ctt api upload 2026-07-02 workspace/runs/ctt_gated_residual_smoke_proxy/metrics.json
Browse files
workspace/runs/ctt_gated_residual_smoke_proxy/metrics.json
ADDED
|
@@ -0,0 +1,472 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"data_hash": "90b431047a1f35cec095cbafa7de5bc0126cb0d0f4db3ff31843dca369977957",
|
| 3 |
+
"k": 16,
|
| 4 |
+
"num_rows": 16,
|
| 5 |
+
"report_type": "ctt_proxy_eval",
|
| 6 |
+
"rows": [
|
| 7 |
+
{
|
| 8 |
+
"chart_id": "ms-003a2225c0892dcda683",
|
| 9 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 10 |
+
"negative_near_at_16_thr_0p40": 0.0,
|
| 11 |
+
"num_proposals": 8,
|
| 12 |
+
"pos_closer_than_neg_at_16": 1.0,
|
| 13 |
+
"pptc_at_16_thr_0p20": 0.0,
|
| 14 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 15 |
+
"proxy_support_distance_at_16": 0.2974270141105481,
|
| 16 |
+
"seed": "unknown",
|
| 17 |
+
"task_id": "PullCube-v1"
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"chart_id": "ms-0093b06390890d181104",
|
| 21 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 22 |
+
"negative_near_at_16_thr_0p40": 0.0,
|
| 23 |
+
"num_proposals": 16,
|
| 24 |
+
"pos_closer_than_neg_at_16": 1.0,
|
| 25 |
+
"pptc_at_16_thr_0p20": 0.0,
|
| 26 |
+
"pptc_at_16_thr_0p40": 0.0,
|
| 27 |
+
"proxy_support_distance_at_16": 0.6211311373590876,
|
| 28 |
+
"seed": "unknown",
|
| 29 |
+
"task_id": "PullCube-v1"
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"chart_id": "ms-016c5de39ed3c111889a",
|
| 33 |
+
"negative_near_at_16_thr_0p20": 0.25,
|
| 34 |
+
"negative_near_at_16_thr_0p40": 1.0,
|
| 35 |
+
"num_proposals": 4,
|
| 36 |
+
"pos_closer_than_neg_at_16": 1.0,
|
| 37 |
+
"pptc_at_16_thr_0p20": 1.0,
|
| 38 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 39 |
+
"proxy_support_distance_at_16": 0.1331443405882652,
|
| 40 |
+
"seed": "unknown",
|
| 41 |
+
"task_id": "StackCube-v1"
|
| 42 |
+
},
|
| 43 |
+
{
|
| 44 |
+
"chart_id": "ms-01e6eaaaf40875e0ef5a",
|
| 45 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 46 |
+
"negative_near_at_16_thr_0p40": 0.45454545454545453,
|
| 47 |
+
"num_proposals": 11,
|
| 48 |
+
"pos_closer_than_neg_at_16": 0.9090909090909091,
|
| 49 |
+
"pptc_at_16_thr_0p20": 1.0,
|
| 50 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 51 |
+
"proxy_support_distance_at_16": 0.12855269473830008,
|
| 52 |
+
"seed": "unknown",
|
| 53 |
+
"task_id": "StackCube-v1"
|
| 54 |
+
},
|
| 55 |
+
{
|
| 56 |
+
"chart_id": "ms-02311c7783bc67e188b4",
|
| 57 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 58 |
+
"negative_near_at_16_thr_0p40": 0.0,
|
| 59 |
+
"num_proposals": 13,
|
| 60 |
+
"pos_closer_than_neg_at_16": null,
|
| 61 |
+
"pptc_at_16_thr_0p20": 0.0,
|
| 62 |
+
"pptc_at_16_thr_0p40": 0.0,
|
| 63 |
+
"proxy_support_distance_at_16": 0.7497789035261392,
|
| 64 |
+
"seed": "unknown",
|
| 65 |
+
"task_id": "PullCube-v1"
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"chart_id": "ms-0285f5815c210ac00bec",
|
| 69 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 70 |
+
"negative_near_at_16_thr_0p40": 0.0,
|
| 71 |
+
"num_proposals": 7,
|
| 72 |
+
"pos_closer_than_neg_at_16": 1.0,
|
| 73 |
+
"pptc_at_16_thr_0p20": 1.0,
|
| 74 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 75 |
+
"proxy_support_distance_at_16": 0.19715026157240312,
|
| 76 |
+
"seed": "unknown",
|
| 77 |
+
"task_id": "PullCube-v1"
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"chart_id": "ms-02fd0338e80ef19ceec1",
|
| 81 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 82 |
+
"negative_near_at_16_thr_0p40": 0.14285714285714285,
|
| 83 |
+
"num_proposals": 14,
|
| 84 |
+
"pos_closer_than_neg_at_16": 0.0,
|
| 85 |
+
"pptc_at_16_thr_0p20": 0.0,
|
| 86 |
+
"pptc_at_16_thr_0p40": 0.0,
|
| 87 |
+
"proxy_support_distance_at_16": 0.7927007012351106,
|
| 88 |
+
"seed": "unknown",
|
| 89 |
+
"task_id": "PullCube-v1"
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"chart_id": "ms-0350672b45f92c437c22",
|
| 93 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 94 |
+
"negative_near_at_16_thr_0p40": 0.1875,
|
| 95 |
+
"num_proposals": 16,
|
| 96 |
+
"pos_closer_than_neg_at_16": 1.0,
|
| 97 |
+
"pptc_at_16_thr_0p20": 0.0,
|
| 98 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 99 |
+
"proxy_support_distance_at_16": 0.2793621804569552,
|
| 100 |
+
"seed": "unknown",
|
| 101 |
+
"task_id": "PushCube-v1"
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"chart_id": "ms-03df3e12e1d61aa9ebbf",
|
| 105 |
+
"negative_near_at_16_thr_0p20": 0.6666666666666666,
|
| 106 |
+
"negative_near_at_16_thr_0p40": 0.8333333333333334,
|
| 107 |
+
"num_proposals": 6,
|
| 108 |
+
"pos_closer_than_neg_at_16": 0.8333333333333334,
|
| 109 |
+
"pptc_at_16_thr_0p20": 1.0,
|
| 110 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 111 |
+
"proxy_support_distance_at_16": 0.16971538889823268,
|
| 112 |
+
"seed": "unknown",
|
| 113 |
+
"task_id": "StackCube-v1"
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"chart_id": "ms-04eec0abab5f55620e4e",
|
| 117 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 118 |
+
"negative_near_at_16_thr_0p40": 0.875,
|
| 119 |
+
"num_proposals": 8,
|
| 120 |
+
"pos_closer_than_neg_at_16": 0.875,
|
| 121 |
+
"pptc_at_16_thr_0p20": 1.0,
|
| 122 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 123 |
+
"proxy_support_distance_at_16": 0.1793190780000492,
|
| 124 |
+
"seed": "unknown",
|
| 125 |
+
"task_id": "StackCube-v1"
|
| 126 |
+
},
|
| 127 |
+
{
|
| 128 |
+
"chart_id": "ms-0541a144afe57b5a7ae6",
|
| 129 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 130 |
+
"negative_near_at_16_thr_0p40": 1.0,
|
| 131 |
+
"num_proposals": 4,
|
| 132 |
+
"pos_closer_than_neg_at_16": 1.0,
|
| 133 |
+
"pptc_at_16_thr_0p20": 1.0,
|
| 134 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 135 |
+
"proxy_support_distance_at_16": 0.1869813893501718,
|
| 136 |
+
"seed": "unknown",
|
| 137 |
+
"task_id": "PickCube-v1"
|
| 138 |
+
},
|
| 139 |
+
{
|
| 140 |
+
"chart_id": "ms-0547201761e8e1873482",
|
| 141 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 142 |
+
"negative_near_at_16_thr_0p40": 0.2,
|
| 143 |
+
"num_proposals": 5,
|
| 144 |
+
"pos_closer_than_neg_at_16": 1.0,
|
| 145 |
+
"pptc_at_16_thr_0p20": 0.0,
|
| 146 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 147 |
+
"proxy_support_distance_at_16": 0.26221705020966296,
|
| 148 |
+
"seed": "unknown",
|
| 149 |
+
"task_id": "StackCube-v1"
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"chart_id": "ms-059d7962a5f998edf668",
|
| 153 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 154 |
+
"negative_near_at_16_thr_0p40": 0.8333333333333334,
|
| 155 |
+
"num_proposals": 12,
|
| 156 |
+
"pos_closer_than_neg_at_16": 0.8333333333333334,
|
| 157 |
+
"pptc_at_16_thr_0p20": 1.0,
|
| 158 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 159 |
+
"proxy_support_distance_at_16": 0.16640479340742947,
|
| 160 |
+
"seed": "unknown",
|
| 161 |
+
"task_id": "LiftPegUpright-v1"
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"chart_id": "ms-062addb9da32a0c2ea97",
|
| 165 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 166 |
+
"negative_near_at_16_thr_0p40": 0.16666666666666666,
|
| 167 |
+
"num_proposals": 6,
|
| 168 |
+
"pos_closer_than_neg_at_16": 0.8333333333333334,
|
| 169 |
+
"pptc_at_16_thr_0p20": 1.0,
|
| 170 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 171 |
+
"proxy_support_distance_at_16": 0.19839601892529432,
|
| 172 |
+
"seed": "unknown",
|
| 173 |
+
"task_id": "LiftPegUpright-v1"
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"chart_id": "ms-0660d5f65df5fc4b6132",
|
| 177 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 178 |
+
"negative_near_at_16_thr_0p40": 1.0,
|
| 179 |
+
"num_proposals": 5,
|
| 180 |
+
"pos_closer_than_neg_at_16": 0.0,
|
| 181 |
+
"pptc_at_16_thr_0p20": 0.0,
|
| 182 |
+
"pptc_at_16_thr_0p40": 1.0,
|
| 183 |
+
"proxy_support_distance_at_16": 0.2560477130376794,
|
| 184 |
+
"seed": "unknown",
|
| 185 |
+
"task_id": "LiftPegUpright-v1"
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"chart_id": "ms-067cac2e4dbeddb12239",
|
| 189 |
+
"negative_near_at_16_thr_0p20": 0.0,
|
| 190 |
+
"negative_near_at_16_thr_0p40": 0.0,
|
| 191 |
+
"num_proposals": 5,
|
| 192 |
+
"pos_closer_than_neg_at_16": 0.0,
|
| 193 |
+
"pptc_at_16_thr_0p20": 0.0,
|
| 194 |
+
"pptc_at_16_thr_0p40": 0.0,
|
| 195 |
+
"proxy_support_distance_at_16": 0.9123656522393015,
|
| 196 |
+
"seed": "unknown",
|
| 197 |
+
"task_id": "PullCube-v1"
|
| 198 |
+
}
|
| 199 |
+
],
|
| 200 |
+
"split_hash": "867a94a1c67686c082765671a7bf8d74a9d633085e1b2619c83197c60bf58372",
|
| 201 |
+
"summary": {
|
| 202 |
+
"negative_near_at_16_thr_0p20": {
|
| 203 |
+
"macro_by_seed": {
|
| 204 |
+
"ci": {
|
| 205 |
+
"confidence": 0.95,
|
| 206 |
+
"high": 0.057291666666666664,
|
| 207 |
+
"low": 0.057291666666666664,
|
| 208 |
+
"mean": 0.057291666666666664,
|
| 209 |
+
"n": 1,
|
| 210 |
+
"num_samples": 200
|
| 211 |
+
},
|
| 212 |
+
"mean": 0.057291666666666664,
|
| 213 |
+
"per_seed": {
|
| 214 |
+
"unknown": 0.057291666666666664
|
| 215 |
+
}
|
| 216 |
+
},
|
| 217 |
+
"macro_by_task": {
|
| 218 |
+
"ci": {
|
| 219 |
+
"confidence": 0.95,
|
| 220 |
+
"high": 0.10999999999999999,
|
| 221 |
+
"low": 0.0,
|
| 222 |
+
"mean": 0.03666666666666667,
|
| 223 |
+
"n": 5,
|
| 224 |
+
"num_samples": 200
|
| 225 |
+
},
|
| 226 |
+
"mean": 0.03666666666666667,
|
| 227 |
+
"per_task": {
|
| 228 |
+
"LiftPegUpright-v1": 0.0,
|
| 229 |
+
"PickCube-v1": 0.0,
|
| 230 |
+
"PullCube-v1": 0.0,
|
| 231 |
+
"PushCube-v1": 0.0,
|
| 232 |
+
"StackCube-v1": 0.18333333333333332
|
| 233 |
+
}
|
| 234 |
+
},
|
| 235 |
+
"metric": "negative_near_at_16_thr_0p20",
|
| 236 |
+
"micro": {
|
| 237 |
+
"confidence": 0.95,
|
| 238 |
+
"high": 0.16666666666666666,
|
| 239 |
+
"low": 0.0,
|
| 240 |
+
"mean": 0.057291666666666664,
|
| 241 |
+
"n": 16,
|
| 242 |
+
"num_samples": 200
|
| 243 |
+
},
|
| 244 |
+
"n": 16
|
| 245 |
+
},
|
| 246 |
+
"negative_near_at_16_thr_0p40": {
|
| 247 |
+
"macro_by_seed": {
|
| 248 |
+
"ci": {
|
| 249 |
+
"confidence": 0.95,
|
| 250 |
+
"high": 0.4183272456709957,
|
| 251 |
+
"low": 0.4183272456709957,
|
| 252 |
+
"mean": 0.4183272456709957,
|
| 253 |
+
"n": 1,
|
| 254 |
+
"num_samples": 200
|
| 255 |
+
},
|
| 256 |
+
"mean": 0.4183272456709957,
|
| 257 |
+
"per_seed": {
|
| 258 |
+
"unknown": 0.4183272456709957
|
| 259 |
+
}
|
| 260 |
+
},
|
| 261 |
+
"macro_by_task": {
|
| 262 |
+
"ci": {
|
| 263 |
+
"confidence": 0.95,
|
| 264 |
+
"high": 0.7720151515151515,
|
| 265 |
+
"low": 0.2178571428571429,
|
| 266 |
+
"mean": 0.5101103896103896,
|
| 267 |
+
"n": 5,
|
| 268 |
+
"num_samples": 200
|
| 269 |
+
},
|
| 270 |
+
"mean": 0.5101103896103896,
|
| 271 |
+
"per_task": {
|
| 272 |
+
"LiftPegUpright-v1": 0.6666666666666666,
|
| 273 |
+
"PickCube-v1": 1.0,
|
| 274 |
+
"PullCube-v1": 0.023809523809523808,
|
| 275 |
+
"PushCube-v1": 0.1875,
|
| 276 |
+
"StackCube-v1": 0.6725757575757576
|
| 277 |
+
}
|
| 278 |
+
},
|
| 279 |
+
"metric": "negative_near_at_16_thr_0p40",
|
| 280 |
+
"micro": {
|
| 281 |
+
"confidence": 0.95,
|
| 282 |
+
"high": 0.5935639880952381,
|
| 283 |
+
"low": 0.2338541666666667,
|
| 284 |
+
"mean": 0.4183272456709957,
|
| 285 |
+
"n": 16,
|
| 286 |
+
"num_samples": 200
|
| 287 |
+
},
|
| 288 |
+
"n": 16
|
| 289 |
+
},
|
| 290 |
+
"pos_closer_than_neg_at_16": {
|
| 291 |
+
"macro_by_seed": {
|
| 292 |
+
"ci": {
|
| 293 |
+
"confidence": 0.95,
|
| 294 |
+
"high": 0.7522727272727273,
|
| 295 |
+
"low": 0.7522727272727273,
|
| 296 |
+
"mean": 0.7522727272727273,
|
| 297 |
+
"n": 1,
|
| 298 |
+
"num_samples": 200
|
| 299 |
+
},
|
| 300 |
+
"mean": 0.7522727272727273,
|
| 301 |
+
"per_seed": {
|
| 302 |
+
"unknown": 0.7522727272727273
|
| 303 |
+
}
|
| 304 |
+
},
|
| 305 |
+
"macro_by_task": {
|
| 306 |
+
"ci": {
|
| 307 |
+
"confidence": 0.95,
|
| 308 |
+
"high": 0.9693939393939394,
|
| 309 |
+
"low": 0.6380303030303031,
|
| 310 |
+
"mean": 0.8158080808080808,
|
| 311 |
+
"n": 5,
|
| 312 |
+
"num_samples": 200
|
| 313 |
+
},
|
| 314 |
+
"mean": 0.8158080808080808,
|
| 315 |
+
"per_task": {
|
| 316 |
+
"LiftPegUpright-v1": 0.5555555555555556,
|
| 317 |
+
"PickCube-v1": 1.0,
|
| 318 |
+
"PullCube-v1": 0.6,
|
| 319 |
+
"PushCube-v1": 1.0,
|
| 320 |
+
"StackCube-v1": 0.9234848484848485
|
| 321 |
+
}
|
| 322 |
+
},
|
| 323 |
+
"metric": "pos_closer_than_neg_at_16",
|
| 324 |
+
"micro": {
|
| 325 |
+
"confidence": 0.95,
|
| 326 |
+
"high": 0.9277777777777777,
|
| 327 |
+
"low": 0.5550505050505051,
|
| 328 |
+
"mean": 0.7522727272727273,
|
| 329 |
+
"n": 15,
|
| 330 |
+
"num_samples": 200
|
| 331 |
+
},
|
| 332 |
+
"n": 15
|
| 333 |
+
},
|
| 334 |
+
"pptc_at_16_thr_0p20": {
|
| 335 |
+
"macro_by_seed": {
|
| 336 |
+
"ci": {
|
| 337 |
+
"confidence": 0.95,
|
| 338 |
+
"high": 0.5,
|
| 339 |
+
"low": 0.5,
|
| 340 |
+
"mean": 0.5,
|
| 341 |
+
"n": 1,
|
| 342 |
+
"num_samples": 200
|
| 343 |
+
},
|
| 344 |
+
"mean": 0.5,
|
| 345 |
+
"per_seed": {
|
| 346 |
+
"unknown": 0.5
|
| 347 |
+
}
|
| 348 |
+
},
|
| 349 |
+
"macro_by_task": {
|
| 350 |
+
"ci": {
|
| 351 |
+
"confidence": 0.95,
|
| 352 |
+
"high": 0.8133333333333332,
|
| 353 |
+
"low": 0.19999999999999998,
|
| 354 |
+
"mean": 0.5266666666666666,
|
| 355 |
+
"n": 5,
|
| 356 |
+
"num_samples": 200
|
| 357 |
+
},
|
| 358 |
+
"mean": 0.5266666666666666,
|
| 359 |
+
"per_task": {
|
| 360 |
+
"LiftPegUpright-v1": 0.6666666666666666,
|
| 361 |
+
"PickCube-v1": 1.0,
|
| 362 |
+
"PullCube-v1": 0.16666666666666666,
|
| 363 |
+
"PushCube-v1": 0.0,
|
| 364 |
+
"StackCube-v1": 0.8
|
| 365 |
+
}
|
| 366 |
+
},
|
| 367 |
+
"metric": "pptc_at_16_thr_0p20",
|
| 368 |
+
"micro": {
|
| 369 |
+
"confidence": 0.95,
|
| 370 |
+
"high": 0.75,
|
| 371 |
+
"low": 0.3125,
|
| 372 |
+
"mean": 0.5,
|
| 373 |
+
"n": 16,
|
| 374 |
+
"num_samples": 200
|
| 375 |
+
},
|
| 376 |
+
"n": 16
|
| 377 |
+
},
|
| 378 |
+
"pptc_at_16_thr_0p40": {
|
| 379 |
+
"macro_by_seed": {
|
| 380 |
+
"ci": {
|
| 381 |
+
"confidence": 0.95,
|
| 382 |
+
"high": 0.75,
|
| 383 |
+
"low": 0.75,
|
| 384 |
+
"mean": 0.75,
|
| 385 |
+
"n": 1,
|
| 386 |
+
"num_samples": 200
|
| 387 |
+
},
|
| 388 |
+
"mean": 0.75,
|
| 389 |
+
"per_seed": {
|
| 390 |
+
"unknown": 0.75
|
| 391 |
+
}
|
| 392 |
+
},
|
| 393 |
+
"macro_by_task": {
|
| 394 |
+
"ci": {
|
| 395 |
+
"confidence": 0.95,
|
| 396 |
+
"high": 1.0,
|
| 397 |
+
"low": 0.6,
|
| 398 |
+
"mean": 0.8666666666666668,
|
| 399 |
+
"n": 5,
|
| 400 |
+
"num_samples": 200
|
| 401 |
+
},
|
| 402 |
+
"mean": 0.8666666666666668,
|
| 403 |
+
"per_task": {
|
| 404 |
+
"LiftPegUpright-v1": 1.0,
|
| 405 |
+
"PickCube-v1": 1.0,
|
| 406 |
+
"PullCube-v1": 0.3333333333333333,
|
| 407 |
+
"PushCube-v1": 1.0,
|
| 408 |
+
"StackCube-v1": 1.0
|
| 409 |
+
}
|
| 410 |
+
},
|
| 411 |
+
"metric": "pptc_at_16_thr_0p40",
|
| 412 |
+
"micro": {
|
| 413 |
+
"confidence": 0.95,
|
| 414 |
+
"high": 0.9375,
|
| 415 |
+
"low": 0.5625,
|
| 416 |
+
"mean": 0.75,
|
| 417 |
+
"n": 16,
|
| 418 |
+
"num_samples": 200
|
| 419 |
+
},
|
| 420 |
+
"n": 16
|
| 421 |
+
},
|
| 422 |
+
"proxy_support_distance_at_16": {
|
| 423 |
+
"macro_by_seed": {
|
| 424 |
+
"ci": {
|
| 425 |
+
"confidence": 0.95,
|
| 426 |
+
"high": 0.3456683948534143,
|
| 427 |
+
"low": 0.3456683948534143,
|
| 428 |
+
"mean": 0.3456683948534143,
|
| 429 |
+
"n": 1,
|
| 430 |
+
"num_samples": 200
|
| 431 |
+
},
|
| 432 |
+
"mean": 0.3456683948534143,
|
| 433 |
+
"per_seed": {
|
| 434 |
+
"unknown": 0.3456683948534143
|
| 435 |
+
}
|
| 436 |
+
},
|
| 437 |
+
"macro_by_task": {
|
| 438 |
+
"ci": {
|
| 439 |
+
"confidence": 0.95,
|
| 440 |
+
"high": 0.44784574519303044,
|
| 441 |
+
"low": 0.18354000585353578,
|
| 442 |
+
"mean": 0.28859501341825233,
|
| 443 |
+
"n": 5,
|
| 444 |
+
"num_samples": 200
|
| 445 |
+
},
|
| 446 |
+
"mean": 0.28859501341825233,
|
| 447 |
+
"per_task": {
|
| 448 |
+
"LiftPegUpright-v1": 0.20694950845680107,
|
| 449 |
+
"PickCube-v1": 0.1869813893501718,
|
| 450 |
+
"PullCube-v1": 0.5950922783404317,
|
| 451 |
+
"PushCube-v1": 0.2793621804569552,
|
| 452 |
+
"StackCube-v1": 0.17458971048690203
|
| 453 |
+
}
|
| 454 |
+
},
|
| 455 |
+
"metric": "proxy_support_distance_at_16",
|
| 456 |
+
"micro": {
|
| 457 |
+
"confidence": 0.95,
|
| 458 |
+
"high": 0.4712772837277821,
|
| 459 |
+
"low": 0.23957205766924433,
|
| 460 |
+
"mean": 0.3456683948534143,
|
| 461 |
+
"n": 16,
|
| 462 |
+
"num_samples": 200
|
| 463 |
+
},
|
| 464 |
+
"n": 16
|
| 465 |
+
}
|
| 466 |
+
},
|
| 467 |
+
"target_split_hash": "867a94a1c67686c082765671a7bf8d74a9d633085e1b2619c83197c60bf58372",
|
| 468 |
+
"thresholds": [
|
| 469 |
+
0.2,
|
| 470 |
+
0.4
|
| 471 |
+
]
|
| 472 |
+
}
|