -0.0497521236538887,
          "SurVAL": 0.9998570829768064
        },
        {
          "task": "pan",
          "epoch": 10,
          "successes": 3,
          "trials": 30,
          "success_rate": 0.1,
          "Loss": -0.02958693640306592,
          "Off_Manifold_Norm_PolicyReference": -0.2626761623489853,
          "MSE_mean_pred": -0.050598155707120895,
          "SurVAL": 0.9992325557185996
        },
        {
          "task": "pan",
          "epoch": 15,
          "successes": 5,
          "trials": 30,
          "success_rate": 0.16666666666666666,
          "Loss": -0.03580035166814923,
          "Off_Manifold_Norm_PolicyReference": -0.2554486009750933,
          "MSE_mean_pred": -0.0537479929625988,
          "SurVAL": 0.9991023716742201
        },
        {
          "task": "pan",
          "epoch": 20,
          "successes": 5,
          "trials": 30,
          "success_rate": 0.16666666666666666,
          "Loss": -0.04076145434752107,
          "Off_Manifold_Norm_PolicyReference": -0.25958551802263874,
          "MSE_mean_pred": -0.05617228150367737,
          "SurVAL": 0.9995179163468613
        },
        {
          "task": "pan",
          "epoch": 25,
          "successes": 5,
          "trials": 30,
          "success_rate": 0.16666666666666666,
          "Loss": -0.047915628738701345,
          "Off_Manifold_Norm_PolicyReference": -0.2602795232711401,
          "MSE_mean_pred": -0.055979687720537186,
          "SurVAL": 0.9996847870827048
        },
        {
          "task": "pan",
          "epoch": 30,
          "successes": 13,
          "trials": 30,
          "success_rate": 0.43333333333333335,
          "Loss": -0.04803862879052758,
          "Off_Manifold_Norm_PolicyReference": -0.259760037376398,
          "MSE_mean_pred": -0.05727466940879822,
          "SurVAL": 1.0
        },
        {
          "task": "pan",
          "epoch": 35,
          "successes": 6,
          "trials": 30,
          "success_rate": 0.2,
          "Loss": -0.05458815947175026,
          "Off_Manifold_Norm_PolicyReference": -0.2608532131022173,
          "MSE_mean_pred": -0.05812300369143486,
          "SurVAL": 1.0
        },
        {
          "task": "pan",
          "epoch": 40,
          "successes": 9,
          "trials": 30,
          "success_rate": 0.3,
          "Loss": -0.05928192753344774,
          "Off_Manifold_Norm_PolicyReference": -0.2569149178140425,
          "MSE_mean_pred": -0.05803405120968819,
          "SurVAL": 1.0
        },
        {
          "task": "pan",
          "epoch": 45,
          "successes": 6,
          "trials": 30,
          "success_rate": 0.2,
          "Loss": -0.06908529058098793,
          "Off_Manifold_Norm_PolicyReference": -0.2573961120226072,
          "MSE_mean_pred": -0.0578404925763607,
          "SurVAL": 1.0
        },
        {
          "task": "pan",
          "epoch": 50,
          "successes": 8,
          "trials": 30,
          "success_rate": 0.26666666666666666,
          "Loss": -0.0758158553391695,
          "Off_Manifold_Norm_PolicyReference": -0.2595104275391801,
          "MSE_mean_pred": -0.05848432704806328,
          "SurVAL": 1.0
        },
        {
          "task": "pet",
          "epoch": 5,
          "successes": 2,
          "trials": 30,
          "success_rate": 0.06666666666666667,
          "Loss": -0.018155722636729477,
          "Off_Manifold_Norm_PolicyReference": -0.27364995920554636,
          "MSE_mean_pred": -0.033309731632471085,
          "SurVAL": 0.9744407105678959
        },
        {
          "task": "pet",
          "epoch": 10,
          "successes": 8,
          "trials": 30,
          "success_rate": 0.26666666666666666,
          "Loss": -0.01959100306034088,
          "Off_Manifold_Norm_PolicyReference": -0.20893292699146682,
          "MSE_mean_pred": -0.03544303774833679,
          "SurVAL": 0.9930918291991415
        },
        {
          "task": "pet",
          "epoch": 15,
          "successes": 11,
          "trials": 30,
          "success_rate": 0.36666666666666664,
          "Loss": -0.023065991625189783,
          "Off_Manifold_Norm_PolicyReference": -0.201476601128545,
          "MSE_mean_pred": -0.03823675587773323,
          "SurVAL": 0.9955062086502038
        },
        {
          "task": "pet",
          "epoch": 20,
          "successes": 12,
          "trials": 30,
          "success_rate": 0.4,
          "Loss": -0.02562894804403186,
          "Off_Manifold_Norm_PolicyReference": -0.1934667661401963,
          "MSE_mean_pred": -0.04035722091794014,
          "SurVAL": 0.9970493376773741
        },
        {
          "task": "pet",
          "epoch": 25,
          "successes": 14,
          "trials": 30,
          "success_rate": 0.4666666666666667,
          "Loss": -0.0317832965310663,
          "Off_Manifold_Norm_PolicyReference": -0.19124530557948147,
          "MSE_mean_pred": -0.04110841453075409,
          "SurVAL": 0.9976609615581727
        },
        {
          "task": "pet",
          "epoch": 30,
          "successes": 15,
          "trials": 30,
          "success_rate": 0.5,
          "Loss": -0.02957350343465805,
          "Off_Manifold_Norm_PolicyReference": -0.18870164179217858,
          "MSE_mean_pred": -0.04133595526218414,
          "SurVAL": 0.9978473359647585
        },
        {
          "task": "pet",
          "epoch": 35,
          "successes": 9,
          "trials": 30,
          "success_rate": 0.3,
          "Loss": -0.03700800932943821,
          "Off_Manifold_Norm_PolicyReference": -0.18964383503382143,
          "MSE_mean_pred": -0.041335828602313995,
          "SurVAL": 0.9973666187179062
        },
        {
          "task": "pet",
          "epoch": 40,
          "successes": 11,
          "trials": 30,
          "success_rate": 0.36666666666666664,
          "Loss": -0.0381337951309979,
          "Off_Manifold_Norm_PolicyReference": -0.18767076712274103,
          "MSE_mean_pred": -0.042108204215765,
          "SurVAL": 0.9954716261720052
        },
        {
          "task": "pet",
          "epoch": 45,
          "successes": 15,
          "trials": 30,
          "success_rate": 0.5,
          "Loss": -0.04090976714156568,
          "Off_Manifold_Norm_PolicyReference": -0.18492974028948983,
          "MSE_mean_pred": -0.041501112282276154,
          "SurVAL": 0.9977441540936831
        },
        {
          "task": "pet",
          "epoch": 50,
          "successes": 12,
          "trials": 30,
          "success_rate": 0.4,
          "Loss": -0.04034906953573227,
          "Off_Manifold_Norm_PolicyReference": -0.18328344267254398,
          "MSE_mean_pred": -0.041629187762737274,
          "SurVAL": 0.9977411890261657
        }
      ],
      "zscore": [
        {
          "task": "apple",
          "epoch": 5,
          "success_rate": -1.286987967375382,
          "Loss": 1.7137414798206152,
          "Off_Manifold_Norm_PolicyReference": -2.8288176878172755,
          "MSE_mean_pred": 2.348503610081748,
          "SurVAL": -2.1973542933637824
        },
        {
          "task": "apple",
          "epoch": 10,
          "success_rate": -1.7466265271523034,
          "Loss": 1.3708635171496097,
          "Off_Manifold_Norm_PolicyReference": 0.11013438952791257,
          "MSE_mean_pred": 1.0054791405798904,
          "SurVAL": -1.0178222966326127
        },
        {
          "task": "apple",
          "epoch": 15,
          "success_rate": -0.597530127709999,
          "Loss": 0.6841925091057298,
          "Off_Manifold_Norm_PolicyReference": 0.244513335525751,
          "MSE_mean_pred": 0.2954992375038529,
          "SurVAL": -0.45353504097824904
        },
        {
          "task": "apple",
          "epoch": 20,
          "success_rate": -0.597530127709999,
          "Loss": 0.1504205543861423,
          "Off_Manifold_Norm_PolicyReference": 0.20564596369286062,
          "MSE_mean_pred": -0.03798931390034047,
          "SurVAL": -0.13150755598769978
        },
        {
          "task": "apple",
          "epoch": 25,
          "success_rate": 0.0919277119553834,
          "Loss": 0.06286253947317019,
          "Off_Manifold_Norm_PolicyReference": 0.45212556292099615,
          "MSE_mean_pred": -0.33971089972324525,
          "SurVAL": 0.09745514253090334
        },
        {
          "task": "apple",
          "epoch": 30,
          "success_rate": 0.7813855516207658,
          "Loss": -0.4143965015890523,
          "Off_Manifold_Norm_PolicyReference": 0.411373457652318,
          "MSE_mean_pred": -0.4486452482727293,
          "SurVAL": 0.4057771431382563
        },
        {
          "task": "apple",
          "epoch": 35,
          "success_rate": 0.7813855516207658,
          "Loss": -0.4411800389541827,
          "Off_Manifold_Norm_PolicyReference": 0.38521652554693075,
          "MSE_mean_pred": -0.5355977380652702,
          "SurVAL": 0.6987691408105253
        },
        {
          "task": "apple",
          "epoch": 40,
          "success_rate": 1.0112048315092266,
          "Loss": -0.901267951981532,
          "Off_Manifold_Norm_PolicyReference": 0.4164380426336675,
          "MSE_mean_pred": -0.6634406467824608,
          "SurVAL": 0.8661588555630182
        },
        {
          "task": "apple",
          "epoch": 45,
          "success_rate": 0.551566271732305,
          "Loss": -0.9810128663233534,
          "Off_Manifold_Norm_PolicyReference": 0.3569024328366507,
          "MSE_mean_pred": -0.7066269799061445,
          "SurVAL": 0.864090862498839
        },
        {
          "task": "apple",
          "epoch": 50,
          "success_rate": 1.0112048315092266,
          "Loss": -1.2442232410871448,
          "Off_Manifold_Norm_PolicyReference": 0.24646797748016985,
          "MSE_mean_pred": -0.9174711615152982,
          "SurVAL": 0.8679680424208028
        },
        {
          "task": "pan",
          "epoch": 5,
          "success_rate": -0.21452908258025843,
          "Loss": 1.4384312936321468,
          "Off_Manifold_Norm_PolicyReference": -2.8298400684039806,
          "MSE_mean_pred": 1.8337370053578075,
          "SurVAL": 0.341394351353305
        },
        {
          "task": "pan",
          "epoch": 10,
          "success_rate": -1.2871744954815496,
          "Loss": 1.1461663224168526,
          "Off_Manifold_Norm_PolicyReference": 0.13448833574166538,
          "MSE_mean_pred": 1.5684748488488116,
          "SurVAL": -1.4714377490604318
        },
        {
          "task": "pan",
          "epoch": 15,
          "success_rate": -0.5720775535473557,
          "Loss": 0.7708574198571209,
          "Off_Manifold_Norm_PolicyReference": 0.5041752516141662,
          "MSE_mean_pred": 0.5808850526367588,
          "SurVAL": -1.8493264825168743
        },
        {
          "task": "pan",
          "epoch": 20,
          "success_rate": -0.5720775535473557,
          "Loss": 0.47119195044217366,
          "Off_Manifold_Norm_PolicyReference": 0.2925735750027513,
          "MSE_mean_pred": -0.17921854385198724,
          "SurVAL": -0.6431137643038257
        },
        {
          "task": "pan",
          "epoch": 25,
          "success_rate": -0.5720775535473557,
          "Loss": 0.03905837737109974,
          "Off_Manifold_Norm_PolicyReference": 0.2570754800513074,
          "MSE_mean_pred": -0.11883331134121654,
          "SurVAL": -0.15873358300412832
        },
        {
          "task": "pan",
          "epoch": 30,
          "success_rate": 2.288310214189421,
          "Loss": 0.03162880563056105,
          "Off_Manifold_Norm_PolicyReference": 0.28364697883652745,
          "MSE_mean_pred": -0.5248576858970415,
          "SurVAL": 0.7562434455061977
        },
        {
          "task": "pan",
          "epoch": 35,
          "success_rate": -0.21452908258025843,
          "Loss": -0.3639824753447202,
          "Off_Manifold_Norm_PolicyReference": 0.22773147128446958,
          "MSE_mean_pred": -0.7908416759751906,
          "SurVAL": 0.7562434455061977
        },
        {
          "task": "pan",
          "epoch": 40,
          "success_rate": 0.8581163303210325,
          "Loss": -0.6475001326957585,
          "Off_Manifold_Norm_PolicyReference": 0.42917372318146546,
          "MSE_mean_pred": -0.762951802246643,
          "SurVAL": 0.7562434455061977
        },
        {
          "task": "pan",
          "epoch": 45,
          "success_rate": -0.21452908258025843,
          "Loss": -1.2396526397787695,
          "Off_Manifold_Norm_PolicyReference": 0.4045608290533601,
          "MSE_mean_pred": -0.702264053725499,
          "SurVAL": 0.7562434455061977
        },
        {
          "task": "pan",
          "epoch": 50,
          "success_rate": 0.5005678593539356,
          "Loss": -1.6461989215307038,
          "Off_Manifold_Norm_PolicyReference": 0.2964144236382931,
          "MSE_mean_pred": -0.9041298338057999,
          "SurVAL": 0.7562434455061977
        },
        {
          "task": "pet",
          "epoch": 5,
          "success_rate": -2.2819679331485334,
          "Loss": 1.433434323239329,
          "Off_Manifold_Norm_PolicyReference": -2.72556804976134,
          "MSE_mean_pred": 2.1012839571394495,
          "SurVAL": -2.781222179989345
        },
        {
          "task": "pet",
          "epoch": 10,
          "success_rate": -0.7435625849585109,
          "Loss": 1.2656792121011662,
          "Off_Manifold_Norm_PolicyReference": -0.3207826307941726,
          "MSE_mean_pred": 1.3927626622547995,
          "SurVAL": -0.1812442501318091
        },
        {
          "task": "pet",
          "epoch": 15,
          "success_rate": 0.02564008913650011,
          "Loss": 0.85952367391492,
          "Off_Manifold_Norm_PolicyReference": -0.0437170253302499,
          "MSE_mean_pred": 0.4649029557356382,
          "SurVAL": 0.15532179779130267
        },
        {
          "task": "pet",
          "epoch": 20,
          "success_rate": 0.2820409805015042,
          "Loss": 0.5599661576190469,
          "Off_Manifold_Norm_PolicyReference": 0.25391612825128085,
          "MSE_mean_pred": -0.2393535149468877,
          "SurVAL": 0.370434976410204
        },
        {
          "task": "pet",
          "epoch": 25,
          "success_rate": 0.7948427632315115,
          "Loss": -0.1593521006399896,
          "Off_Manifold_Norm_PolicyReference": 0.3364621873404286,
          "MSE_mean_pred": -0.4888426376291397,
          "SurVAL": 0.4556957386674432
        },
        {
          "task": "pet",
          "epoch": 30,
          "success_rate": 1.0512436545965151,
          "Loss": 0.09892780089141831,
          "Off_Manifold_Norm_PolicyReference": 0.4309808229636798,
          "MSE_mean_pred": -0.5644142831950641,
          "SurVAL": 0.481676450784167
        },
        {
          "task": "pet",
          "epoch": 35,
          "success_rate": -0.48716169359350725,
          "Loss": -0.7700148406434464,
          "Off_Manifold_Norm_PolicyReference": 0.395970370663316,
          "MSE_mean_pred": -0.5643722164637406,
          "SurVAL": 0.41466415747195956
        },
        {
          "task": "pet",
          "epoch": 40,
          "success_rate": 0.02564008913650011,
          "Loss": -0.901596320576976,
          "Off_Manifold_Norm_PolicyReference": 0.4692865407791313,
          "MSE_mean_pred": -0.8208963813314518,
          "SurVAL": 0.1505009780939062
        },
        {
          "task": "pet",
          "epoch": 45,
          "success_rate": 1.0512436545965151,
          "Loss": -1.226051030254888,
          "Off_Manifold_Norm_PolicyReference": 0.5711388837573403,
          "MSE_mean_pred": -0.6192668263273445,
          "SurVAL": 0.4672928316103559
        },
        {
          "task": "pet",
          "epoch": 50,
          "success_rate": 0.2820409805015042,
          "Loss": -1.1605168756505804,
          "Off_Manifold_Norm_PolicyReference": 0.6323127721305848,
          "MSE_mean_pred": -0.6618037152362588,
          "SurVAL": 0.4668794992918928
        }
      ],
      "metrics": {
        "Loss": {
          "nregret": 0.8444444444422808,
          "mmrv": 0.3011111111111111,
          "spearman": -0.7004659272862394,
          "balanced_loss": 0.8210748055441363,
          "regret_pp": 33.333333333333336,
          "selected_success_rate": 0.2111111111111111
        },
        "Off_Manifold_Norm_PolicyReference": {
          "nregret": 0.4547008546995994,
          "mmrv": 0.11222222222222222,
          "spearman": 0.5272078232721212,
          "balanced_loss": 0.33070752802972664,
          "regret_pp": 16.666666666666668,
          "selected_success_rate": 0.37777777777777777
        },
        "MSE_mean_pred": {
          "nregret": 0.8444444444422808,
          "mmrv": 0.3011111111111111,
          "spearman": -0.6923106401371552,
          "balanced_loss": 0.819715591019289,
          "regret_pp": 33.333333333333336,
          "selected_success_rate": 0.2111111111111111
        },
        "SurVAL": {
          "nregret": 0.0,
          "mmrv": 0.05333333333333334,
          "spearman": 0.9010583223617212,
          "balanced_loss": 0.06493757305367324,
          "regret_pp": 0.0,
          "selected_success_rate": 0.5444444444444444
        }
      },
      "checkpoint": "epoch",
      "steps": [
        5,
        10,
        15,
        20,
        25,
        30,
        35,
        40,
        45,
        50
      ],
      "legend": [
        "apple",
        "pan",
        "pet"
      ],
      "tasks": {
        "apple": {
          "name": "PnP Apple",
          "color": "#4285F4",
          "marker": "square"
        },
        "pan": {
          "name": "Put Pan on Stove",
          "color": "#EA4335",
          "marker": "triangle"
        },
        "pet": {
          "name": "Put Pet on Shelf",
          "color": "#34A853",
          "marker": "diamond"
        }
      },
      "axes": {
        "xlim": [
          -3.5,
          3.0
        ],
        "ylim": [
          -3.0,
          5.5
        ]
      },
      "method_keys": {
        "race": "SurVAL",
        "loss": "Loss",
        "mse": "MSE_mean_pred",
        "omn": "Off_Manifold_Norm_PolicyReference"
      },
      "task_metrics": {
        "apple": {
          "Loss": {
            "spearman": -0.8991204081865278,
            "mmrv": 0.3333333333333333,
            "nregret": 0.83333333333125
          },
          "Off_Manifold_Norm_PolicyReference": {
            "spearman": 0.7033935166085081,
            "mmrv": 0.07999999999999999,
            "nregret": 0.3333333333325
          },
          "MSE_mean_pred": {
            "spearman": -0.8991204081865278,
            "mmrv": 0.3333333333333333,
            "nregret": 0.83333333333125
          },
          "SurVAL": {
            "spearman": 0.9419356657192196,
            "mmrv": 0.023333333333333327,
            "nregret": 0.0
          }
        },
        "pan": {
          "Loss": {
            "spearman": -0.6150966989381312,
            "mmrv": 0.22666666666666666,
            "nregret": 0.6999999999979
          },
          "Off_Manifold_Norm_PolicyReference": {
            "spearman": 0.22988462485566521,
            "mmrv": 0.16000000000000003,
            "nregret": 0.7999999999976001
          },
          "MSE_mean_pred": {
            "spearman": -0.6150966989381312,
            "mmrv": 0.22666666666666666,
            "nregret": 0.6999999999979
          },
          "SurVAL": {
            "spearman": 0.8682353585412291,
            "mmrv": 0.09000000000000001,
            "nregret": 0.0
          }
        },
        "pet": {
          "Loss": {
            "spearman": -0.587180674734059,
            "mmrv": 0.3433333333333334,
            "nregret": 0.9999999999976924
          },
          "Off_Manifold_Norm_PolicyReference": {
            "spearman": 0.6483453283521902,
            "mmrv": 0.09666666666666668,
            "nregret": 0.23076923076869818
          },
          "MSE_mean_pred": {
            "spearman": -0.5627148132868065,
            "mmrv": 0.3433333333333334,
            "nregret": 0.9999999999976924
          },
          "SurVAL": {
            "spearman": 0.8930039428247147,
            "mmrv": 0.04666666666666667,
            "nregret": 0.0
          }
        }
      }
    }
  },
  "simulation": [
    {
      "name": "Held-out Demos",
      "methods": {
        "race": {
          "hit1": {
            "mean": 0.44,
            "lower": 0.23,
            "upper": 0.65
          },
          "hit3": {
            "mean": 0.75,
            "lower": 0.59,
            "upper": 0.9
          },
          "hit5": {
            "mean": 0.95,
            "lower": 0.88,
            "upper": 1.0
          },
          "spearman": {
            "mean": 0.83,
            "lower": 0.72,
            "upper": 0.92
          },
          "mmrv": {
            "mean": 0.03,
            "lower": 0.012,
            "upper": 0.053
          },
          "nregret": {
            "mean": 0.082,
            "lower": 0.035,
            "upper": 0.142
          }
        },
        "loss": {
          "hit1": {
            "mean": 0.31,
            "lower": 0.12,
            "upper": 0.53
          },
          "hit3": {
            "mean": 0.59,
            "lower": 0.33,
            "upper": 0.82
          },
          "hit5": {
            "mean": 0.75,
            "lower": 0.49,
            "upper": 0.97
          },
          "spearman": {
            "mean": 0.52,
            "lower": 0.01,
            "upper": 0.93
          },
          "mmrv": {
            "mean": 0.042,
            "lower": 0.011,
            "upper": 0.088
          },
          "nregret": {
            "mean": 0.238,
            "lower": 0.03,
            "upper": 0.504
          }
        },
        "mse": {
          "hit1": {
            "mean": 0.0,
            "lower": 0.0,
            "upper": 0.0
          },
          "hit3": {
            "mean": 0.08,
            "lower": 0.0,
            "upper": 0.17
          },
          "hit5": {
            "mean": 0.2,
            "lower": 0.04,
            "upper": 0.4
          },
          "spearman": {
            "mean": -0.22,
            "lower": -0.62,
            "upper": 0.2
          },
          "mmrv": {
            "mean": 0.122,
            "lower": 0.068,
            "upper": 0.18
          },
          "nregret": {
            "mean": 0.495,
            "lower": 0.325,
            "upper": 0.66
          }
        },
        "omn": {
          "hit1": {
            "mean": 0.25,
            "lower": 0.11,
            "upper": 0.4
          },
          "hit3": {
            "mean": 0.54,
            "lower": 0.33,
            "upper": 0.74
          },
          "hit5": {
            "mean": 0.72,
            "lower": 0.46,
            "upper": 0.94
          },
          "spearman": {
            "mean": 0.6,
            "lower": 0.22,
            "upper": 0.87
          },
          "mmrv": {
            "mean": 0.042,
            "lower": 0.022,
            "upper": 0.067
          },
          "nregret": {
            "mean": 0.146,
            "lower": 0.052,
            "upper": 0.288
          }
        }
      }
    },
    {
      "name": "Rollouts",
      "methods": {
        "race": {
          "hit1": {
            "mean": 0.27,
            "lower": 0.12,
            "upper": 0.45
          },
          "hit3": {
            "mean": 0.62,
            "lower": 0.38,
            "upper": 0.83
          },
          "hit5": {
            "mean": 0.9,
            "lower": 0.77,
            "upper": 1.0
          },
          "spearman": {
            "mean": 0.79,
            "lower": 0.69,
            "upper": 0.88
          },
          "mmrv": {
            "mean": 0.034,
            "lower": 0.021,
            "upper": 0.047
          },
          "nregret": {
            "mean": 0.113,
            "lower": 0.049,
            "upper": 0.186
          }
        },
        "loss": {
          "hit1": {
            "mean": 0.03,
            "lower": 0.0,
            "upper": 0.1
          },
          "hit3": {
            "mean": 0.12,
            "lower": 0.02,
            "upper": 0.27
          },
          "hit5": {
            "mean": 0.15,
            "lower": 0.03,
            "upper": 0.3
          },
          "spearman": {
            "mean": -0.1,
            "lower": -0.45,
            "upper": 0.25
          },
          "mmrv": {
            "mean": 0.136,
            "lower": 0.086,
            "upper": 0.19
          },
          "nregret": {
            "mean": 0.402,
            "lower": 0.298,
            "upper": 0.514
          }
        },
        "mse": {
          "hit1": {
            "mean": 0.08,
            "lower": 0.02,
            "upper": 0.18
          },
          "hit3": {
            "mean": 0.22,
            "lower": 0.05,
            "upper": 0.4
          },
          "hit5": {
            "mean": 0.3,
            "lower": 0.12,
            "upper": 0.5
          },
          "spearman": {
            "mean": -0.06,
            "lower": -0.27,
            "upper": 0.16
          },
          "mmrv": {
            "mean": 0.129,
            "lower": 0.084,
            "upper": 0.177
          },
          "nregret": {
            "mean": 0.48,
            "lower": 0.343,
            "upper": 0.633
          }
        },
        "omn": {
          "hit1": {
            "mean": 0.17,
            "lower": 0.07,
            "upper": 0.27
          },
          "hit3": {
            "mean": 0.37,
            "lower": 0.23,
            "upper": 0.5
          },
          "hit5": {
            "mean": 0.55,
            "lower": 0.42,
            "upper": 0.68
          },
          "spearman": {
            "mean": 0.53,
            "lower": 0.39,
            "upper": 0.66
          },
          "mmrv": {
            "mean": 0.068,
            "lower": 0.047,
            "upper": 0.094
          },
          "nregret": {
            "mean": 0.164,
            "lower": 0.123,
            "upper": 0.207
          }
        }
      }
    },
    {
      "name": "Noised Rollouts",
      "methods": {
        "race": {
          "hit1": {
            "mean": 0.32,
            "lower": 0.19,
            "upper": 0.49
          },
          "hit3": {
            "mean": 0.69,
            "lower": 0.52,
            "upper": 0.85
          },
          "hit5": {
            "mean": 0.91,
            "lower": 0.81,
            "upper": 0.99
          },
          "spearman": {
            "mean": 0.78,
            "lower": 0.68,
            "upper": 0.86
          },
          "mmrv": {
            "mean": 0.035,
            "lower": 0.021,
            "upper": 0.053
          },
          "nregret": {
            "mean": 0.081,
            "lower": 0.053,
            "upper": 0.116
          }
        },
        "loss": {
          "hit1": {
            "mean": 0.0,
            "lower": 0.0,
            "upper": 0.0
          },
          "hit3": {
            "mean": 0.04,
            "lower": 0.0,
            "upper": 0.11
          },
          "hit5": {
            "mean": 0.09,
            "lower": 0.01,
            "upper": 0.19
          },
          "spearman": {
            "mean": -0.47,
            "lower": -0.68,
            "upper": -0.15
          },
          "mmrv": {
            "mean": 0.154,
            "lower": 0.105,
            "upper": 0.202
          },
          "nregret": {
            "mean": 0.613,
            "lower": 0.462,
            "upper": 0.773
          }
        },
        "mse": {
          "hit1": {
            "mean": 0.09,
            "lower": 0.0,
            "upper": 0.21
          },
          "hit3": {
            "mean": 0.29,
            "lower": 0.11,
            "upper": 0.49
          },
          "hit5": {
            "mean": 0.39,
            "lower": 0.17,
            "upper": 0.61
          },
          "spearman": {
            "mean": 0.3,
            "lower": -0.05,
            "upper": 0.61
          },
          "mmrv": {
            "mean": 0.083,
            "lower": 0.041,
            "upper": 0.13
          },
          "nregret": {
            "mean": 0.288,
            "lower": 0.148,
            "upper": 0.443
          }
        },
        "omn": {
          "hit1": {
            "mean": 0.06,
            "lower": 0.01,
            "upper": 0.14
          },
          "hit3": {
            "mean": 0.31,
            "lower": 0.16,
            "upper": 0.47
          },
          "hit5": {
            "mean": 0.42,
            "lower": 0.25,
            "upper": 0.59
          },
          "spearman": {
            "mean": 0.3,
            "lower": 0.05,
            "upper": 0.52
          },
          "mmrv": {
            "mean": 0.085,
            "lower": 0.055,
            "upper": 0.122
          },
          "nregret": {
            "mean": 0.276,
            "lower": 0.185,
            "upper": 0.405
          }
        }
      }
    }
  ]
}
