Scheduled Commit
Browse files- ablation_runs/eval_all.json +952 -0
- ablation_runs/eval_d.json +261 -0
- ablation_runs/findings.html +43 -0
- ablation_runs/track_force_cascade.html +160 -32
ablation_runs/eval_all.json
ADDED
|
@@ -0,0 +1,952 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"full": {
|
| 3 |
+
"checkpoint": "checkpoints/ablate_full_3k/checkpoint-3000",
|
| 4 |
+
"use_track": true,
|
| 5 |
+
"use_force": true,
|
| 6 |
+
"use_deform_tactile": false,
|
| 7 |
+
"forward_tau_losses": {
|
| 8 |
+
"0.8": {
|
| 9 |
+
"action_loss": [
|
| 10 |
+
0.13885498046875,
|
| 11 |
+
0.021893762307867347
|
| 12 |
+
],
|
| 13 |
+
"dynamics_loss": [
|
| 14 |
+
0.214599609375,
|
| 15 |
+
0.028883325232585192
|
| 16 |
+
],
|
| 17 |
+
"track_flow_loss": [
|
| 18 |
+
0.01959228515625,
|
| 19 |
+
0.005815500664137739
|
| 20 |
+
],
|
| 21 |
+
"force_loss": [
|
| 22 |
+
0.6427408854166666,
|
| 23 |
+
0.07174448889029805
|
| 24 |
+
]
|
| 25 |
+
},
|
| 26 |
+
"0.6": {
|
| 27 |
+
"action_loss": [
|
| 28 |
+
0.20402018229166666,
|
| 29 |
+
0.027474107787100906
|
| 30 |
+
],
|
| 31 |
+
"dynamics_loss": [
|
| 32 |
+
0.2801106770833333,
|
| 33 |
+
0.02749254221890766
|
| 34 |
+
],
|
| 35 |
+
"track_flow_loss": [
|
| 36 |
+
0.025583902994791668,
|
| 37 |
+
0.009258714454018649
|
| 38 |
+
],
|
| 39 |
+
"force_loss": [
|
| 40 |
+
0.6427408854166666,
|
| 41 |
+
0.07174448889029805
|
| 42 |
+
]
|
| 43 |
+
},
|
| 44 |
+
"0.4": {
|
| 45 |
+
"action_loss": [
|
| 46 |
+
0.3986002604166667,
|
| 47 |
+
0.0390743661447396
|
| 48 |
+
],
|
| 49 |
+
"dynamics_loss": [
|
| 50 |
+
0.4138997395833333,
|
| 51 |
+
0.025458871788911118
|
| 52 |
+
],
|
| 53 |
+
"track_flow_loss": [
|
| 54 |
+
0.051778157552083336,
|
| 55 |
+
0.017947368985419505
|
| 56 |
+
],
|
| 57 |
+
"force_loss": [
|
| 58 |
+
0.6427408854166666,
|
| 59 |
+
0.07174448889029805
|
| 60 |
+
]
|
| 61 |
+
},
|
| 62 |
+
"0.2": {
|
| 63 |
+
"action_loss": [
|
| 64 |
+
0.7521158854166666,
|
| 65 |
+
0.048475014437045466
|
| 66 |
+
],
|
| 67 |
+
"dynamics_loss": [
|
| 68 |
+
0.7236328125,
|
| 69 |
+
0.020585171554378095
|
| 70 |
+
],
|
| 71 |
+
"track_flow_loss": [
|
| 72 |
+
0.270751953125,
|
| 73 |
+
0.02853870589963502
|
| 74 |
+
],
|
| 75 |
+
"force_loss": [
|
| 76 |
+
0.6427408854166666,
|
| 77 |
+
0.07174448889029805
|
| 78 |
+
]
|
| 79 |
+
},
|
| 80 |
+
"0.05": {
|
| 81 |
+
"action_loss": [
|
| 82 |
+
1.0758463541666667,
|
| 83 |
+
0.05334471104376999
|
| 84 |
+
],
|
| 85 |
+
"dynamics_loss": [
|
| 86 |
+
1.1305338541666667,
|
| 87 |
+
0.01504103522415094
|
| 88 |
+
],
|
| 89 |
+
"track_flow_loss": [
|
| 90 |
+
0.912109375,
|
| 91 |
+
0.026373214974647258
|
| 92 |
+
],
|
| 93 |
+
"force_loss": [
|
| 94 |
+
0.6427408854166666,
|
| 95 |
+
0.07174448889029805
|
| 96 |
+
]
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"openloop_action_mse": {
|
| 100 |
+
"cascade": [
|
| 101 |
+
0.20543859526515007,
|
| 102 |
+
0.10553651615425272,
|
| 103 |
+
16
|
| 104 |
+
],
|
| 105 |
+
"coarse_only": [
|
| 106 |
+
0.3859300408512354,
|
| 107 |
+
0.09367721956997326,
|
| 108 |
+
16
|
| 109 |
+
]
|
| 110 |
+
},
|
| 111 |
+
"openloop_action_mse_raw": {
|
| 112 |
+
"cascade": [
|
| 113 |
+
0.2599409520626068,
|
| 114 |
+
0.16157826781272888,
|
| 115 |
+
0.15067900717258453,
|
| 116 |
+
0.15890291333198547,
|
| 117 |
+
0.13666832447052002,
|
| 118 |
+
0.14541499316692352,
|
| 119 |
+
0.1762809008359909,
|
| 120 |
+
0.1548680067062378,
|
| 121 |
+
0.14876900613307953,
|
| 122 |
+
0.2613345682621002,
|
| 123 |
+
0.14742904901504517,
|
| 124 |
+
0.15507377684116364,
|
| 125 |
+
0.17896708846092224,
|
| 126 |
+
0.19943304359912872,
|
| 127 |
+
0.27371665835380554,
|
| 128 |
+
0.5779609680175781
|
| 129 |
+
],
|
| 130 |
+
"coarse_only": [
|
| 131 |
+
0.44467893242836,
|
| 132 |
+
0.35982707142829895,
|
| 133 |
+
0.32907775044441223,
|
| 134 |
+
0.36753952503204346,
|
| 135 |
+
0.33504316210746765,
|
| 136 |
+
0.3160090148448944,
|
| 137 |
+
0.34967127442359924,
|
| 138 |
+
0.3433721661567688,
|
| 139 |
+
0.2978774905204773,
|
| 140 |
+
0.3976256549358368,
|
| 141 |
+
0.34054020047187805,
|
| 142 |
+
0.3271501660346985,
|
| 143 |
+
0.3467199206352234,
|
| 144 |
+
0.42348822951316833,
|
| 145 |
+
0.5095632076263428,
|
| 146 |
+
0.6866968870162964
|
| 147 |
+
]
|
| 148 |
+
},
|
| 149 |
+
"openloop_action_mse_split": {
|
| 150 |
+
"cascade_arm": 0.29095884039998055,
|
| 151 |
+
"cascade_hand": 0.1704530455172062,
|
| 152 |
+
"coarse_arm": 0.42187667451798916,
|
| 153 |
+
"coarse_hand": 0.3712246026843786
|
| 154 |
+
},
|
| 155 |
+
"stage1_attention": {
|
| 156 |
+
"0": {
|
| 157 |
+
"action": {
|
| 158 |
+
"cond_obs": 0.028233155608177185,
|
| 159 |
+
"obs": 0.10685919225215912,
|
| 160 |
+
"action": 0.06159074231982231,
|
| 161 |
+
"state": 0.001643829746171832,
|
| 162 |
+
"track_past": 0.4548226296901703,
|
| 163 |
+
"track_future": 0.3468504846096039
|
| 164 |
+
},
|
| 165 |
+
"obs": {
|
| 166 |
+
"cond_obs": 0.053328514099121094,
|
| 167 |
+
"obs": 0.5826746821403503,
|
| 168 |
+
"action": 0.009129117242991924,
|
| 169 |
+
"state": 0.0005130452918820083,
|
| 170 |
+
"track_past": 0.1784334033727646,
|
| 171 |
+
"track_future": 0.1759212762117386
|
| 172 |
+
}
|
| 173 |
+
},
|
| 174 |
+
"14": {
|
| 175 |
+
"action": {
|
| 176 |
+
"cond_obs": 0.04729510098695755,
|
| 177 |
+
"obs": 0.08689284324645996,
|
| 178 |
+
"action": 0.20308324694633484,
|
| 179 |
+
"state": 0.0032673801761120558,
|
| 180 |
+
"track_past": 0.6379822492599487,
|
| 181 |
+
"track_future": 0.021479224786162376
|
| 182 |
+
},
|
| 183 |
+
"obs": {
|
| 184 |
+
"cond_obs": 0.16180963814258575,
|
| 185 |
+
"obs": 0.6352365016937256,
|
| 186 |
+
"action": 0.013975701294839382,
|
| 187 |
+
"state": 0.006121945567429066,
|
| 188 |
+
"track_past": 0.167484849691391,
|
| 189 |
+
"track_future": 0.015371335670351982
|
| 190 |
+
}
|
| 191 |
+
},
|
| 192 |
+
"29": {
|
| 193 |
+
"action": {
|
| 194 |
+
"cond_obs": 0.05335542559623718,
|
| 195 |
+
"obs": 0.09940896928310394,
|
| 196 |
+
"action": 0.18759986758232117,
|
| 197 |
+
"state": 0.0025236550718545914,
|
| 198 |
+
"track_past": 0.5753425359725952,
|
| 199 |
+
"track_future": 0.08176963031291962
|
| 200 |
+
},
|
| 201 |
+
"obs": {
|
| 202 |
+
"cond_obs": 0.09543927758932114,
|
| 203 |
+
"obs": 0.7705966234207153,
|
| 204 |
+
"action": 0.01101355254650116,
|
| 205 |
+
"state": 0.0004360256134532392,
|
| 206 |
+
"track_past": 0.07644421607255936,
|
| 207 |
+
"track_future": 0.046070314943790436
|
| 208 |
+
}
|
| 209 |
+
}
|
| 210 |
+
},
|
| 211 |
+
"stage2_attention": [
|
| 212 |
+
{
|
| 213 |
+
"action": 0.033935546875,
|
| 214 |
+
"f6": 0.0230712890625,
|
| 215 |
+
"vq": 0.024169921875,
|
| 216 |
+
"memory": 0.91796875
|
| 217 |
+
},
|
| 218 |
+
{
|
| 219 |
+
"action": 0.0322265625,
|
| 220 |
+
"f6": 0.0208740234375,
|
| 221 |
+
"vq": 0.023681640625,
|
| 222 |
+
"memory": 0.921875
|
| 223 |
+
},
|
| 224 |
+
{
|
| 225 |
+
"action": 0.0274658203125,
|
| 226 |
+
"f6": 0.01708984375,
|
| 227 |
+
"vq": 0.0203857421875,
|
| 228 |
+
"memory": 0.93359375
|
| 229 |
+
},
|
| 230 |
+
{
|
| 231 |
+
"action": 0.0250244140625,
|
| 232 |
+
"f6": 0.01513671875,
|
| 233 |
+
"vq": 0.0185546875,
|
| 234 |
+
"memory": 0.94140625
|
| 235 |
+
},
|
| 236 |
+
{
|
| 237 |
+
"action": 0.024169921875,
|
| 238 |
+
"f6": 0.01416015625,
|
| 239 |
+
"vq": 0.017333984375,
|
| 240 |
+
"memory": 0.9453125
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"action": 0.0228271484375,
|
| 244 |
+
"f6": 0.0130615234375,
|
| 245 |
+
"vq": 0.0157470703125,
|
| 246 |
+
"memory": 0.94921875
|
| 247 |
+
}
|
| 248 |
+
],
|
| 249 |
+
"tactile_sensitivity": {
|
| 250 |
+
"mean_abs_delta_vs_masked": 0.014084354043006897,
|
| 251 |
+
"max_abs_delta_vs_masked": 0.06640625,
|
| 252 |
+
"mean_abs_refinement": 0.2452070713043213
|
| 253 |
+
}
|
| 254 |
+
},
|
| 255 |
+
"notrack": {
|
| 256 |
+
"checkpoint": "checkpoints/ablate_notrack_3k/checkpoint-3000",
|
| 257 |
+
"use_track": false,
|
| 258 |
+
"use_force": true,
|
| 259 |
+
"use_deform_tactile": false,
|
| 260 |
+
"forward_tau_losses": {
|
| 261 |
+
"0.8": {
|
| 262 |
+
"action_loss": [
|
| 263 |
+
0.12394205729166667,
|
| 264 |
+
0.02089633950925863
|
| 265 |
+
],
|
| 266 |
+
"dynamics_loss": [
|
| 267 |
+
0.178466796875,
|
| 268 |
+
0.028673447679287072
|
| 269 |
+
],
|
| 270 |
+
"track_flow_loss": [
|
| 271 |
+
0.0,
|
| 272 |
+
0.0
|
| 273 |
+
],
|
| 274 |
+
"force_loss": [
|
| 275 |
+
0.7783203125,
|
| 276 |
+
0.043793139306208324
|
| 277 |
+
]
|
| 278 |
+
},
|
| 279 |
+
"0.6": {
|
| 280 |
+
"action_loss": [
|
| 281 |
+
0.19099934895833334,
|
| 282 |
+
0.028310372080589903
|
| 283 |
+
],
|
| 284 |
+
"dynamics_loss": [
|
| 285 |
+
0.22908528645833334,
|
| 286 |
+
0.028607776677988368
|
| 287 |
+
],
|
| 288 |
+
"track_flow_loss": [
|
| 289 |
+
0.0,
|
| 290 |
+
0.0
|
| 291 |
+
],
|
| 292 |
+
"force_loss": [
|
| 293 |
+
0.7783203125,
|
| 294 |
+
0.043793139306208324
|
| 295 |
+
]
|
| 296 |
+
},
|
| 297 |
+
"0.4": {
|
| 298 |
+
"action_loss": [
|
| 299 |
+
0.3876953125,
|
| 300 |
+
0.035242046640337726
|
| 301 |
+
],
|
| 302 |
+
"dynamics_loss": [
|
| 303 |
+
0.3413899739583333,
|
| 304 |
+
0.029684818436720592
|
| 305 |
+
],
|
| 306 |
+
"track_flow_loss": [
|
| 307 |
+
0.0,
|
| 308 |
+
0.0
|
| 309 |
+
],
|
| 310 |
+
"force_loss": [
|
| 311 |
+
0.7783203125,
|
| 312 |
+
0.043793139306208324
|
| 313 |
+
]
|
| 314 |
+
},
|
| 315 |
+
"0.2": {
|
| 316 |
+
"action_loss": [
|
| 317 |
+
0.7447916666666666,
|
| 318 |
+
0.0426817276121973
|
| 319 |
+
],
|
| 320 |
+
"dynamics_loss": [
|
| 321 |
+
0.6276041666666666,
|
| 322 |
+
0.024927330363381052
|
| 323 |
+
],
|
| 324 |
+
"track_flow_loss": [
|
| 325 |
+
0.0,
|
| 326 |
+
0.0
|
| 327 |
+
],
|
| 328 |
+
"force_loss": [
|
| 329 |
+
0.7783203125,
|
| 330 |
+
0.043793139306208324
|
| 331 |
+
]
|
| 332 |
+
},
|
| 333 |
+
"0.05": {
|
| 334 |
+
"action_loss": [
|
| 335 |
+
1.0777994791666667,
|
| 336 |
+
0.054336633312902415
|
| 337 |
+
],
|
| 338 |
+
"dynamics_loss": [
|
| 339 |
+
1.0641276041666667,
|
| 340 |
+
0.01625975747883096
|
| 341 |
+
],
|
| 342 |
+
"track_flow_loss": [
|
| 343 |
+
0.0,
|
| 344 |
+
0.0
|
| 345 |
+
],
|
| 346 |
+
"force_loss": [
|
| 347 |
+
0.7783203125,
|
| 348 |
+
0.043793139306208324
|
| 349 |
+
]
|
| 350 |
+
}
|
| 351 |
+
},
|
| 352 |
+
"openloop_action_mse": {
|
| 353 |
+
"cascade": [
|
| 354 |
+
0.22997993603348732,
|
| 355 |
+
0.1032914996028219,
|
| 356 |
+
16
|
| 357 |
+
],
|
| 358 |
+
"coarse_only": [
|
| 359 |
+
0.38339873775839806,
|
| 360 |
+
0.09786209841435693,
|
| 361 |
+
16
|
| 362 |
+
]
|
| 363 |
+
},
|
| 364 |
+
"openloop_action_mse_raw": {
|
| 365 |
+
"cascade": [
|
| 366 |
+
0.2659914791584015,
|
| 367 |
+
0.1821313500404358,
|
| 368 |
+
0.16993063688278198,
|
| 369 |
+
0.19111782312393188,
|
| 370 |
+
0.17413336038589478,
|
| 371 |
+
0.1670207381248474,
|
| 372 |
+
0.20298124849796295,
|
| 373 |
+
0.17849621176719666,
|
| 374 |
+
0.15974009037017822,
|
| 375 |
+
0.2785947620868683,
|
| 376 |
+
0.16809789836406708,
|
| 377 |
+
0.19190159440040588,
|
| 378 |
+
0.2161150574684143,
|
| 379 |
+
0.21241432428359985,
|
| 380 |
+
0.33366072177886963,
|
| 381 |
+
0.5873516798019409
|
| 382 |
+
],
|
| 383 |
+
"coarse_only": [
|
| 384 |
+
0.43769222497940063,
|
| 385 |
+
0.353431761264801,
|
| 386 |
+
0.32726842164993286,
|
| 387 |
+
0.3657931089401245,
|
| 388 |
+
0.33935850858688354,
|
| 389 |
+
0.30753999948501587,
|
| 390 |
+
0.3536660075187683,
|
| 391 |
+
0.34271180629730225,
|
| 392 |
+
0.2882019579410553,
|
| 393 |
+
0.3864496946334839,
|
| 394 |
+
0.32534658908843994,
|
| 395 |
+
0.320208877325058,
|
| 396 |
+
0.3476864695549011,
|
| 397 |
+
0.40916377305984497,
|
| 398 |
+
0.5438520312309265,
|
| 399 |
+
0.6860085725784302
|
| 400 |
+
]
|
| 401 |
+
},
|
| 402 |
+
"openloop_action_mse_split": {
|
| 403 |
+
"cascade_arm": 0.29843927826732397,
|
| 404 |
+
"cascade_hand": 0.2019738396629691,
|
| 405 |
+
"coarse_arm": 0.4064813628792763,
|
| 406 |
+
"coarse_hand": 0.373955849558115
|
| 407 |
+
},
|
| 408 |
+
"stage1_attention": {
|
| 409 |
+
"0": {
|
| 410 |
+
"action": {
|
| 411 |
+
"cond_obs": 0.20962195098400116,
|
| 412 |
+
"obs": 0.507138729095459,
|
| 413 |
+
"action": 0.27125489711761475,
|
| 414 |
+
"state": 0.011984430253505707,
|
| 415 |
+
"track_past": 0.0,
|
| 416 |
+
"track_future": 0.0
|
| 417 |
+
},
|
| 418 |
+
"obs": {
|
| 419 |
+
"cond_obs": 0.16306498646736145,
|
| 420 |
+
"obs": 0.7571282386779785,
|
| 421 |
+
"action": 0.07827480137348175,
|
| 422 |
+
"state": 0.001532059395685792,
|
| 423 |
+
"track_past": 0.0,
|
| 424 |
+
"track_future": 0.0
|
| 425 |
+
}
|
| 426 |
+
},
|
| 427 |
+
"14": {
|
| 428 |
+
"action": {
|
| 429 |
+
"cond_obs": 0.09229698032140732,
|
| 430 |
+
"obs": 0.18946343660354614,
|
| 431 |
+
"action": 0.6833343505859375,
|
| 432 |
+
"state": 0.03490526229143143,
|
| 433 |
+
"track_past": 0.0,
|
| 434 |
+
"track_future": 0.0
|
| 435 |
+
},
|
| 436 |
+
"obs": {
|
| 437 |
+
"cond_obs": 0.26676031947135925,
|
| 438 |
+
"obs": 0.6440349221229553,
|
| 439 |
+
"action": 0.054011691361665726,
|
| 440 |
+
"state": 0.03519308567047119,
|
| 441 |
+
"track_past": 0.0,
|
| 442 |
+
"track_future": 0.0
|
| 443 |
+
}
|
| 444 |
+
},
|
| 445 |
+
"29": {
|
| 446 |
+
"action": {
|
| 447 |
+
"cond_obs": 0.10308176279067993,
|
| 448 |
+
"obs": 0.24439243972301483,
|
| 449 |
+
"action": 0.6442670822143555,
|
| 450 |
+
"state": 0.008258762769401073,
|
| 451 |
+
"track_past": 0.0,
|
| 452 |
+
"track_future": 0.0
|
| 453 |
+
},
|
| 454 |
+
"obs": {
|
| 455 |
+
"cond_obs": 0.15114127099514008,
|
| 456 |
+
"obs": 0.8275548219680786,
|
| 457 |
+
"action": 0.019980238750576973,
|
| 458 |
+
"state": 0.0013237263774499297,
|
| 459 |
+
"track_past": 0.0,
|
| 460 |
+
"track_future": 0.0
|
| 461 |
+
}
|
| 462 |
+
}
|
| 463 |
+
},
|
| 464 |
+
"stage2_attention": [
|
| 465 |
+
{
|
| 466 |
+
"action": 0.1103515625,
|
| 467 |
+
"f6": 0.0625,
|
| 468 |
+
"vq": 0.06640625,
|
| 469 |
+
"memory": 0.76171875
|
| 470 |
+
},
|
| 471 |
+
{
|
| 472 |
+
"action": 0.1904296875,
|
| 473 |
+
"f6": 0.09814453125,
|
| 474 |
+
"vq": 0.0966796875,
|
| 475 |
+
"memory": 0.61328125
|
| 476 |
+
},
|
| 477 |
+
{
|
| 478 |
+
"action": 0.3359375,
|
| 479 |
+
"f6": 0.146484375,
|
| 480 |
+
"vq": 0.1513671875,
|
| 481 |
+
"memory": 0.365234375
|
| 482 |
+
},
|
| 483 |
+
{
|
| 484 |
+
"action": 0.283203125,
|
| 485 |
+
"f6": 0.12353515625,
|
| 486 |
+
"vq": 0.1259765625,
|
| 487 |
+
"memory": 0.466796875
|
| 488 |
+
},
|
| 489 |
+
{
|
| 490 |
+
"action": 0.193359375,
|
| 491 |
+
"f6": 0.08984375,
|
| 492 |
+
"vq": 0.09716796875,
|
| 493 |
+
"memory": 0.62109375
|
| 494 |
+
},
|
| 495 |
+
{
|
| 496 |
+
"action": 0.1298828125,
|
| 497 |
+
"f6": 0.072265625,
|
| 498 |
+
"vq": 0.0771484375,
|
| 499 |
+
"memory": 0.71875
|
| 500 |
+
}
|
| 501 |
+
],
|
| 502 |
+
"tactile_sensitivity": {
|
| 503 |
+
"mean_abs_delta_vs_masked": 0.025584066286683083,
|
| 504 |
+
"max_abs_delta_vs_masked": 0.130859375,
|
| 505 |
+
"mean_abs_refinement": 0.24716320633888245
|
| 506 |
+
}
|
| 507 |
+
},
|
| 508 |
+
"noforce": {
|
| 509 |
+
"checkpoint": "checkpoints/ablate_noforce_3k/checkpoint-3000",
|
| 510 |
+
"use_track": true,
|
| 511 |
+
"use_force": false,
|
| 512 |
+
"use_deform_tactile": false,
|
| 513 |
+
"forward_tau_losses": {
|
| 514 |
+
"0.8": {
|
| 515 |
+
"action_loss": [
|
| 516 |
+
0.1353759765625,
|
| 517 |
+
0.02098532703746221
|
| 518 |
+
],
|
| 519 |
+
"dynamics_loss": [
|
| 520 |
+
0.21435546875,
|
| 521 |
+
0.02948615073426333
|
| 522 |
+
],
|
| 523 |
+
"track_flow_loss": [
|
| 524 |
+
0.018323262532552082,
|
| 525 |
+
0.005343945969003892
|
| 526 |
+
],
|
| 527 |
+
"force_loss": [
|
| 528 |
+
0.0,
|
| 529 |
+
0.0
|
| 530 |
+
]
|
| 531 |
+
},
|
| 532 |
+
"0.6": {
|
| 533 |
+
"action_loss": [
|
| 534 |
+
0.20084635416666666,
|
| 535 |
+
0.028057648818818122
|
| 536 |
+
],
|
| 537 |
+
"dynamics_loss": [
|
| 538 |
+
0.2789306640625,
|
| 539 |
+
0.027221132281491568
|
| 540 |
+
],
|
| 541 |
+
"track_flow_loss": [
|
| 542 |
+
0.023956298828125,
|
| 543 |
+
0.008590738896058613
|
| 544 |
+
],
|
| 545 |
+
"force_loss": [
|
| 546 |
+
0.0,
|
| 547 |
+
0.0
|
| 548 |
+
]
|
| 549 |
+
},
|
| 550 |
+
"0.4": {
|
| 551 |
+
"action_loss": [
|
| 552 |
+
0.396240234375,
|
| 553 |
+
0.04582674094675714
|
| 554 |
+
],
|
| 555 |
+
"dynamics_loss": [
|
| 556 |
+
0.41259765625,
|
| 557 |
+
0.025191085421407685
|
| 558 |
+
],
|
| 559 |
+
"track_flow_loss": [
|
| 560 |
+
0.049285888671875,
|
| 561 |
+
0.01706895005166986
|
| 562 |
+
],
|
| 563 |
+
"force_loss": [
|
| 564 |
+
0.0,
|
| 565 |
+
0.0
|
| 566 |
+
]
|
| 567 |
+
},
|
| 568 |
+
"0.2": {
|
| 569 |
+
"action_loss": [
|
| 570 |
+
0.7443033854166666,
|
| 571 |
+
0.04709123258240433
|
| 572 |
+
],
|
| 573 |
+
"dynamics_loss": [
|
| 574 |
+
0.72216796875,
|
| 575 |
+
0.02011853062365024
|
| 576 |
+
],
|
| 577 |
+
"track_flow_loss": [
|
| 578 |
+
0.2680257161458333,
|
| 579 |
+
0.02835253639775798
|
| 580 |
+
],
|
| 581 |
+
"force_loss": [
|
| 582 |
+
0.0,
|
| 583 |
+
0.0
|
| 584 |
+
]
|
| 585 |
+
},
|
| 586 |
+
"0.05": {
|
| 587 |
+
"action_loss": [
|
| 588 |
+
1.0846354166666667,
|
| 589 |
+
0.04728911405555957
|
| 590 |
+
],
|
| 591 |
+
"dynamics_loss": [
|
| 592 |
+
1.1276041666666667,
|
| 593 |
+
0.014023866685246106
|
| 594 |
+
],
|
| 595 |
+
"track_flow_loss": [
|
| 596 |
+
0.90869140625,
|
| 597 |
+
0.026584810349459555
|
| 598 |
+
],
|
| 599 |
+
"force_loss": [
|
| 600 |
+
0.0,
|
| 601 |
+
0.0
|
| 602 |
+
]
|
| 603 |
+
}
|
| 604 |
+
},
|
| 605 |
+
"openloop_action_mse": {
|
| 606 |
+
"coarse_only": [
|
| 607 |
+
0.3794044554233551,
|
| 608 |
+
0.09120250836697517,
|
| 609 |
+
16
|
| 610 |
+
]
|
| 611 |
+
},
|
| 612 |
+
"openloop_action_mse_raw": {
|
| 613 |
+
"coarse_only": [
|
| 614 |
+
0.44760891795158386,
|
| 615 |
+
0.3614843189716339,
|
| 616 |
+
0.3224659860134125,
|
| 617 |
+
0.3654090464115143,
|
| 618 |
+
0.3293968141078949,
|
| 619 |
+
0.3092474639415741,
|
| 620 |
+
0.3457678556442261,
|
| 621 |
+
0.33577266335487366,
|
| 622 |
+
0.29082638025283813,
|
| 623 |
+
0.38748833537101746,
|
| 624 |
+
0.3380368947982788,
|
| 625 |
+
0.3234664499759674,
|
| 626 |
+
0.34388527274131775,
|
| 627 |
+
0.4068928360939026,
|
| 628 |
+
0.4868208169937134,
|
| 629 |
+
0.6759012341499329
|
| 630 |
+
]
|
| 631 |
+
},
|
| 632 |
+
"openloop_action_mse_split": {
|
| 633 |
+
"coarse_arm": 0.4130093362182379,
|
| 634 |
+
"coarse_hand": 0.36565699987113476
|
| 635 |
+
},
|
| 636 |
+
"stage1_attention": {
|
| 637 |
+
"0": {
|
| 638 |
+
"action": {
|
| 639 |
+
"cond_obs": 0.029441919177770615,
|
| 640 |
+
"obs": 0.10868024826049805,
|
| 641 |
+
"action": 0.057694464921951294,
|
| 642 |
+
"state": 0.0018944537732750177,
|
| 643 |
+
"track_past": 0.4593313932418823,
|
| 644 |
+
"track_future": 0.3429574966430664
|
| 645 |
+
},
|
| 646 |
+
"obs": {
|
| 647 |
+
"cond_obs": 0.0534292496740818,
|
| 648 |
+
"obs": 0.5819946527481079,
|
| 649 |
+
"action": 0.009241057559847832,
|
| 650 |
+
"state": 0.0005529792397283018,
|
| 651 |
+
"track_past": 0.17858824133872986,
|
| 652 |
+
"track_future": 0.17619383335113525
|
| 653 |
+
}
|
| 654 |
+
},
|
| 655 |
+
"14": {
|
| 656 |
+
"action": {
|
| 657 |
+
"cond_obs": 0.052035462111234665,
|
| 658 |
+
"obs": 0.10198719799518585,
|
| 659 |
+
"action": 0.19155430793762207,
|
| 660 |
+
"state": 0.0012913987739011645,
|
| 661 |
+
"track_past": 0.6307284832000732,
|
| 662 |
+
"track_future": 0.02240319363772869
|
| 663 |
+
},
|
| 664 |
+
"obs": {
|
| 665 |
+
"cond_obs": 0.16328012943267822,
|
| 666 |
+
"obs": 0.6316336393356323,
|
| 667 |
+
"action": 0.015135358087718487,
|
| 668 |
+
"state": 0.002750932238996029,
|
| 669 |
+
"track_past": 0.1714852750301361,
|
| 670 |
+
"track_future": 0.015714673325419426
|
| 671 |
+
}
|
| 672 |
+
},
|
| 673 |
+
"29": {
|
| 674 |
+
"action": {
|
| 675 |
+
"cond_obs": 0.05154501646757126,
|
| 676 |
+
"obs": 0.09563900530338287,
|
| 677 |
+
"action": 0.19082441926002502,
|
| 678 |
+
"state": 0.0022456112783402205,
|
| 679 |
+
"track_past": 0.5803318023681641,
|
| 680 |
+
"track_future": 0.07941415905952454
|
| 681 |
+
},
|
| 682 |
+
"obs": {
|
| 683 |
+
"cond_obs": 0.09326619654893875,
|
| 684 |
+
"obs": 0.7765737175941467,
|
| 685 |
+
"action": 0.010622730478644371,
|
| 686 |
+
"state": 0.00036804963019676507,
|
| 687 |
+
"track_past": 0.07503131777048111,
|
| 688 |
+
"track_future": 0.044137973338365555
|
| 689 |
+
}
|
| 690 |
+
}
|
| 691 |
+
}
|
| 692 |
+
},
|
| 693 |
+
"deform": {
|
| 694 |
+
"checkpoint": "checkpoints/ablate_deform_3k/checkpoint-3000",
|
| 695 |
+
"use_track": true,
|
| 696 |
+
"use_force": true,
|
| 697 |
+
"use_deform_tactile": true,
|
| 698 |
+
"forward_tau_losses": {
|
| 699 |
+
"0.8": {
|
| 700 |
+
"action_loss": [
|
| 701 |
+
0.14005533854166666,
|
| 702 |
+
0.02105122113243602
|
| 703 |
+
],
|
| 704 |
+
"dynamics_loss": [
|
| 705 |
+
0.2144775390625,
|
| 706 |
+
0.02896240896365155
|
| 707 |
+
],
|
| 708 |
+
"track_flow_loss": [
|
| 709 |
+
0.01992034912109375,
|
| 710 |
+
0.006087289540532924
|
| 711 |
+
],
|
| 712 |
+
"force_loss": [
|
| 713 |
+
0.6385091145833334,
|
| 714 |
+
0.0693288689056317
|
| 715 |
+
]
|
| 716 |
+
},
|
| 717 |
+
"0.6": {
|
| 718 |
+
"action_loss": [
|
| 719 |
+
0.205078125,
|
| 720 |
+
0.02730302957532159
|
| 721 |
+
],
|
| 722 |
+
"dynamics_loss": [
|
| 723 |
+
0.2793782552083333,
|
| 724 |
+
0.027562192277447207
|
| 725 |
+
],
|
| 726 |
+
"track_flow_loss": [
|
| 727 |
+
0.025990804036458332,
|
| 728 |
+
0.009586659999509849
|
| 729 |
+
],
|
| 730 |
+
"force_loss": [
|
| 731 |
+
0.6385091145833334,
|
| 732 |
+
0.0693288689056317
|
| 733 |
+
]
|
| 734 |
+
},
|
| 735 |
+
"0.4": {
|
| 736 |
+
"action_loss": [
|
| 737 |
+
0.3975423177083333,
|
| 738 |
+
0.03933005362916035
|
| 739 |
+
],
|
| 740 |
+
"dynamics_loss": [
|
| 741 |
+
0.412841796875,
|
| 742 |
+
0.02506180118084375
|
| 743 |
+
],
|
| 744 |
+
"track_flow_loss": [
|
| 745 |
+
0.052449544270833336,
|
| 746 |
+
0.018266911258591405
|
| 747 |
+
],
|
| 748 |
+
"force_loss": [
|
| 749 |
+
0.6385091145833334,
|
| 750 |
+
0.0693288689056317
|
| 751 |
+
]
|
| 752 |
+
},
|
| 753 |
+
"0.2": {
|
| 754 |
+
"action_loss": [
|
| 755 |
+
0.7540690104166666,
|
| 756 |
+
0.04711147986173765
|
| 757 |
+
],
|
| 758 |
+
"dynamics_loss": [
|
| 759 |
+
0.7234700520833334,
|
| 760 |
+
0.020390574072347532
|
| 761 |
+
],
|
| 762 |
+
"track_flow_loss": [
|
| 763 |
+
0.2725016276041667,
|
| 764 |
+
0.028762320444776262
|
| 765 |
+
],
|
| 766 |
+
"force_loss": [
|
| 767 |
+
0.6385091145833334,
|
| 768 |
+
0.0693288689056317
|
| 769 |
+
]
|
| 770 |
+
},
|
| 771 |
+
"0.05": {
|
| 772 |
+
"action_loss": [
|
| 773 |
+
1.0712890625,
|
| 774 |
+
0.05129510151042538
|
| 775 |
+
],
|
| 776 |
+
"dynamics_loss": [
|
| 777 |
+
1.1285807291666667,
|
| 778 |
+
0.015292551623107222
|
| 779 |
+
],
|
| 780 |
+
"track_flow_loss": [
|
| 781 |
+
0.9132486979166666,
|
| 782 |
+
0.026444936649788933
|
| 783 |
+
],
|
| 784 |
+
"force_loss": [
|
| 785 |
+
0.6385091145833334,
|
| 786 |
+
0.0693288689056317
|
| 787 |
+
]
|
| 788 |
+
}
|
| 789 |
+
},
|
| 790 |
+
"openloop_action_mse": {
|
| 791 |
+
"cascade": [
|
| 792 |
+
0.20164753962308168,
|
| 793 |
+
0.10256788707075791,
|
| 794 |
+
16
|
| 795 |
+
],
|
| 796 |
+
"coarse_only": [
|
| 797 |
+
0.38597522117197514,
|
| 798 |
+
0.09435245328291447,
|
| 799 |
+
16
|
| 800 |
+
]
|
| 801 |
+
},
|
| 802 |
+
"openloop_action_mse_raw": {
|
| 803 |
+
"cascade": [
|
| 804 |
+
0.239223912358284,
|
| 805 |
+
0.16198278963565826,
|
| 806 |
+
0.14795620739459991,
|
| 807 |
+
0.1607406884431839,
|
| 808 |
+
0.13953106105327606,
|
| 809 |
+
0.14302007853984833,
|
| 810 |
+
0.17535518109798431,
|
| 811 |
+
0.14668907225131989,
|
| 812 |
+
0.14609704911708832,
|
| 813 |
+
0.255096971988678,
|
| 814 |
+
0.14849339425563812,
|
| 815 |
+
0.1533306986093521,
|
| 816 |
+
0.18130069971084595,
|
| 817 |
+
0.1971125304698944,
|
| 818 |
+
0.26189425587654114,
|
| 819 |
+
0.5685360431671143
|
| 820 |
+
],
|
| 821 |
+
"coarse_only": [
|
| 822 |
+
0.4410543143749237,
|
| 823 |
+
0.35842546820640564,
|
| 824 |
+
0.32901355624198914,
|
| 825 |
+
0.3685969412326813,
|
| 826 |
+
0.33627718687057495,
|
| 827 |
+
0.3183954656124115,
|
| 828 |
+
0.3497931659221649,
|
| 829 |
+
0.34323593974113464,
|
| 830 |
+
0.3002097010612488,
|
| 831 |
+
0.39490291476249695,
|
| 832 |
+
0.34140878915786743,
|
| 833 |
+
0.3261561691761017,
|
| 834 |
+
0.34672024846076965,
|
| 835 |
+
0.4223487675189972,
|
| 836 |
+
0.5050632953643799,
|
| 837 |
+
0.6940016150474548
|
| 838 |
+
]
|
| 839 |
+
},
|
| 840 |
+
"openloop_action_mse_split": {
|
| 841 |
+
"cascade_arm": 0.2864319020882249,
|
| 842 |
+
"cascade_hand": 0.16696302872151136,
|
| 843 |
+
"coarse_arm": 0.4209964629262686,
|
| 844 |
+
"coarse_hand": 0.3716483376920223
|
| 845 |
+
},
|
| 846 |
+
"stage1_attention": {
|
| 847 |
+
"0": {
|
| 848 |
+
"action": {
|
| 849 |
+
"cond_obs": 0.027012702077627182,
|
| 850 |
+
"obs": 0.10060404241085052,
|
| 851 |
+
"action": 0.06752102077007294,
|
| 852 |
+
"state": 0.0014766933163627982,
|
| 853 |
+
"track_past": 0.45365235209465027,
|
| 854 |
+
"track_future": 0.34973323345184326
|
| 855 |
+
},
|
| 856 |
+
"obs": {
|
| 857 |
+
"cond_obs": 0.053314413875341415,
|
| 858 |
+
"obs": 0.5820602774620056,
|
| 859 |
+
"action": 0.008973240852355957,
|
| 860 |
+
"state": 0.0005145587492734194,
|
| 861 |
+
"track_past": 0.17881375551223755,
|
| 862 |
+
"track_future": 0.1763237565755844
|
| 863 |
+
}
|
| 864 |
+
},
|
| 865 |
+
"14": {
|
| 866 |
+
"action": {
|
| 867 |
+
"cond_obs": 0.04775896668434143,
|
| 868 |
+
"obs": 0.08490505069494247,
|
| 869 |
+
"action": 0.20447513461112976,
|
| 870 |
+
"state": 0.005050991661846638,
|
| 871 |
+
"track_past": 0.6356136202812195,
|
| 872 |
+
"track_future": 0.02219625934958458
|
| 873 |
+
},
|
| 874 |
+
"obs": {
|
| 875 |
+
"cond_obs": 0.1583261936903,
|
| 876 |
+
"obs": 0.6368992328643799,
|
| 877 |
+
"action": 0.014521163888275623,
|
| 878 |
+
"state": 0.007083813659846783,
|
| 879 |
+
"track_past": 0.16779881715774536,
|
| 880 |
+
"track_future": 0.015370833687484264
|
| 881 |
+
}
|
| 882 |
+
},
|
| 883 |
+
"29": {
|
| 884 |
+
"action": {
|
| 885 |
+
"cond_obs": 0.052446119487285614,
|
| 886 |
+
"obs": 0.09761743247509003,
|
| 887 |
+
"action": 0.17695856094360352,
|
| 888 |
+
"state": 0.0029112468473613262,
|
| 889 |
+
"track_past": 0.5859178900718689,
|
| 890 |
+
"track_future": 0.08414878696203232
|
| 891 |
+
},
|
| 892 |
+
"obs": {
|
| 893 |
+
"cond_obs": 0.09640265256166458,
|
| 894 |
+
"obs": 0.7698864340782166,
|
| 895 |
+
"action": 0.010517171584069729,
|
| 896 |
+
"state": 0.0004369919188320637,
|
| 897 |
+
"track_past": 0.07651636004447937,
|
| 898 |
+
"track_future": 0.04624041169881821
|
| 899 |
+
}
|
| 900 |
+
}
|
| 901 |
+
},
|
| 902 |
+
"stage2_attention": [
|
| 903 |
+
{
|
| 904 |
+
"action": 0.026611328125,
|
| 905 |
+
"f6": 0.0159912109375,
|
| 906 |
+
"vq": 0.0179443359375,
|
| 907 |
+
"deform": 0.017822265625,
|
| 908 |
+
"memory": 0.921875
|
| 909 |
+
},
|
| 910 |
+
{
|
| 911 |
+
"action": 0.0267333984375,
|
| 912 |
+
"f6": 0.0147705078125,
|
| 913 |
+
"vq": 0.0167236328125,
|
| 914 |
+
"deform": 0.0179443359375,
|
| 915 |
+
"memory": 0.921875
|
| 916 |
+
},
|
| 917 |
+
{
|
| 918 |
+
"action": 0.027587890625,
|
| 919 |
+
"f6": 0.01470947265625,
|
| 920 |
+
"vq": 0.0164794921875,
|
| 921 |
+
"deform": 0.0172119140625,
|
| 922 |
+
"memory": 0.92578125
|
| 923 |
+
},
|
| 924 |
+
{
|
| 925 |
+
"action": 0.0269775390625,
|
| 926 |
+
"f6": 0.01409912109375,
|
| 927 |
+
"vq": 0.015869140625,
|
| 928 |
+
"deform": 0.016357421875,
|
| 929 |
+
"memory": 0.92578125
|
| 930 |
+
},
|
| 931 |
+
{
|
| 932 |
+
"action": 0.0250244140625,
|
| 933 |
+
"f6": 0.01324462890625,
|
| 934 |
+
"vq": 0.01446533203125,
|
| 935 |
+
"deform": 0.01495361328125,
|
| 936 |
+
"memory": 0.93359375
|
| 937 |
+
},
|
| 938 |
+
{
|
| 939 |
+
"action": 0.0225830078125,
|
| 940 |
+
"f6": 0.0125732421875,
|
| 941 |
+
"vq": 0.01300048828125,
|
| 942 |
+
"deform": 0.01348876953125,
|
| 943 |
+
"memory": 0.9375
|
| 944 |
+
}
|
| 945 |
+
],
|
| 946 |
+
"tactile_sensitivity": {
|
| 947 |
+
"mean_abs_delta_vs_masked": 0.01314772292971611,
|
| 948 |
+
"max_abs_delta_vs_masked": 0.0546875,
|
| 949 |
+
"mean_abs_refinement": 0.2471647560596466
|
| 950 |
+
}
|
| 951 |
+
}
|
| 952 |
+
}
|
ablation_runs/eval_d.json
ADDED
|
@@ -0,0 +1,261 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"deform": {
|
| 3 |
+
"checkpoint": "checkpoints/ablate_deform_3k/checkpoint-3000",
|
| 4 |
+
"use_track": true,
|
| 5 |
+
"use_force": true,
|
| 6 |
+
"use_deform_tactile": true,
|
| 7 |
+
"forward_tau_losses": {
|
| 8 |
+
"0.8": {
|
| 9 |
+
"action_loss": [
|
| 10 |
+
0.14005533854166666,
|
| 11 |
+
0.02105122113243602
|
| 12 |
+
],
|
| 13 |
+
"dynamics_loss": [
|
| 14 |
+
0.2144775390625,
|
| 15 |
+
0.02896240896365155
|
| 16 |
+
],
|
| 17 |
+
"track_flow_loss": [
|
| 18 |
+
0.01992034912109375,
|
| 19 |
+
0.006087289540532924
|
| 20 |
+
],
|
| 21 |
+
"force_loss": [
|
| 22 |
+
0.6385091145833334,
|
| 23 |
+
0.0693288689056317
|
| 24 |
+
]
|
| 25 |
+
},
|
| 26 |
+
"0.6": {
|
| 27 |
+
"action_loss": [
|
| 28 |
+
0.205078125,
|
| 29 |
+
0.02730302957532159
|
| 30 |
+
],
|
| 31 |
+
"dynamics_loss": [
|
| 32 |
+
0.2793782552083333,
|
| 33 |
+
0.027562192277447207
|
| 34 |
+
],
|
| 35 |
+
"track_flow_loss": [
|
| 36 |
+
0.025990804036458332,
|
| 37 |
+
0.009586659999509849
|
| 38 |
+
],
|
| 39 |
+
"force_loss": [
|
| 40 |
+
0.6385091145833334,
|
| 41 |
+
0.0693288689056317
|
| 42 |
+
]
|
| 43 |
+
},
|
| 44 |
+
"0.4": {
|
| 45 |
+
"action_loss": [
|
| 46 |
+
0.3975423177083333,
|
| 47 |
+
0.03933005362916035
|
| 48 |
+
],
|
| 49 |
+
"dynamics_loss": [
|
| 50 |
+
0.412841796875,
|
| 51 |
+
0.02506180118084375
|
| 52 |
+
],
|
| 53 |
+
"track_flow_loss": [
|
| 54 |
+
0.052449544270833336,
|
| 55 |
+
0.018266911258591405
|
| 56 |
+
],
|
| 57 |
+
"force_loss": [
|
| 58 |
+
0.6385091145833334,
|
| 59 |
+
0.0693288689056317
|
| 60 |
+
]
|
| 61 |
+
},
|
| 62 |
+
"0.2": {
|
| 63 |
+
"action_loss": [
|
| 64 |
+
0.7540690104166666,
|
| 65 |
+
0.04711147986173765
|
| 66 |
+
],
|
| 67 |
+
"dynamics_loss": [
|
| 68 |
+
0.7234700520833334,
|
| 69 |
+
0.020390574072347532
|
| 70 |
+
],
|
| 71 |
+
"track_flow_loss": [
|
| 72 |
+
0.2725016276041667,
|
| 73 |
+
0.028762320444776262
|
| 74 |
+
],
|
| 75 |
+
"force_loss": [
|
| 76 |
+
0.6385091145833334,
|
| 77 |
+
0.0693288689056317
|
| 78 |
+
]
|
| 79 |
+
},
|
| 80 |
+
"0.05": {
|
| 81 |
+
"action_loss": [
|
| 82 |
+
1.0712890625,
|
| 83 |
+
0.05129510151042538
|
| 84 |
+
],
|
| 85 |
+
"dynamics_loss": [
|
| 86 |
+
1.1285807291666667,
|
| 87 |
+
0.015292551623107222
|
| 88 |
+
],
|
| 89 |
+
"track_flow_loss": [
|
| 90 |
+
0.9132486979166666,
|
| 91 |
+
0.026444936649788933
|
| 92 |
+
],
|
| 93 |
+
"force_loss": [
|
| 94 |
+
0.6385091145833334,
|
| 95 |
+
0.0693288689056317
|
| 96 |
+
]
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"openloop_action_mse": {
|
| 100 |
+
"cascade": [
|
| 101 |
+
0.20164753962308168,
|
| 102 |
+
0.10256788707075791,
|
| 103 |
+
16
|
| 104 |
+
],
|
| 105 |
+
"coarse_only": [
|
| 106 |
+
0.38597522117197514,
|
| 107 |
+
0.09435245328291447,
|
| 108 |
+
16
|
| 109 |
+
]
|
| 110 |
+
},
|
| 111 |
+
"openloop_action_mse_raw": {
|
| 112 |
+
"cascade": [
|
| 113 |
+
0.239223912358284,
|
| 114 |
+
0.16198278963565826,
|
| 115 |
+
0.14795620739459991,
|
| 116 |
+
0.1607406884431839,
|
| 117 |
+
0.13953106105327606,
|
| 118 |
+
0.14302007853984833,
|
| 119 |
+
0.17535518109798431,
|
| 120 |
+
0.14668907225131989,
|
| 121 |
+
0.14609704911708832,
|
| 122 |
+
0.255096971988678,
|
| 123 |
+
0.14849339425563812,
|
| 124 |
+
0.1533306986093521,
|
| 125 |
+
0.18130069971084595,
|
| 126 |
+
0.1971125304698944,
|
| 127 |
+
0.26189425587654114,
|
| 128 |
+
0.5685360431671143
|
| 129 |
+
],
|
| 130 |
+
"coarse_only": [
|
| 131 |
+
0.4410543143749237,
|
| 132 |
+
0.35842546820640564,
|
| 133 |
+
0.32901355624198914,
|
| 134 |
+
0.3685969412326813,
|
| 135 |
+
0.33627718687057495,
|
| 136 |
+
0.3183954656124115,
|
| 137 |
+
0.3497931659221649,
|
| 138 |
+
0.34323593974113464,
|
| 139 |
+
0.3002097010612488,
|
| 140 |
+
0.39490291476249695,
|
| 141 |
+
0.34140878915786743,
|
| 142 |
+
0.3261561691761017,
|
| 143 |
+
0.34672024846076965,
|
| 144 |
+
0.4223487675189972,
|
| 145 |
+
0.5050632953643799,
|
| 146 |
+
0.6940016150474548
|
| 147 |
+
]
|
| 148 |
+
},
|
| 149 |
+
"openloop_action_mse_split": {
|
| 150 |
+
"cascade_arm": 0.2864319020882249,
|
| 151 |
+
"cascade_hand": 0.16696302872151136,
|
| 152 |
+
"coarse_arm": 0.4209964629262686,
|
| 153 |
+
"coarse_hand": 0.3716483376920223
|
| 154 |
+
},
|
| 155 |
+
"stage1_attention": {
|
| 156 |
+
"0": {
|
| 157 |
+
"action": {
|
| 158 |
+
"cond_obs": 0.027012702077627182,
|
| 159 |
+
"obs": 0.10060404241085052,
|
| 160 |
+
"action": 0.06752102077007294,
|
| 161 |
+
"state": 0.0014766933163627982,
|
| 162 |
+
"track_past": 0.45365235209465027,
|
| 163 |
+
"track_future": 0.34973323345184326
|
| 164 |
+
},
|
| 165 |
+
"obs": {
|
| 166 |
+
"cond_obs": 0.053314413875341415,
|
| 167 |
+
"obs": 0.5820602774620056,
|
| 168 |
+
"action": 0.008973240852355957,
|
| 169 |
+
"state": 0.0005145587492734194,
|
| 170 |
+
"track_past": 0.17881375551223755,
|
| 171 |
+
"track_future": 0.1763237565755844
|
| 172 |
+
}
|
| 173 |
+
},
|
| 174 |
+
"14": {
|
| 175 |
+
"action": {
|
| 176 |
+
"cond_obs": 0.04775896668434143,
|
| 177 |
+
"obs": 0.08490505069494247,
|
| 178 |
+
"action": 0.20447513461112976,
|
| 179 |
+
"state": 0.005050991661846638,
|
| 180 |
+
"track_past": 0.6356136202812195,
|
| 181 |
+
"track_future": 0.02219625934958458
|
| 182 |
+
},
|
| 183 |
+
"obs": {
|
| 184 |
+
"cond_obs": 0.1583261936903,
|
| 185 |
+
"obs": 0.6368992328643799,
|
| 186 |
+
"action": 0.014521163888275623,
|
| 187 |
+
"state": 0.007083813659846783,
|
| 188 |
+
"track_past": 0.16779881715774536,
|
| 189 |
+
"track_future": 0.015370833687484264
|
| 190 |
+
}
|
| 191 |
+
},
|
| 192 |
+
"29": {
|
| 193 |
+
"action": {
|
| 194 |
+
"cond_obs": 0.052446119487285614,
|
| 195 |
+
"obs": 0.09761743247509003,
|
| 196 |
+
"action": 0.17695856094360352,
|
| 197 |
+
"state": 0.0029112468473613262,
|
| 198 |
+
"track_past": 0.5859178900718689,
|
| 199 |
+
"track_future": 0.08414878696203232
|
| 200 |
+
},
|
| 201 |
+
"obs": {
|
| 202 |
+
"cond_obs": 0.09640265256166458,
|
| 203 |
+
"obs": 0.7698864340782166,
|
| 204 |
+
"action": 0.010517171584069729,
|
| 205 |
+
"state": 0.0004369919188320637,
|
| 206 |
+
"track_past": 0.07651636004447937,
|
| 207 |
+
"track_future": 0.04624041169881821
|
| 208 |
+
}
|
| 209 |
+
}
|
| 210 |
+
},
|
| 211 |
+
"stage2_attention": [
|
| 212 |
+
{
|
| 213 |
+
"action": 0.026611328125,
|
| 214 |
+
"f6": 0.0159912109375,
|
| 215 |
+
"vq": 0.0179443359375,
|
| 216 |
+
"deform": 0.017822265625,
|
| 217 |
+
"memory": 0.921875
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"action": 0.0267333984375,
|
| 221 |
+
"f6": 0.0147705078125,
|
| 222 |
+
"vq": 0.0167236328125,
|
| 223 |
+
"deform": 0.0179443359375,
|
| 224 |
+
"memory": 0.921875
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"action": 0.027587890625,
|
| 228 |
+
"f6": 0.01470947265625,
|
| 229 |
+
"vq": 0.0164794921875,
|
| 230 |
+
"deform": 0.0172119140625,
|
| 231 |
+
"memory": 0.92578125
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"action": 0.0269775390625,
|
| 235 |
+
"f6": 0.01409912109375,
|
| 236 |
+
"vq": 0.015869140625,
|
| 237 |
+
"deform": 0.016357421875,
|
| 238 |
+
"memory": 0.92578125
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"action": 0.0250244140625,
|
| 242 |
+
"f6": 0.01324462890625,
|
| 243 |
+
"vq": 0.01446533203125,
|
| 244 |
+
"deform": 0.01495361328125,
|
| 245 |
+
"memory": 0.93359375
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"action": 0.0225830078125,
|
| 249 |
+
"f6": 0.0125732421875,
|
| 250 |
+
"vq": 0.01300048828125,
|
| 251 |
+
"deform": 0.01348876953125,
|
| 252 |
+
"memory": 0.9375
|
| 253 |
+
}
|
| 254 |
+
],
|
| 255 |
+
"tactile_sensitivity": {
|
| 256 |
+
"mean_abs_delta_vs_masked": 0.01314772292971611,
|
| 257 |
+
"max_abs_delta_vs_masked": 0.0546875,
|
| 258 |
+
"mean_abs_refinement": 0.2471647560596466
|
| 259 |
+
}
|
| 260 |
+
}
|
| 261 |
+
}
|
ablation_runs/findings.html
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<h3>What the three questions get answered</h3>
|
| 2 |
+
<div class="prose">
|
| 3 |
+
<p><b>1 · Do the track tokens help?</b> Yes — and the mechanism is
|
| 4 |
+
visible. With identical budgets, removing the track pathway leaves the
|
| 5 |
+
single-step action flow MSE almost unchanged (0.745 vs 0.752 at τ=0.2)
|
| 6 |
+
but costs <b>+12% open-loop cascade error</b> (0.230 vs 0.205, paired
|
| 7 |
+
+0.0245 ± 0.0033 over the same 16 anchors and seeds), with the
|
| 8 |
+
largest hit on the hand dimensions (+19%). The routing explains why: in the
|
| 9 |
+
full model the action queries put <b>~80% of their self-attention mass on
|
| 10 |
+
track tokens</b> (past + future), so the plan flows through the geometric
|
| 11 |
+
pathway; without it they fall back onto raw observation patches and
|
| 12 |
+
themselves. The loss also propagates downstream — the force stage reads a
|
| 13 |
+
poorer coarse memory and its flow MSE degrades from 0.643 to 0.778, while
|
| 14 |
+
its reliance on tactile tokens roughly triples (attention mass 13–15% vs
|
| 15 |
+
4–5%; tactile sensitivity 0.026 vs 0.014) to compensate.</p>
|
| 16 |
+
<p><b>2 · Does the force stage earn its place as an output?</b> This is
|
| 17 |
+
the largest effect in the study. Running the tactile refinement cuts
|
| 18 |
+
open-loop action error <b>almost in half</b>: 0.386 → 0.205 (paired
|
| 19 |
+
−0.1805 ± 0.0077), and on the hand dimensions from 0.371 to
|
| 20 |
+
0.170 (−54%). Critically, variant C’s coarse model — trained without
|
| 21 |
+
any force stage — performs the same as A’s coarse readout
|
| 22 |
+
(0.379 vs 0.386), so the gain is produced by the refinement stage itself,
|
| 23 |
+
not by a training side-effect on the world model. The low-noise segment
|
| 24 |
+
τ ∈ (0, 0.4] is simply better served by a small,
|
| 25 |
+
tactile-conditioned expert than by six more Euler steps of the 5B coarse
|
| 26 |
+
model.</p>
|
| 27 |
+
<p><b>3 · F6 alone vs deformation tactile + F6?</b> Adding ten per-finger
|
| 28 |
+
deformation-image tokens gives a small but consistent further gain:
|
| 29 |
+
open-loop cascade 0.2016 vs 0.2054 (paired −0.0038 ± 0.0015),
|
| 30 |
+
with the force-stage flow MSE edging from 0.643 to 0.639 and training force
|
| 31 |
+
loss from 0.64 to 0.61. The refiner does read the new tokens — they
|
| 32 |
+
receive the same attention mass as each existing tactile stream
|
| 33 |
+
(≈1.6%, vs F6 1.4% and VQ 1.6%) — but coarse memory still dominates
|
| 34 |
+
at ~93%. On a single-episode overfit benchmark the headroom above F6 is
|
| 35 |
+
inherently small; the direction of every metric favors deformation input,
|
| 36 |
+
and the honest read is “cheap to add, consistently non-negative,
|
| 37 |
+
needs multi-episode training to show its real value.”</p>
|
| 38 |
+
<p style="font-size:13.5px;color:var(--muted);">Scope note: all runs overfit
|
| 39 |
+
one 902-frame episode (~537 anchors) for 3,000 steps; the comparisons
|
| 40 |
+
isolate representation and architecture effects under identical budgets,
|
| 41 |
+
not generalization. The 4-GPU 20k-step production run of the full model is
|
| 42 |
+
reported separately above.</p>
|
| 43 |
+
</div>
|
ablation_runs/track_force_cascade.html
CHANGED
|
@@ -1686,8 +1686,10 @@
|
|
| 1686 |
<circle class="vdot v-full" cx="508.5" cy="198.1" r="3.5"></circle>
|
| 1687 |
<path class="vw v-notrack" d="M46.0 30.7 L49.1 30.4 L52.2 32.6 L55.3 37.6 L58.4 43.3 L61.5 49.0 L64.6 54.4 L67.7 59.6 L70.7 64.1 L73.8 68.5 L76.9 72.6 L80.0 76.2 L83.1 79.7 L86.2 82.8 L89.3 85.7 L92.4 88.1 L95.5 90.4 L98.6 92.7 L101.7 94.6 L104.8 96.4 L107.9 98.0 L111.0 99.6 L114.1 101.0 L117.1 102.4 L120.2 103.7 L123.3 104.9 L126.4 106.2 L129.5 107.4 L132.6 108.6 L135.7 110.0 L138.8 110.9 L141.9 111.9 L145.0 113.2 L148.1 114.6 L151.2 115.8 L154.3 117.1 L157.4 118.4 L160.5 120.0 L163.5 121.1 L166.6 122.1 L169.7 123.5 L172.8 125.8 L175.9 127.2 L179.0 128.5 L182.1 130.0 L185.2 131.4 L188.3 132.8 L191.4 134.4 L194.5 135.5 L197.6 136.5 L200.7 138.1 L203.8 139.8 L206.9 141.5 L209.9 142.3 L213.0 143.6 L216.1 145.9 L219.2 148.1 L222.3 149.2 L225.4 151.1 L228.5 152.1 L231.6 152.9 L234.7 154.2 L237.8 155.7 L240.9 157.2 L244.0 159.3 L247.1 160.4 L250.2 162.7 L253.3 165.1 L256.3 166.4 L259.4 167.1 L262.5 168.0 L265.6 169.0 L268.7 169.6 L271.8 171.4 L274.9 172.0 L278.0 172.7 L281.1 174.5 L284.2 175.0 L287.3 176.4 L290.4 175.0 L293.5 176.0 L296.6 176.5 L299.7 177.5 L302.7 178.4 L305.8 180.5 L308.9 181.0 L312.0 182.2 L315.1 182.0 L318.2 184.0 L321.3 183.9 L324.4 182.7 L327.5 183.2 L330.6 183.9 L333.7 185.0 L336.8 186.3 L339.9 187.0 L343.0 188.0 L346.1 188.2 L349.1 186.5 L352.2 187.1 L355.3 187.9 L358.4 187.8 L361.5 189.6 L364.6 189.3 L367.7 192.2 L370.8 191.2 L373.9 191.5 L377.0 192.8 L380.1 193.0 L383.2 194.9 L386.3 195.4 L389.4 196.5 L392.5 196.9 L395.5 197.0 L398.6 194.9 L401.7 195.6 L404.8 195.7 L407.9 195.8 L411.0 196.6 L414.1 197.3 L417.2 196.3 L420.3 198.0 L423.4 197.2 L426.5 197.7 L429.6 197.8 L432.7 197.2 L435.8 196.2 L438.9 197.0 L441.9 195.6 L445.0 195.5 L448.1 194.6 L451.2 194.0 L454.3 193.7 L457.4 194.0 L460.5 194.2 L463.6 195.3 L466.7 194.2 L469.8 193.7 L472.9 192.2 L476.0 194.7 L479.1 194.5 L482.2 194.4 L485.3 196.0 L488.3 196.9 L491.4 194.9 L494.5 197.3 L497.6 200.0 L500.7 199.8 L503.8 199.9 L506.9 198.8 L508.5 198.2"></path>
|
| 1688 |
<circle class="vdot v-notrack" cx="508.5" cy="198.2" r="3.5"></circle>
|
| 1689 |
-
<path class="vw v-noforce" d="M46.0 26.5
|
| 1690 |
-
<circle class="vdot v-noforce" cx="
|
|
|
|
|
|
|
| 1691 |
</svg></div>
|
| 1692 |
<div><svg viewBox="0 0 520 230" role="img" aria-label="dynamics (video flow) MSE">
|
| 1693 |
<text class="th" x="46" y="15">dynamics (video flow) MSE</text>
|
|
@@ -1710,19 +1712,21 @@
|
|
| 1710 |
<circle class="vdot v-full" cx="508.5" cy="187.2" r="3.5"></circle>
|
| 1711 |
<path class="vw v-notrack" d="M46.0 28.4 L49.1 29.7 L52.2 30.3 L55.3 32.5 L58.4 35.4 L61.5 38.5 L64.6 43.1 L67.7 47.8 L70.7 53.1 L73.8 59.4 L76.9 66.2 L80.0 73.0 L83.1 80.1 L86.2 87.1 L89.3 94.4 L92.4 101.1 L95.5 107.6 L98.6 113.1 L101.7 117.8 L104.8 122.6 L107.9 127.3 L111.0 131.2 L114.1 135.5 L117.1 139.6 L120.2 143.0 L123.3 146.5 L126.4 150.4 L129.5 153.2 L132.6 155.2 L135.7 158.0 L138.8 158.5 L141.9 159.2 L145.0 160.8 L148.1 163.4 L151.2 164.5 L154.3 166.7 L157.4 168.3 L160.5 169.4 L163.5 169.8 L166.6 169.6 L169.7 170.2 L172.8 173.7 L175.9 174.5 L179.0 175.2 L182.1 175.1 L185.2 175.9 L188.3 176.4 L191.4 177.2 L194.5 177.1 L197.6 176.4 L200.7 177.3 L203.8 178.1 L206.9 178.1 L209.9 178.1 L213.0 178.2 L216.1 180.4 L219.2 181.8 L222.3 181.4 L225.4 183.0 L228.5 182.8 L231.6 182.4 L234.7 182.6 L237.8 183.6 L240.9 184.7 L244.0 186.5 L247.1 187.0 L250.2 188.7 L253.3 190.4 L256.3 190.6 L259.4 191.0 L262.5 190.9 L265.6 190.8 L268.7 190.5 L271.8 192.1 L274.9 191.2 L278.0 190.8 L281.1 191.9 L284.2 191.9 L287.3 192.5 L290.4 190.3 L293.5 190.5 L296.6 190.5 L299.7 190.8 L302.7 191.3 L305.8 193.0 L308.9 192.8 L312.0 193.9 L315.1 192.7 L318.2 194.6 L321.3 194.2 L324.4 192.5 L327.5 192.5 L330.6 192.8 L333.7 193.5 L336.8 194.1 L339.9 194.2 L343.0 194.7 L346.1 194.7 L349.1 193.1 L352.2 193.3 L355.3 193.9 L358.4 193.6 L361.5 194.6 L364.6 193.7 L367.7 196.1 L370.8 195.1 L373.9 195.4 L377.0 196.2 L380.1 196.5 L383.2 198.2 L386.3 198.5 L389.4 199.1 L392.5 199.1 L395.5 199.4 L398.6 197.6 L401.7 197.8 L404.8 197.6 L407.9 197.2 L411.0 197.9 L414.1 198.4 L417.2 197.6 L420.3 199.4 L423.4 198.6 L426.5 198.6 L429.6 198.5 L432.7 197.4 L435.8 196.6 L438.9 197.4 L441.9 196.3 L445.0 196.4 L448.1 195.3 L451.2 195.0 L454.3 194.8 L457.4 195.0 L460.5 195.3 L463.6 196.2 L466.7 194.8 L469.8 194.1 L472.9 192.5 L476.0 194.7 L479.1 194.5 L482.2 194.2 L485.3 196.0 L488.3 197.3 L491.4 195.2 L494.5 197.2 L497.6 200.0 L500.7 199.8 L503.8 199.8 L506.9 198.7 L508.5 198.1"></path>
|
| 1712 |
<circle class="vdot v-notrack" cx="508.5" cy="198.1" r="3.5"></circle>
|
| 1713 |
-
<path class="vw v-noforce" d="M46.0 26.0
|
| 1714 |
-
<circle class="vdot v-noforce" cx="
|
|
|
|
|
|
|
| 1715 |
</svg></div>
|
| 1716 |
-
<div><svg viewBox="0 0 520 230" role="img" aria-label="track loss (
|
| 1717 |
-
<text class="th" x="46" y="15">track loss (
|
| 1718 |
-
<line class="grid-cell" x1="46" y1="185.
|
| 1719 |
-
<text class="ts" x="40" y="188.
|
| 1720 |
-
<line class="grid-cell" x1="46" y1="
|
| 1721 |
-
<text class="ts" x="40" y="
|
| 1722 |
-
<line class="grid-cell" x1="46" y1="114.
|
| 1723 |
-
<text class="ts" x="40" y="117.
|
| 1724 |
-
<line class="grid-cell" x1="46" y1="84.
|
| 1725 |
-
<text class="ts" x="40" y="87.
|
| 1726 |
<line class="grid-cell" x1="46" y1="43.8" x2="510" y2="43.8"></line>
|
| 1727 |
<text class="ts" x="40" y="46.8" text-anchor="end">5</text>
|
| 1728 |
<line class="grid-cell" x1="46.0" y1="26" x2="46.0" y2="200"></line>
|
|
@@ -1734,15 +1738,17 @@
|
|
| 1734 |
<line class="grid-cell" x1="510.0" y1="26" x2="510.0" y2="200"></line>
|
| 1735 |
<text class="ts" x="510.0" y="214" text-anchor="middle">3k</text>
|
| 1736 |
<text class="ts" x="510" y="224" text-anchor="end">step</text>
|
| 1737 |
-
<path class="vw v-full" d="M46.0 38.6 L49.1 34.8 L52.2 32.1 L55.3 30.1 L58.4 28.2 L61.5 26.9 L64.6 26.0 L67.7 26.6 L70.7 27.8 L73.8 29.5 L76.9 31.8 L80.0 33.
|
| 1738 |
-
<circle class="vdot v-full" cx="508.5" cy="
|
| 1739 |
-
<path class="vw v-noforce" d="M46.0 38.6
|
| 1740 |
-
<circle class="vdot v-noforce" cx="
|
|
|
|
|
|
|
| 1741 |
</svg></div>
|
| 1742 |
<div><svg viewBox="0 0 520 230" role="img" aria-label="force flow MSE">
|
| 1743 |
<text class="th" x="46" y="15">force flow MSE</text>
|
| 1744 |
-
<line class="grid-cell" x1="46" y1="
|
| 1745 |
-
<text class="ts" x="40" y="
|
| 1746 |
<line class="grid-cell" x1="46.0" y1="26" x2="46.0" y2="200"></line>
|
| 1747 |
<text class="ts" x="46.0" y="214" text-anchor="middle">0k</text>
|
| 1748 |
<line class="grid-cell" x1="200.7" y1="26" x2="200.7" y2="200"></line>
|
|
@@ -1752,10 +1758,12 @@
|
|
| 1752 |
<line class="grid-cell" x1="510.0" y1="26" x2="510.0" y2="200"></line>
|
| 1753 |
<text class="ts" x="510.0" y="214" text-anchor="middle">3k</text>
|
| 1754 |
<text class="ts" x="510" y="224" text-anchor="end">step</text>
|
| 1755 |
-
<path class="vw v-full" d="M46.0
|
| 1756 |
-
<circle class="vdot v-full" cx="508.5" cy="
|
| 1757 |
-
<path class="vw v-notrack" d="M46.0
|
| 1758 |
-
<circle class="vdot v-notrack" cx="508.5" cy="167.
|
|
|
|
|
|
|
| 1759 |
</svg></div>
|
| 1760 |
</div>
|
| 1761 |
<figcaption><b>Figure A1 — Same-step training loss convergence.</b>
|
|
@@ -1771,7 +1779,8 @@
|
|
| 1771 |
<th>dynamics</th><th>track</th><th>force</th></tr></thead>
|
| 1772 |
<tbody><tr><td><span class="chip"><span class="vs vs-full"></span> A · full (track+force)</span></td><td class="shape">0.2819</td><td class="shape">0.3325</td><td class="shape">0.1501</td><td class="shape">0.6187</td></tr>
|
| 1773 |
<tr><td><span class="chip"><span class="vs vs-notrack"></span> B · no track</span></td><td class="shape">0.2749</td><td class="shape">0.2837</td><td class="shape">—</td><td class="shape">0.7535</td></tr>
|
| 1774 |
-
<tr><td><span class="chip"><span class="vs vs-noforce"></span> C · no force stage</span></td><td class="shape">
|
|
|
|
| 1775 |
</table>
|
| 1776 |
</div>
|
| 1777 |
<h3>Controlled evaluation on frozen checkpoints</h3>
|
|
@@ -1780,7 +1789,7 @@
|
|
| 1780 |
(<code>scripts/eval/trex_ablation_eval.py</code>). Open-loop: each variant
|
| 1781 |
runs its own full sampling loop from pure noise on 16 anchors and is scored
|
| 1782 |
against the ground-truth normalized delta-base chunk.</p></div>
|
| 1783 |
-
<figure><svg viewBox="0 0 560
|
| 1784 |
<text class="th" x="0" y="14">open-loop action MSE ↓ (16 anchors, normalized 62-D)</text>
|
| 1785 |
<text class="t" x="232" y="39" text-anchor="end">A · cascade (+force stage)</text>
|
| 1786 |
<rect class="v-full" x="240" y="26" width="133.1" height="16" rx="3"></rect>
|
|
@@ -1792,17 +1801,35 @@
|
|
| 1792 |
<rect class="v-notrack" x="240" y="82" width="149.0" height="16" rx="3"></rect>
|
| 1793 |
<text class="ts" x="395.0" y="95">0.230</text>
|
| 1794 |
<text class="t" x="232" y="123" text-anchor="end">B · coarse only</text>
|
| 1795 |
-
<rect class="v-notrack vfaded" x="240" y="110" width="248.
|
| 1796 |
-
<text class="ts" x="494.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1797 |
</svg><figcaption><b>Figure A2 — Open-loop reconstruction.</b> Faded bars: the coarse action read out at τ=0 without tactile refinement (for C this is its only mode). Solid bars: the full cascade.</figcaption></figure>
|
| 1798 |
|
| 1799 |
<div class="tbl-wrap">
|
| 1800 |
<table>
|
| 1801 |
<thead><tr><th>flow MSE @ fixed τ</th><th>τ=0.8</th><th>τ=0.6</th><th>τ=0.4</th><th>τ=0.2</th><th>τ=0.05</th></tr></thead>
|
| 1802 |
-
<tbody><tr><td><span class="chip"><span class="vs vs-full"></span> A · action</span></td><td class="shape">0.139</td><td class="shape">0.204</td><td class="shape">0.399</td><td class="shape">0.752</td><td class="shape">1.076</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span> B · action</span></td><td class="shape">0.124</td><td class="shape">0.191</td><td class="shape">0.388</td><td class="shape">0.745</td><td class="shape">1.078</td></tr><tr><td><span class="chip"><span class="vs vs-full"></span> A · dynamics</span></td><td class="shape">0.215</td><td class="shape">0.280</td><td class="shape">0.414</td><td class="shape">0.724</td><td class="shape">1.131</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span> B · dynamics</span></td><td class="shape">0.178</td><td class="shape">0.229</td><td class="shape">0.341</td><td class="shape">0.628</td><td class="shape">1.064</td></tr></tbody>
|
| 1803 |
</table>
|
| 1804 |
</div>
|
| 1805 |
-
<figure>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1806 |
<text class="th" x="0" y="14">Stage-1: where action queries attend (mass share)</text>
|
| 1807 |
<text class="t" x="102" y="39" text-anchor="end">A · layer 0</text>
|
| 1808 |
<rect class="sg-cond" x="110.0" y="26" width="12.6" height="18" rx="3"></rect>
|
|
@@ -1841,8 +1868,59 @@
|
|
| 1841 |
<rect class="sg-act" x="290.0" y="110" width="331.7" height="18" rx="3"></rect>
|
| 1842 |
<text class="ts" x="456.9" y="123" text-anchor="middle">64%</text>
|
| 1843 |
<rect class="sg-st" x="623.7" y="110" width="2.3" height="18" rx="3"></rect>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1844 |
</svg><figcaption><b>Figure A3 — Stage-1 attention routing.</b> Share of self-attention mass from action queries onto each token group at τ=0.4 (mean over heads/blocks).</figcaption></figure>
|
| 1845 |
-
<figure>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1846 |
<text class="th" x="0" y="14">Stage-2: where action queries attend (mass share)</text>
|
| 1847 |
<text class="t" x="102" y="39" text-anchor="end">A · L0–L5 mean</text>
|
| 1848 |
<rect class="sg-act" x="110.0" y="26" width="12.3" height="18" rx="3"></rect>
|
|
@@ -1859,6 +1937,13 @@
|
|
| 1859 |
<text class="ts" x="295.1" y="67" text-anchor="middle">10%</text>
|
| 1860 |
<rect class="sg-mem" x="321.7" y="54" width="304.3" height="18" rx="3"></rect>
|
| 1861 |
<text class="ts" x="474.8" y="67" text-anchor="middle">59%</text>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1862 |
</svg><figcaption><b>Figure A4 — Stage-2 attention routing.</b> Action-query mass over F6 / VQ-history / deform / coarse-memory tokens during real cascaded refinement.</figcaption></figure>
|
| 1863 |
|
| 1864 |
<div class="tbl-wrap">
|
|
@@ -1866,9 +1951,52 @@
|
|
| 1866 |
<thead><tr><th>Tactile sensitivity (normalized units)</th>
|
| 1867 |
<th>mean |Δa| tactile vs masked</th><th>max |Δa|</th>
|
| 1868 |
<th>mean |refinement|</th></tr></thead>
|
| 1869 |
-
<tbody><tr><td><span class="chip"><span class="vs vs-full"></span> A · full (track+force)</span></td><td class="shape">0.0141</td><td class="shape">0.0664</td><td class="shape">0.2452</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span> B · no track</span></td><td class="shape">0.0256</td><td class="shape">0.1309</td><td class="shape">0.2472</td></tr></tbody>
|
| 1870 |
</table>
|
| 1871 |
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1872 |
|
| 1873 |
</section>
|
| 1874 |
</div>
|
|
|
|
| 1686 |
<circle class="vdot v-full" cx="508.5" cy="198.1" r="3.5"></circle>
|
| 1687 |
<path class="vw v-notrack" d="M46.0 30.7 L49.1 30.4 L52.2 32.6 L55.3 37.6 L58.4 43.3 L61.5 49.0 L64.6 54.4 L67.7 59.6 L70.7 64.1 L73.8 68.5 L76.9 72.6 L80.0 76.2 L83.1 79.7 L86.2 82.8 L89.3 85.7 L92.4 88.1 L95.5 90.4 L98.6 92.7 L101.7 94.6 L104.8 96.4 L107.9 98.0 L111.0 99.6 L114.1 101.0 L117.1 102.4 L120.2 103.7 L123.3 104.9 L126.4 106.2 L129.5 107.4 L132.6 108.6 L135.7 110.0 L138.8 110.9 L141.9 111.9 L145.0 113.2 L148.1 114.6 L151.2 115.8 L154.3 117.1 L157.4 118.4 L160.5 120.0 L163.5 121.1 L166.6 122.1 L169.7 123.5 L172.8 125.8 L175.9 127.2 L179.0 128.5 L182.1 130.0 L185.2 131.4 L188.3 132.8 L191.4 134.4 L194.5 135.5 L197.6 136.5 L200.7 138.1 L203.8 139.8 L206.9 141.5 L209.9 142.3 L213.0 143.6 L216.1 145.9 L219.2 148.1 L222.3 149.2 L225.4 151.1 L228.5 152.1 L231.6 152.9 L234.7 154.2 L237.8 155.7 L240.9 157.2 L244.0 159.3 L247.1 160.4 L250.2 162.7 L253.3 165.1 L256.3 166.4 L259.4 167.1 L262.5 168.0 L265.6 169.0 L268.7 169.6 L271.8 171.4 L274.9 172.0 L278.0 172.7 L281.1 174.5 L284.2 175.0 L287.3 176.4 L290.4 175.0 L293.5 176.0 L296.6 176.5 L299.7 177.5 L302.7 178.4 L305.8 180.5 L308.9 181.0 L312.0 182.2 L315.1 182.0 L318.2 184.0 L321.3 183.9 L324.4 182.7 L327.5 183.2 L330.6 183.9 L333.7 185.0 L336.8 186.3 L339.9 187.0 L343.0 188.0 L346.1 188.2 L349.1 186.5 L352.2 187.1 L355.3 187.9 L358.4 187.8 L361.5 189.6 L364.6 189.3 L367.7 192.2 L370.8 191.2 L373.9 191.5 L377.0 192.8 L380.1 193.0 L383.2 194.9 L386.3 195.4 L389.4 196.5 L392.5 196.9 L395.5 197.0 L398.6 194.9 L401.7 195.6 L404.8 195.7 L407.9 195.8 L411.0 196.6 L414.1 197.3 L417.2 196.3 L420.3 198.0 L423.4 197.2 L426.5 197.7 L429.6 197.8 L432.7 197.2 L435.8 196.2 L438.9 197.0 L441.9 195.6 L445.0 195.5 L448.1 194.6 L451.2 194.0 L454.3 193.7 L457.4 194.0 L460.5 194.2 L463.6 195.3 L466.7 194.2 L469.8 193.7 L472.9 192.2 L476.0 194.7 L479.1 194.5 L482.2 194.4 L485.3 196.0 L488.3 196.9 L491.4 194.9 L494.5 197.3 L497.6 200.0 L500.7 199.8 L503.8 199.9 L506.9 198.8 L508.5 198.2"></path>
|
| 1688 |
<circle class="vdot v-notrack" cx="508.5" cy="198.2" r="3.5"></circle>
|
| 1689 |
+
<path class="vw v-noforce" d="M46.0 26.5 L49.1 26.2 L52.2 27.9 L55.3 32.7 L58.4 38.5 L61.5 44.3 L64.6 49.9 L67.7 55.5 L70.7 60.2 L73.8 64.7 L76.9 69.0 L80.0 72.9 L83.1 76.8 L86.2 80.0 L89.3 83.0 L92.4 85.8 L95.5 88.4 L98.6 90.8 L101.7 93.0 L104.8 94.9 L107.9 96.7 L111.0 98.4 L114.1 99.7 L117.1 101.2 L120.2 102.6 L123.3 103.6 L126.4 104.9 L129.5 106.1 L132.6 107.2 L135.7 108.5 L138.8 109.8 L141.9 111.2 L145.0 112.3 L148.1 113.4 L151.2 114.5 L154.3 115.5 L157.4 117.0 L160.5 118.6 L163.5 119.9 L166.6 121.5 L169.7 123.2 L172.8 124.6 L175.9 126.0 L179.0 127.3 L182.1 128.7 L185.2 130.1 L188.3 131.7 L191.4 133.5 L194.5 135.7 L197.6 137.3 L200.7 139.0 L203.8 139.9 L206.9 141.5 L209.9 142.8 L213.0 144.6 L216.1 146.6 L219.2 147.9 L222.3 150.0 L225.4 150.8 L228.5 152.8 L231.6 153.7 L234.7 154.7 L237.8 154.8 L240.9 156.1 L244.0 158.0 L247.1 159.8 L250.2 160.2 L253.3 161.5 L256.3 163.1 L259.4 164.3 L262.5 165.0 L265.6 166.1 L268.7 167.3 L271.8 168.1 L274.9 169.0 L278.0 169.9 L281.1 171.4 L284.2 172.7 L287.3 173.6 L290.4 173.5 L293.5 175.0 L296.6 176.4 L299.7 177.2 L302.7 179.3 L305.8 179.6 L308.9 179.5 L312.0 180.7 L315.1 181.2 L318.2 181.7 L321.3 181.9 L324.4 182.8 L327.5 182.2 L330.6 182.5 L333.7 182.7 L336.8 182.3 L339.9 183.4 L343.0 184.1 L346.1 183.6 L349.1 181.5 L352.2 184.5 L355.3 186.2 L358.4 188.5 L361.5 187.2 L364.6 188.5 L367.7 188.3 L370.8 188.5 L373.9 187.7 L377.0 188.8 L380.1 189.2 L383.2 190.8 L386.3 187.8 L389.4 188.3 L392.5 190.5 L395.5 190.7 L398.6 189.1 L401.7 189.5 L404.8 190.1 L407.9 189.7 L411.0 191.7 L414.1 191.1 L417.2 192.2 L420.3 192.9 L423.4 192.7 L426.5 191.2 L429.6 192.5 L432.7 195.1 L435.8 194.7 L438.9 195.2 L441.9 195.2 L445.0 192.8 L448.1 194.7 L451.2 196.7 L454.3 197.7 L457.4 197.0 L460.5 196.3 L463.6 196.0 L466.7 194.8 L469.8 193.5 L472.9 192.8 L476.0 195.6 L479.1 195.8 L482.2 195.0 L485.3 194.8 L488.3 194.7 L491.4 194.9 L494.5 196.1 L497.6 193.6 L500.7 193.4 L503.8 192.9 L506.9 193.5 L508.5 194.2"></path>
|
| 1690 |
+
<circle class="vdot v-noforce" cx="508.5" cy="194.2" r="3.5"></circle>
|
| 1691 |
+
<path class="vw v-deform" d="M46.0 26.5 L49.1 26.0 L52.2 28.0 L55.3 32.8 L58.4 38.5 L61.5 44.3 L64.6 49.8 L67.7 55.2 L70.7 60.0 L73.8 64.6 L76.9 68.9 L80.0 72.8 L83.1 76.5 L86.2 79.8 L89.3 82.9 L92.4 85.5 L95.5 88.1 L98.6 90.6 L101.7 92.8 L104.8 94.7 L107.9 96.4 L111.0 98.0 L114.1 99.5 L117.1 100.9 L120.2 102.3 L123.3 103.5 L126.4 104.7 L129.5 105.7 L132.6 106.7 L135.7 107.9 L138.8 109.3 L141.9 110.7 L145.0 111.5 L148.1 112.8 L151.2 114.4 L154.3 115.9 L157.4 117.0 L160.5 118.1 L163.5 119.4 L166.6 121.0 L169.7 122.4 L172.8 123.6 L175.9 125.1 L179.0 126.3 L182.1 127.9 L185.2 129.3 L188.3 130.7 L191.4 131.8 L194.5 133.8 L197.6 135.4 L200.7 137.4 L203.8 139.6 L206.9 141.0 L209.9 141.9 L213.0 144.2 L216.1 146.3 L219.2 148.0 L222.3 149.6 L225.4 150.7 L228.5 152.2 L231.6 153.9 L234.7 154.9 L237.8 156.0 L240.9 157.7 L244.0 158.8 L247.1 160.2 L250.2 161.6 L253.3 161.2 L256.3 161.3 L259.4 161.8 L262.5 163.1 L265.6 164.4 L268.7 165.5 L271.8 165.7 L274.9 166.6 L278.0 168.3 L281.1 170.0 L284.2 170.6 L287.3 171.0 L290.4 172.3 L293.5 172.1 L296.6 172.3 L299.7 173.2 L302.7 174.4 L305.8 176.2 L308.9 177.2 L312.0 177.6 L315.1 177.9 L318.2 179.1 L321.3 180.0 L324.4 181.0 L327.5 181.6 L330.6 181.8 L333.7 183.7 L336.8 185.4 L339.9 184.7 L343.0 183.3 L346.1 183.8 L349.1 182.9 L352.2 184.5 L355.3 184.1 L358.4 184.8 L361.5 184.1 L364.6 184.3 L367.7 184.6 L370.8 186.9 L373.9 187.2 L377.0 188.9 L380.1 189.1 L383.2 187.5 L386.3 188.4 L389.4 189.2 L392.5 190.8 L395.5 191.1 L398.6 189.6 L401.7 187.7 L404.8 188.8 L407.9 190.4 L411.0 190.8 L414.1 191.7 L417.2 191.5 L420.3 189.8 L423.4 190.5 L426.5 191.2 L429.6 191.8 L432.7 192.7 L435.8 193.2 L438.9 193.9 L441.9 192.9 L445.0 194.1 L448.1 195.0 L451.2 195.3 L454.3 195.7 L457.4 196.2 L460.5 197.1 L463.6 195.4 L466.7 195.6 L469.8 197.1 L472.9 196.7 L476.0 196.8 L479.1 196.5 L482.2 196.1 L485.3 196.2 L488.3 195.9 L491.4 197.2 L494.5 197.2 L497.6 196.9 L500.7 194.4 L503.8 196.5 L506.9 199.6 L508.5 198.2"></path>
|
| 1692 |
+
<circle class="vdot v-deform" cx="508.5" cy="198.2" r="3.5"></circle>
|
| 1693 |
</svg></div>
|
| 1694 |
<div><svg viewBox="0 0 520 230" role="img" aria-label="dynamics (video flow) MSE">
|
| 1695 |
<text class="th" x="46" y="15">dynamics (video flow) MSE</text>
|
|
|
|
| 1712 |
<circle class="vdot v-full" cx="508.5" cy="187.2" r="3.5"></circle>
|
| 1713 |
<path class="vw v-notrack" d="M46.0 28.4 L49.1 29.7 L52.2 30.3 L55.3 32.5 L58.4 35.4 L61.5 38.5 L64.6 43.1 L67.7 47.8 L70.7 53.1 L73.8 59.4 L76.9 66.2 L80.0 73.0 L83.1 80.1 L86.2 87.1 L89.3 94.4 L92.4 101.1 L95.5 107.6 L98.6 113.1 L101.7 117.8 L104.8 122.6 L107.9 127.3 L111.0 131.2 L114.1 135.5 L117.1 139.6 L120.2 143.0 L123.3 146.5 L126.4 150.4 L129.5 153.2 L132.6 155.2 L135.7 158.0 L138.8 158.5 L141.9 159.2 L145.0 160.8 L148.1 163.4 L151.2 164.5 L154.3 166.7 L157.4 168.3 L160.5 169.4 L163.5 169.8 L166.6 169.6 L169.7 170.2 L172.8 173.7 L175.9 174.5 L179.0 175.2 L182.1 175.1 L185.2 175.9 L188.3 176.4 L191.4 177.2 L194.5 177.1 L197.6 176.4 L200.7 177.3 L203.8 178.1 L206.9 178.1 L209.9 178.1 L213.0 178.2 L216.1 180.4 L219.2 181.8 L222.3 181.4 L225.4 183.0 L228.5 182.8 L231.6 182.4 L234.7 182.6 L237.8 183.6 L240.9 184.7 L244.0 186.5 L247.1 187.0 L250.2 188.7 L253.3 190.4 L256.3 190.6 L259.4 191.0 L262.5 190.9 L265.6 190.8 L268.7 190.5 L271.8 192.1 L274.9 191.2 L278.0 190.8 L281.1 191.9 L284.2 191.9 L287.3 192.5 L290.4 190.3 L293.5 190.5 L296.6 190.5 L299.7 190.8 L302.7 191.3 L305.8 193.0 L308.9 192.8 L312.0 193.9 L315.1 192.7 L318.2 194.6 L321.3 194.2 L324.4 192.5 L327.5 192.5 L330.6 192.8 L333.7 193.5 L336.8 194.1 L339.9 194.2 L343.0 194.7 L346.1 194.7 L349.1 193.1 L352.2 193.3 L355.3 193.9 L358.4 193.6 L361.5 194.6 L364.6 193.7 L367.7 196.1 L370.8 195.1 L373.9 195.4 L377.0 196.2 L380.1 196.5 L383.2 198.2 L386.3 198.5 L389.4 199.1 L392.5 199.1 L395.5 199.4 L398.6 197.6 L401.7 197.8 L404.8 197.6 L407.9 197.2 L411.0 197.9 L414.1 198.4 L417.2 197.6 L420.3 199.4 L423.4 198.6 L426.5 198.6 L429.6 198.5 L432.7 197.4 L435.8 196.6 L438.9 197.4 L441.9 196.3 L445.0 196.4 L448.1 195.3 L451.2 195.0 L454.3 194.8 L457.4 195.0 L460.5 195.3 L463.6 196.2 L466.7 194.8 L469.8 194.1 L472.9 192.5 L476.0 194.7 L479.1 194.5 L482.2 194.2 L485.3 196.0 L488.3 197.3 L491.4 195.2 L494.5 197.2 L497.6 200.0 L500.7 199.8 L503.8 199.8 L506.9 198.7 L508.5 198.1"></path>
|
| 1714 |
<circle class="vdot v-notrack" cx="508.5" cy="198.1" r="3.5"></circle>
|
| 1715 |
+
<path class="vw v-noforce" d="M46.0 26.0 L49.1 26.9 L52.2 27.9 L55.3 29.9 L58.4 32.4 L61.5 35.1 L64.6 38.6 L67.7 42.5 L70.7 46.4 L73.8 51.0 L76.9 56.5 L80.0 62.9 L83.1 70.1 L86.2 77.1 L89.3 83.5 L92.4 90.3 L95.5 96.4 L98.6 102.5 L101.7 108.4 L104.8 114.6 L107.9 118.7 L111.0 123.2 L114.1 127.2 L117.1 131.5 L120.2 134.9 L123.3 137.3 L126.4 141.5 L129.5 144.0 L132.6 147.0 L135.7 148.9 L138.8 151.6 L141.9 153.9 L145.0 155.6 L148.1 157.6 L151.2 159.1 L154.3 159.3 L157.4 161.1 L160.5 162.6 L163.5 163.8 L166.6 165.6 L169.7 167.5 L172.8 167.6 L175.9 167.7 L179.0 167.3 L182.1 167.2 L185.2 167.2 L188.3 167.4 L191.4 168.3 L194.5 170.8 L197.6 171.0 L200.7 172.3 L203.8 171.3 L206.9 171.6 L209.9 171.3 L213.0 171.9 L216.1 173.4 L219.2 173.1 L222.3 173.9 L225.4 173.7 L228.5 174.8 L231.6 174.1 L234.7 173.6 L237.8 172.0 L240.9 172.4 L244.0 173.5 L247.1 175.2 L250.2 174.7 L253.3 175.1 L256.3 176.2 L259.4 177.1 L262.5 176.6 L265.6 176.5 L268.7 177.0 L271.8 177.0 L274.9 176.9 L278.0 176.9 L281.1 177.6 L284.2 178.8 L287.3 179.1 L290.4 178.2 L293.5 179.7 L296.6 180.8 L299.7 180.9 L302.7 182.0 L305.8 181.9 L308.9 181.1 L312.0 181.9 L315.1 181.7 L318.2 181.6 L321.3 181.5 L324.4 181.5 L327.5 180.4 L330.6 180.4 L333.7 180.0 L336.8 179.3 L339.9 180.3 L343.0 180.5 L346.1 180.1 L349.1 177.8 L352.2 180.3 L355.3 181.5 L358.4 183.1 L361.5 181.5 L364.6 182.0 L367.7 182.0 L370.8 182.0 L373.9 181.2 L377.0 181.9 L380.1 182.2 L383.2 183.6 L386.3 180.5 L389.4 180.7 L392.5 182.6 L395.5 182.6 L398.6 181.3 L401.7 181.3 L404.8 181.6 L407.9 180.8 L411.0 182.3 L414.1 181.6 L417.2 182.6 L420.3 183.4 L423.4 183.3 L426.5 181.8 L429.6 182.6 L432.7 184.5 L435.8 184.1 L438.9 184.5 L441.9 184.7 L445.0 183.1 L448.1 184.4 L451.2 186.1 L454.3 186.9 L457.4 186.3 L460.5 185.8 L463.6 185.5 L466.7 184.0 L469.8 183.0 L472.9 182.4 L476.0 184.5 L479.1 184.6 L482.2 183.9 L485.3 184.0 L488.3 184.3 L491.4 184.4 L494.5 185.4 L497.6 183.4 L500.7 183.4 L503.8 182.6 L506.9 183.0 L508.5 183.6"></path>
|
| 1716 |
+
<circle class="vdot v-noforce" cx="508.5" cy="183.6" r="3.5"></circle>
|
| 1717 |
+
<path class="vw v-deform" d="M46.0 26.0 L49.1 26.7 L52.2 27.7 L55.3 29.9 L58.4 32.4 L61.5 35.6 L64.6 39.2 L67.7 42.6 L70.7 46.5 L73.8 51.2 L76.9 56.4 L80.0 62.8 L83.1 70.2 L86.2 76.9 L89.3 83.8 L92.4 90.2 L95.5 96.4 L98.6 102.2 L101.7 107.8 L104.8 113.8 L107.9 117.9 L111.0 121.4 L114.1 125.7 L117.1 129.7 L120.2 133.8 L123.3 136.2 L126.4 139.4 L129.5 141.7 L132.6 144.4 L135.7 146.6 L138.8 149.2 L141.9 152.1 L145.0 152.3 L148.1 154.4 L151.2 156.8 L154.3 159.6 L157.4 159.8 L160.5 160.6 L163.5 162.0 L166.6 163.5 L169.7 164.5 L172.8 164.5 L175.9 165.2 L179.0 165.2 L182.1 166.1 L185.2 166.6 L188.3 167.2 L191.4 166.8 L194.5 168.8 L197.6 169.3 L200.7 170.9 L203.8 172.4 L206.9 172.5 L209.9 171.9 L213.0 173.9 L216.1 175.4 L219.2 175.6 L222.3 176.1 L225.4 175.7 L228.5 176.0 L231.6 177.0 L234.7 176.8 L237.8 176.9 L240.9 178.1 L244.0 178.1 L247.1 178.7 L250.2 179.3 L253.3 177.3 L256.3 175.8 L259.4 175.9 L262.5 176.1 L265.6 176.2 L268.7 176.7 L271.8 176.2 L274.9 176.0 L278.0 177.0 L281.1 177.6 L284.2 178.0 L287.3 177.5 L290.4 178.4 L293.5 177.9 L296.6 177.5 L299.7 178.0 L302.7 178.4 L305.8 179.7 L308.9 180.3 L312.0 180.7 L315.1 180.1 L318.2 180.9 L321.3 181.5 L324.4 181.9 L327.5 181.9 L330.6 181.7 L333.7 183.2 L336.8 184.3 L339.9 183.2 L343.0 181.2 L346.1 181.6 L349.1 180.5 L352.2 181.6 L355.3 180.8 L358.4 181.2 L361.5 180.0 L364.6 179.8 L367.7 179.7 L370.8 181.8 L373.9 181.7 L377.0 182.9 L380.1 182.9 L383.2 181.5 L386.3 182.2 L389.4 182.8 L392.5 184.0 L395.5 184.2 L398.6 182.7 L401.7 180.6 L404.8 181.3 L407.9 182.3 L411.0 182.7 L414.1 183.3 L417.2 183.3 L420.3 181.8 L423.4 182.1 L426.5 182.7 L429.6 183.0 L432.7 183.8 L435.8 184.4 L438.9 184.6 L441.9 183.5 L445.0 184.9 L448.1 185.5 L451.2 185.8 L454.3 186.1 L457.4 186.4 L460.5 187.2 L463.6 185.7 L466.7 185.5 L469.8 186.8 L472.9 186.4 L476.0 186.4 L479.1 186.0 L482.2 185.6 L485.3 185.9 L488.3 185.9 L491.4 186.9 L494.5 186.6 L497.6 186.7 L500.7 184.7 L503.8 186.1 L506.9 188.4 L508.5 187.4"></path>
|
| 1718 |
+
<circle class="vdot v-deform" cx="508.5" cy="187.4" r="3.5"></circle>
|
| 1719 |
</svg></div>
|
| 1720 |
+
<div><svg viewBox="0 0 520 230" role="img" aria-label="track loss (flow + position + smoothness)">
|
| 1721 |
+
<text class="th" x="46" y="15">track loss (flow + position + smoothness)</text>
|
| 1722 |
+
<line class="grid-cell" x1="46" y1="185.2" x2="510" y2="185.2"></line>
|
| 1723 |
+
<text class="ts" x="40" y="188.2" text-anchor="end">0.2</text>
|
| 1724 |
+
<line class="grid-cell" x1="46" y1="144.9" x2="510" y2="144.9"></line>
|
| 1725 |
+
<text class="ts" x="40" y="147.9" text-anchor="end">0.5</text>
|
| 1726 |
+
<line class="grid-cell" x1="46" y1="114.5" x2="510" y2="114.5"></line>
|
| 1727 |
+
<text class="ts" x="40" y="117.5" text-anchor="end">1</text>
|
| 1728 |
+
<line class="grid-cell" x1="46" y1="84.0" x2="510" y2="84.0"></line>
|
| 1729 |
+
<text class="ts" x="40" y="87.0" text-anchor="end">2</text>
|
| 1730 |
<line class="grid-cell" x1="46" y1="43.8" x2="510" y2="43.8"></line>
|
| 1731 |
<text class="ts" x="40" y="46.8" text-anchor="end">5</text>
|
| 1732 |
<line class="grid-cell" x1="46.0" y1="26" x2="46.0" y2="200"></line>
|
|
|
|
| 1738 |
<line class="grid-cell" x1="510.0" y1="26" x2="510.0" y2="200"></line>
|
| 1739 |
<text class="ts" x="510.0" y="214" text-anchor="middle">3k</text>
|
| 1740 |
<text class="ts" x="510" y="224" text-anchor="end">step</text>
|
| 1741 |
+
<path class="vw v-full" d="M46.0 38.6 L49.1 34.8 L52.2 32.1 L55.3 30.1 L58.4 28.2 L61.5 26.9 L64.6 26.0 L67.7 26.6 L70.7 27.8 L73.8 29.5 L76.9 31.8 L80.0 33.5 L83.1 35.1 L86.2 37.4 L89.3 39.5 L92.4 42.0 L95.5 44.6 L98.6 47.2 L101.7 49.6 L104.8 52.0 L107.9 55.1 L111.0 58.2 L114.1 61.1 L117.1 63.9 L120.2 66.8 L123.3 69.7 L126.4 72.6 L129.5 75.5 L132.6 78.5 L135.7 81.2 L138.8 83.8 L141.9 86.4 L145.0 89.2 L148.1 91.8 L151.2 94.3 L154.3 96.7 L157.4 99.2 L160.5 101.7 L163.5 104.3 L166.6 106.8 L169.7 109.3 L172.8 111.5 L175.9 113.5 L179.0 115.8 L182.1 118.1 L185.2 120.5 L188.3 122.6 L191.4 124.8 L194.5 127.1 L197.6 129.5 L200.7 131.6 L203.8 134.1 L206.9 136.3 L209.9 138.0 L213.0 140.2 L216.1 142.1 L219.2 143.9 L222.3 145.9 L225.4 147.3 L228.5 149.0 L231.6 150.7 L234.7 152.2 L237.8 153.9 L240.9 155.8 L244.0 157.3 L247.1 158.5 L250.2 159.9 L253.3 160.1 L256.3 160.7 L259.4 161.8 L262.5 163.2 L265.6 164.1 L268.7 165.4 L271.8 165.7 L274.9 166.6 L278.0 168.3 L281.1 169.5 L284.2 170.6 L287.3 171.1 L290.4 172.3 L293.5 173.1 L296.6 173.0 L299.7 174.5 L302.7 175.9 L305.8 177.5 L308.9 178.2 L312.0 178.9 L315.1 179.5 L318.2 180.8 L321.3 181.7 L324.4 182.7 L327.5 183.2 L330.6 183.8 L333.7 185.0 L336.8 186.3 L339.9 185.8 L343.0 184.6 L346.1 186.0 L349.1 186.0 L352.2 187.3 L355.3 187.7 L358.4 188.5 L361.5 188.7 L364.6 189.4 L367.7 189.7 L370.8 190.9 L373.9 190.6 L377.0 192.3 L380.1 192.5 L383.2 191.0 L386.3 192.5 L389.4 193.7 L392.5 195.0 L395.5 195.0 L398.6 193.8 L401.7 192.7 L404.8 194.0 L407.9 195.0 L411.0 195.1 L414.1 195.8 L417.2 194.7 L420.3 193.1 L423.4 193.5 L426.5 194.7 L429.6 194.7 L432.7 196.4 L435.8 196.3 L438.9 196.8 L441.9 194.8 L445.0 196.3 L448.1 197.7 L451.2 197.3 L454.3 196.8 L457.4 197.0 L460.5 198.5 L463.6 197.6 L466.7 197.1 L469.8 199.0 L472.9 198.8 L476.0 198.8 L479.1 198.8 L482.2 198.4 L485.3 197.7 L488.3 198.2 L491.4 198.9 L494.5 199.0 L497.6 198.9 L500.7 196.5 L503.8 198.1 L506.9 199.7 L508.5 197.9"></path>
|
| 1742 |
+
<circle class="vdot v-full" cx="508.5" cy="197.9" r="3.5"></circle>
|
| 1743 |
+
<path class="vw v-noforce" d="M46.0 38.6 L49.1 34.3 L52.2 31.6 L55.3 29.7 L58.4 28.5 L61.5 27.7 L64.6 27.1 L67.7 27.1 L70.7 27.9 L73.8 29.7 L76.9 31.3 L80.0 33.5 L83.1 35.5 L86.2 37.7 L89.3 40.4 L92.4 42.6 L95.5 45.2 L98.6 47.7 L101.7 50.2 L104.8 52.3 L107.9 55.4 L111.0 58.2 L114.1 61.1 L117.1 64.0 L120.2 67.0 L123.3 69.9 L126.4 72.7 L129.5 75.6 L132.6 78.4 L135.7 81.0 L138.8 83.7 L141.9 86.1 L145.0 88.8 L148.1 91.7 L151.2 94.3 L154.3 96.8 L157.4 99.3 L160.5 101.8 L163.5 104.3 L166.6 107.0 L169.7 109.5 L172.8 111.7 L175.9 114.0 L179.0 116.4 L182.1 118.7 L185.2 120.9 L188.3 122.8 L191.4 125.2 L194.5 127.7 L197.6 129.9 L200.7 132.4 L203.8 134.1 L206.9 136.2 L209.9 137.9 L213.0 139.8 L216.1 141.8 L219.2 143.5 L222.3 145.6 L225.4 146.9 L228.5 148.9 L231.6 150.0 L234.7 151.3 L237.8 152.1 L240.9 153.7 L244.0 155.4 L247.1 157.4 L250.2 158.7 L253.3 159.8 L256.3 161.9 L259.4 163.4 L262.5 164.5 L265.6 165.4 L268.7 166.4 L271.8 167.5 L274.9 168.4 L278.0 169.2 L281.1 170.6 L284.2 172.6 L287.3 173.8 L290.4 173.8 L293.5 176.0 L296.6 176.7 L299.7 177.8 L302.7 179.6 L305.8 180.5 L308.9 180.2 L312.0 181.4 L315.1 182.3 L318.2 182.4 L321.3 183.1 L324.4 183.7 L327.5 183.1 L330.6 183.7 L333.7 184.1 L336.8 184.1 L339.9 185.1 L343.0 185.7 L346.1 187.1 L349.1 186.1 L352.2 187.3 L355.3 189.0 L358.4 190.7 L361.5 190.5 L364.6 191.7 L367.7 192.5 L370.8 192.1 L373.9 191.0 L377.0 191.4 L380.1 191.8 L383.2 193.6 L386.3 190.0 L389.4 191.2 L392.5 192.9 L395.5 192.5 L398.6 192.5 L401.7 192.8 L404.8 193.4 L407.9 192.4 L411.0 194.6 L414.1 194.9 L417.2 195.5 L420.3 196.1 L423.4 196.1 L426.5 194.6 L429.6 194.9 L432.7 197.2 L435.8 197.0 L438.9 197.1 L441.9 197.7 L445.0 196.8 L448.1 197.8 L451.2 199.2 L454.3 199.0 L457.4 199.0 L460.5 198.6 L463.6 199.0 L466.7 197.3 L469.8 197.6 L472.9 196.9 L476.0 198.5 L479.1 198.4 L482.2 198.4 L485.3 198.0 L488.3 199.0 L491.4 199.1 L494.5 200.0 L497.6 196.9 L500.7 197.1 L503.8 196.1 L506.9 196.8 L508.5 197.6"></path>
|
| 1744 |
+
<circle class="vdot v-noforce" cx="508.5" cy="197.6" r="3.5"></circle>
|
| 1745 |
+
<path class="vw v-deform" d="M46.0 38.6 L49.1 34.8 L52.2 32.1 L55.3 30.1 L58.4 28.2 L61.5 26.9 L64.6 26.0 L67.7 26.6 L70.7 27.8 L73.8 29.5 L76.9 31.8 L80.0 33.5 L83.1 35.0 L86.2 37.3 L89.3 39.5 L92.4 41.9 L95.5 44.6 L98.6 47.1 L101.7 49.5 L104.8 52.0 L107.9 55.0 L111.0 58.1 L114.1 61.0 L117.1 63.8 L120.2 66.6 L123.3 69.6 L126.4 72.5 L129.5 75.4 L132.6 78.4 L135.7 81.1 L138.8 83.7 L141.9 86.3 L145.0 89.1 L148.1 91.7 L151.2 94.2 L154.3 96.7 L157.4 99.1 L160.5 101.7 L163.5 104.2 L166.6 106.7 L169.7 109.2 L172.8 111.4 L175.9 113.4 L179.0 115.7 L182.1 118.1 L185.2 120.4 L188.3 122.5 L191.4 124.7 L194.5 127.1 L197.6 129.5 L200.7 131.5 L203.8 134.0 L206.9 136.2 L209.9 137.9 L213.0 140.1 L216.1 142.0 L219.2 143.8 L222.3 145.8 L225.4 147.2 L228.5 148.9 L231.6 150.6 L234.7 152.1 L237.8 153.8 L240.9 155.6 L244.0 157.1 L247.1 158.4 L250.2 159.7 L253.3 160.0 L256.3 160.5 L259.4 161.6 L262.5 163.0 L265.6 163.9 L268.7 165.1 L271.8 165.4 L274.9 166.3 L278.0 168.0 L281.1 169.1 L284.2 170.3 L287.3 170.7 L290.4 171.9 L293.5 172.7 L296.6 172.6 L299.7 174.1 L302.7 175.5 L305.8 177.1 L308.9 177.8 L312.0 178.4 L315.1 179.1 L318.2 180.3 L321.3 181.2 L324.4 182.2 L327.5 182.8 L330.6 183.3 L333.7 184.5 L336.8 185.8 L339.9 185.4 L343.0 184.1 L346.1 185.6 L349.1 185.6 L352.2 186.9 L355.3 187.3 L358.4 188.1 L361.5 188.2 L364.6 189.0 L367.7 189.2 L370.8 190.5 L373.9 190.2 L377.0 191.8 L380.1 192.1 L383.2 190.6 L386.3 192.1 L389.4 193.3 L392.5 194.6 L395.5 194.6 L398.6 193.4 L401.7 192.4 L404.8 193.6 L407.9 194.6 L411.0 194.7 L414.1 195.4 L417.2 194.3 L420.3 192.7 L423.4 193.1 L426.5 194.4 L429.6 194.4 L432.7 196.0 L435.8 195.9 L438.9 196.4 L441.9 194.5 L445.0 195.9 L448.1 197.3 L451.2 196.9 L454.3 196.4 L457.4 196.6 L460.5 198.1 L463.6 197.2 L466.7 196.7 L469.8 198.7 L472.9 198.4 L476.0 198.4 L479.1 198.5 L482.2 198.0 L485.3 197.4 L488.3 197.8 L491.4 198.6 L494.5 198.6 L497.6 198.5 L500.7 196.1 L503.8 197.7 L506.9 199.3 L508.5 197.5"></path>
|
| 1746 |
+
<circle class="vdot v-deform" cx="508.5" cy="197.5" r="3.5"></circle>
|
| 1747 |
</svg></div>
|
| 1748 |
<div><svg viewBox="0 0 520 230" role="img" aria-label="force flow MSE">
|
| 1749 |
<text class="th" x="46" y="15">force flow MSE</text>
|
| 1750 |
+
<line class="grid-cell" x1="46" y1="122.3" x2="510" y2="122.3"></line>
|
| 1751 |
+
<text class="ts" x="40" y="125.3" text-anchor="end">1</text>
|
| 1752 |
<line class="grid-cell" x1="46.0" y1="26" x2="46.0" y2="200"></line>
|
| 1753 |
<text class="ts" x="46.0" y="214" text-anchor="middle">0k</text>
|
| 1754 |
<line class="grid-cell" x1="200.7" y1="26" x2="200.7" y2="200"></line>
|
|
|
|
| 1758 |
<line class="grid-cell" x1="510.0" y1="26" x2="510.0" y2="200"></line>
|
| 1759 |
<text class="ts" x="510.0" y="214" text-anchor="middle">3k</text>
|
| 1760 |
<text class="ts" x="510" y="224" text-anchor="end">step</text>
|
| 1761 |
+
<path class="vw v-full" d="M46.0 29.4 L49.1 30.8 L52.2 33.4 L55.3 38.4 L58.4 43.6 L61.5 48.5 L64.6 52.3 L67.7 57.0 L70.7 59.0 L73.8 61.7 L76.9 64.8 L80.0 67.7 L83.1 71.0 L86.2 73.9 L89.3 77.4 L92.4 80.4 L95.5 83.6 L98.6 87.0 L101.7 89.8 L104.8 91.9 L107.9 93.9 L111.0 95.9 L114.1 97.7 L117.1 99.2 L120.2 100.9 L123.3 102.1 L126.4 103.2 L129.5 104.3 L132.6 105.5 L135.7 106.6 L138.8 107.7 L141.9 108.3 L145.0 108.6 L148.1 109.2 L151.2 109.9 L154.3 110.6 L157.4 111.1 L160.5 111.9 L163.5 112.3 L166.6 112.7 L169.7 113.1 L172.8 113.4 L175.9 113.8 L179.0 114.3 L182.1 114.5 L185.2 114.8 L188.3 115.2 L191.4 115.6 L194.5 115.9 L197.6 116.3 L200.7 116.5 L203.8 117.0 L206.9 117.3 L209.9 117.5 L213.0 117.9 L216.1 118.5 L219.2 118.7 L222.3 119.3 L225.4 119.9 L228.5 120.3 L231.6 120.7 L234.7 121.1 L237.8 121.5 L240.9 122.0 L244.0 123.0 L247.1 123.8 L250.2 124.6 L253.3 125.3 L256.3 125.9 L259.4 126.4 L262.5 127.2 L265.6 127.9 L268.7 128.7 L271.8 130.0 L274.9 130.9 L278.0 132.0 L281.1 133.1 L284.2 133.9 L287.3 135.0 L290.4 136.2 L293.5 137.0 L296.6 138.8 L299.7 140.4 L302.7 141.4 L305.8 143.0 L308.9 144.5 L312.0 146.4 L315.1 148.3 L318.2 149.8 L321.3 151.3 L324.4 152.7 L327.5 154.2 L330.6 155.5 L333.7 157.1 L336.8 158.9 L339.9 160.2 L343.0 161.9 L346.1 163.4 L349.1 164.8 L352.2 166.1 L355.3 167.2 L358.4 168.7 L361.5 169.8 L364.6 171.7 L367.7 173.5 L370.8 173.7 L373.9 175.1 L377.0 176.4 L380.1 178.0 L383.2 179.6 L386.3 180.9 L389.4 181.9 L392.5 183.2 L395.5 184.3 L398.6 184.7 L401.7 184.9 L404.8 185.9 L407.9 187.0 L411.0 188.6 L414.1 189.0 L417.2 189.8 L420.3 189.9 L423.4 190.2 L426.5 191.3 L429.6 192.3 L432.7 193.1 L435.8 194.5 L438.9 195.0 L441.9 196.0 L445.0 195.9 L448.1 196.6 L451.2 196.9 L454.3 196.4 L457.4 197.1 L460.5 197.5 L463.6 197.4 L466.7 197.5 L469.8 197.9 L472.9 198.1 L476.0 198.4 L479.1 198.7 L482.2 198.4 L485.3 198.1 L488.3 198.3 L491.4 198.5 L494.5 198.9 L497.6 199.0 L500.7 197.8 L503.8 198.1 L506.9 197.9 L508.5 198.2"></path>
|
| 1762 |
+
<circle class="vdot v-full" cx="508.5" cy="198.2" r="3.5"></circle>
|
| 1763 |
+
<path class="vw v-notrack" d="M46.0 32.9 L49.1 32.4 L52.2 34.3 L55.3 39.7 L58.4 44.7 L61.5 49.9 L64.6 54.0 L67.7 59.1 L70.7 61.8 L73.8 65.5 L76.9 69.3 L80.0 73.4 L83.1 77.7 L86.2 81.0 L89.3 84.7 L92.4 87.2 L95.5 89.6 L98.6 92.5 L101.7 94.7 L104.8 96.4 L107.9 97.8 L111.0 99.3 L114.1 100.5 L117.1 102.1 L120.2 103.5 L123.3 104.1 L126.4 105.2 L129.5 106.0 L132.6 106.9 L135.7 107.8 L138.8 108.8 L141.9 109.4 L145.0 109.7 L148.1 110.2 L151.2 110.5 L154.3 111.1 L157.4 111.5 L160.5 112.0 L163.5 111.9 L166.6 112.4 L169.7 112.5 L172.8 112.7 L175.9 112.8 L179.0 113.2 L182.1 113.3 L185.2 113.7 L188.3 114.0 L191.4 114.7 L194.5 114.8 L197.6 114.6 L200.7 114.9 L203.8 115.4 L206.9 115.8 L209.9 115.9 L213.0 116.4 L216.1 117.2 L219.2 117.4 L222.3 117.6 L225.4 117.8 L228.5 118.1 L231.6 118.6 L234.7 118.4 L237.8 118.8 L240.9 119.1 L244.0 119.6 L247.1 119.7 L250.2 120.2 L253.3 120.7 L256.3 121.0 L259.4 121.3 L262.5 121.8 L265.6 122.5 L268.7 123.0 L271.8 123.6 L274.9 123.5 L278.0 124.1 L281.1 125.1 L284.2 125.3 L287.3 125.7 L290.4 126.2 L293.5 126.5 L296.6 127.1 L299.7 127.8 L302.7 128.6 L305.8 129.1 L308.9 129.7 L312.0 130.0 L315.1 130.7 L318.2 131.6 L321.3 132.4 L324.4 133.2 L327.5 134.1 L330.6 134.8 L333.7 135.7 L336.8 136.8 L339.9 137.7 L343.0 138.6 L346.1 139.2 L349.1 140.4 L352.2 140.9 L355.3 142.2 L358.4 143.1 L361.5 143.5 L364.6 144.3 L367.7 145.2 L370.8 145.9 L373.9 146.9 L377.0 148.2 L380.1 149.6 L383.2 150.1 L386.3 151.4 L389.4 152.3 L392.5 152.8 L395.5 153.3 L398.6 153.8 L401.7 154.5 L404.8 155.6 L407.9 156.3 L411.0 156.9 L414.1 157.5 L417.2 158.2 L420.3 158.7 L423.4 159.4 L426.5 159.7 L429.6 160.9 L432.7 161.5 L435.8 161.5 L438.9 161.7 L441.9 162.2 L445.0 162.4 L448.1 163.2 L451.2 164.1 L454.3 164.3 L457.4 164.6 L460.5 164.8 L463.6 165.1 L466.7 166.1 L469.8 166.9 L472.9 166.9 L476.0 167.5 L479.1 167.5 L482.2 167.9 L485.3 167.5 L488.3 166.9 L491.4 166.8 L494.5 166.7 L497.6 166.9 L500.7 167.1 L503.8 167.1 L506.9 167.3 L508.5 167.5"></path>
|
| 1764 |
+
<circle class="vdot v-notrack" cx="508.5" cy="167.5" r="3.5"></circle>
|
| 1765 |
+
<path class="vw v-deform" d="M46.0 26.0 L49.1 27.0 L52.2 29.1 L55.3 34.2 L58.4 40.0 L61.5 45.5 L64.6 49.8 L67.7 54.8 L70.7 57.4 L73.8 60.3 L76.9 63.8 L80.0 67.0 L83.1 70.9 L86.2 74.3 L89.3 78.1 L92.4 81.1 L95.5 84.3 L98.6 87.8 L101.7 90.4 L104.8 92.3 L107.9 94.3 L111.0 96.3 L114.1 98.0 L117.1 99.5 L120.2 101.1 L123.3 102.2 L126.4 103.3 L129.5 104.2 L132.6 105.4 L135.7 106.5 L138.8 107.5 L141.9 108.2 L145.0 108.5 L148.1 109.1 L151.2 109.8 L154.3 110.5 L157.4 111.0 L160.5 111.8 L163.5 112.3 L166.6 112.7 L169.7 113.1 L172.8 113.4 L175.9 113.7 L179.0 114.2 L182.1 114.4 L185.2 114.8 L188.3 115.2 L191.4 115.6 L194.5 116.0 L197.6 116.6 L200.7 116.8 L203.8 117.2 L206.9 117.6 L209.9 117.9 L213.0 118.3 L216.1 118.9 L219.2 119.2 L222.3 119.8 L225.4 120.4 L228.5 120.8 L231.6 121.2 L234.7 121.7 L237.8 122.2 L240.9 122.7 L244.0 123.7 L247.1 124.6 L250.2 125.6 L253.3 126.5 L256.3 127.2 L259.4 127.8 L262.5 128.8 L265.6 129.7 L268.7 130.6 L271.8 131.9 L274.9 132.9 L278.0 134.2 L281.1 135.5 L284.2 136.3 L287.3 137.5 L290.4 138.8 L293.5 139.6 L296.6 141.6 L299.7 143.1 L302.7 144.2 L305.8 145.9 L308.9 147.5 L312.0 149.4 L315.1 151.4 L318.2 152.8 L321.3 154.3 L324.4 155.6 L327.5 157.1 L330.6 158.5 L333.7 160.0 L336.8 161.8 L339.9 163.1 L343.0 164.7 L346.1 166.0 L349.1 167.4 L352.2 168.4 L355.3 169.4 L358.4 171.0 L361.5 172.2 L364.6 174.0 L367.7 175.8 L370.8 176.0 L373.9 177.3 L377.0 178.5 L380.1 180.1 L383.2 181.7 L386.3 183.0 L389.4 183.8 L392.5 185.1 L395.5 186.2 L398.6 186.5 L401.7 186.7 L404.8 187.4 L407.9 188.6 L411.0 190.2 L414.1 190.7 L417.2 191.5 L420.3 191.5 L423.4 191.8 L426.5 192.8 L429.6 193.6 L432.7 194.4 L435.8 195.8 L438.9 196.2 L441.9 197.3 L445.0 197.3 L448.1 197.8 L451.2 198.2 L454.3 197.7 L457.4 198.4 L460.5 198.8 L463.6 198.7 L466.7 198.7 L469.8 199.1 L472.9 199.3 L476.0 199.5 L479.1 199.9 L482.2 199.5 L485.3 199.3 L488.3 199.4 L491.4 199.6 L494.5 199.9 L497.6 200.0 L500.7 198.8 L503.8 199.2 L506.9 199.0 L508.5 199.4"></path>
|
| 1766 |
+
<circle class="vdot v-deform" cx="508.5" cy="199.4" r="3.5"></circle>
|
| 1767 |
</svg></div>
|
| 1768 |
</div>
|
| 1769 |
<figcaption><b>Figure A1 — Same-step training loss convergence.</b>
|
|
|
|
| 1779 |
<th>dynamics</th><th>track</th><th>force</th></tr></thead>
|
| 1780 |
<tbody><tr><td><span class="chip"><span class="vs vs-full"></span> A · full (track+force)</span></td><td class="shape">0.2819</td><td class="shape">0.3325</td><td class="shape">0.1501</td><td class="shape">0.6187</td></tr>
|
| 1781 |
<tr><td><span class="chip"><span class="vs vs-notrack"></span> B · no track</span></td><td class="shape">0.2749</td><td class="shape">0.2837</td><td class="shape">—</td><td class="shape">0.7535</td></tr>
|
| 1782 |
+
<tr><td><span class="chip"><span class="vs vs-noforce"></span> C · no force stage</span></td><td class="shape">0.2979</td><td class="shape">0.3479</td><td class="shape">0.1511</td><td class="shape">—</td></tr>
|
| 1783 |
+
<tr><td><span class="chip"><span class="vs vs-deform"></span> D · deform tactile + force</span></td><td class="shape">0.2818</td><td class="shape">0.3320</td><td class="shape">0.1514</td><td class="shape">0.6145</td></tr></tbody>
|
| 1784 |
</table>
|
| 1785 |
</div>
|
| 1786 |
<h3>Controlled evaluation on frozen checkpoints</h3>
|
|
|
|
| 1789 |
(<code>scripts/eval/trex_ablation_eval.py</code>). Open-loop: each variant
|
| 1790 |
runs its own full sampling loop from pure noise on 16 anchors and is scored
|
| 1791 |
against the ground-truth normalized delta-base chunk.</p></div>
|
| 1792 |
+
<figure><svg viewBox="0 0 560 222" role="img" aria-label="open-loop action MSE ↓ (16 anchors, normalized 62-D)">
|
| 1793 |
<text class="th" x="0" y="14">open-loop action MSE ↓ (16 anchors, normalized 62-D)</text>
|
| 1794 |
<text class="t" x="232" y="39" text-anchor="end">A · cascade (+force stage)</text>
|
| 1795 |
<rect class="v-full" x="240" y="26" width="133.1" height="16" rx="3"></rect>
|
|
|
|
| 1801 |
<rect class="v-notrack" x="240" y="82" width="149.0" height="16" rx="3"></rect>
|
| 1802 |
<text class="ts" x="395.0" y="95">0.230</text>
|
| 1803 |
<text class="t" x="232" y="123" text-anchor="end">B · coarse only</text>
|
| 1804 |
+
<rect class="v-notrack vfaded" x="240" y="110" width="248.3" height="16" rx="3"></rect>
|
| 1805 |
+
<text class="ts" x="494.3" y="123">0.383</text>
|
| 1806 |
+
<text class="t" x="232" y="151" text-anchor="end">C · coarse only</text>
|
| 1807 |
+
<rect class="v-noforce vfaded" x="240" y="138" width="245.7" height="16" rx="3"></rect>
|
| 1808 |
+
<text class="ts" x="491.7" y="151">0.379</text>
|
| 1809 |
+
<text class="t" x="232" y="179" text-anchor="end">D · cascade (+force stage)</text>
|
| 1810 |
+
<rect class="v-deform" x="240" y="166" width="130.6" height="16" rx="3"></rect>
|
| 1811 |
+
<text class="ts" x="376.6" y="179">0.202</text>
|
| 1812 |
+
<text class="t" x="232" y="207" text-anchor="end">D · coarse only</text>
|
| 1813 |
+
<rect class="v-deform vfaded" x="240" y="194" width="250.0" height="16" rx="3"></rect>
|
| 1814 |
+
<text class="ts" x="496.0" y="207">0.386</text>
|
| 1815 |
</svg><figcaption><b>Figure A2 — Open-loop reconstruction.</b> Faded bars: the coarse action read out at τ=0 without tactile refinement (for C this is its only mode). Solid bars: the full cascade.</figcaption></figure>
|
| 1816 |
|
| 1817 |
<div class="tbl-wrap">
|
| 1818 |
<table>
|
| 1819 |
<thead><tr><th>flow MSE @ fixed τ</th><th>τ=0.8</th><th>τ=0.6</th><th>τ=0.4</th><th>τ=0.2</th><th>τ=0.05</th></tr></thead>
|
| 1820 |
+
<tbody><tr><td><span class="chip"><span class="vs vs-full"></span> A · action</span></td><td class="shape">0.139</td><td class="shape">0.204</td><td class="shape">0.399</td><td class="shape">0.752</td><td class="shape">1.076</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span> B · action</span></td><td class="shape">0.124</td><td class="shape">0.191</td><td class="shape">0.388</td><td class="shape">0.745</td><td class="shape">1.078</td></tr><tr><td><span class="chip"><span class="vs vs-noforce"></span> C · action</span></td><td class="shape">0.135</td><td class="shape">0.201</td><td class="shape">0.396</td><td class="shape">0.744</td><td class="shape">1.085</td></tr><tr><td><span class="chip"><span class="vs vs-deform"></span> D · action</span></td><td class="shape">0.140</td><td class="shape">0.205</td><td class="shape">0.398</td><td class="shape">0.754</td><td class="shape">1.071</td></tr><tr><td><span class="chip"><span class="vs vs-full"></span> A · dynamics</span></td><td class="shape">0.215</td><td class="shape">0.280</td><td class="shape">0.414</td><td class="shape">0.724</td><td class="shape">1.131</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span> B · dynamics</span></td><td class="shape">0.178</td><td class="shape">0.229</td><td class="shape">0.341</td><td class="shape">0.628</td><td class="shape">1.064</td></tr><tr><td><span class="chip"><span class="vs vs-noforce"></span> C · dynamics</span></td><td class="shape">0.214</td><td class="shape">0.279</td><td class="shape">0.413</td><td class="shape">0.722</td><td class="shape">1.128</td></tr><tr><td><span class="chip"><span class="vs vs-deform"></span> D · dynamics</span></td><td class="shape">0.214</td><td class="shape">0.279</td><td class="shape">0.413</td><td class="shape">0.723</td><td class="shape">1.129</td></tr></tbody>
|
| 1821 |
</table>
|
| 1822 |
</div>
|
| 1823 |
+
<figure>
|
| 1824 |
+
<div class="legend" role="list" aria-label="Stage-1 key token groups">
|
| 1825 |
+
<span class="k"><span class="sw" style="background:var(--c-obs);border-color:var(--c-obs);opacity:.45"></span>cond obs</span>
|
| 1826 |
+
<span class="k"><span class="sw" style="background:var(--c-obs);border-color:var(--c-obs)"></span>obs</span>
|
| 1827 |
+
<span class="k"><span class="sw" style="background:var(--c-act);border-color:var(--c-act)"></span>action</span>
|
| 1828 |
+
<span class="k"><span class="sw" style="background:var(--c-st);border-color:var(--c-st)"></span>state</span>
|
| 1829 |
+
<span class="k"><span class="sw" style="background:var(--c-trk);border-color:var(--c-trk)"></span>track past</span>
|
| 1830 |
+
<span class="k"><span class="sw" style="background:var(--c-trk);border-color:var(--c-trk);opacity:.45"></span>track future</span>
|
| 1831 |
+
</div>
|
| 1832 |
+
<svg viewBox="0 0 640 258" role="img" aria-label="Stage-1: where action queries attend (mass share)">
|
| 1833 |
<text class="th" x="0" y="14">Stage-1: where action queries attend (mass share)</text>
|
| 1834 |
<text class="t" x="102" y="39" text-anchor="end">A · layer 0</text>
|
| 1835 |
<rect class="sg-cond" x="110.0" y="26" width="12.6" height="18" rx="3"></rect>
|
|
|
|
| 1868 |
<rect class="sg-act" x="290.0" y="110" width="331.7" height="18" rx="3"></rect>
|
| 1869 |
<text class="ts" x="456.9" y="123" text-anchor="middle">64%</text>
|
| 1870 |
<rect class="sg-st" x="623.7" y="110" width="2.3" height="18" rx="3"></rect>
|
| 1871 |
+
<text class="t" x="102" y="151" text-anchor="end">C · layer 0</text>
|
| 1872 |
+
<rect class="sg-cond" x="110.0" y="138" width="13.3" height="18" rx="3"></rect>
|
| 1873 |
+
<rect class="sg-obs" x="125.3" y="138" width="54.3" height="18" rx="3"></rect>
|
| 1874 |
+
<text class="ts" x="153.4" y="151" text-anchor="middle">11%</text>
|
| 1875 |
+
<rect class="sg-act" x="181.5" y="138" width="27.9" height="18" rx="3"></rect>
|
| 1876 |
+
<rect class="sg-st" x="211.4" y="138" width="0.5" height="18" rx="3"></rect>
|
| 1877 |
+
<rect class="sg-trkp" x="212.4" y="138" width="235.9" height="18" rx="3"></rect>
|
| 1878 |
+
<text class="ts" x="331.4" y="151" text-anchor="middle">46%</text>
|
| 1879 |
+
<rect class="sg-trkf" x="450.3" y="138" width="175.7" height="18" rx="3"></rect>
|
| 1880 |
+
<text class="ts" x="539.2" y="151" text-anchor="middle">34%</text>
|
| 1881 |
+
<text class="t" x="102" y="179" text-anchor="end">C · layer 29</text>
|
| 1882 |
+
<rect class="sg-cond" x="110.0" y="166" width="24.7" height="18" rx="3"></rect>
|
| 1883 |
+
<rect class="sg-obs" x="136.7" y="166" width="47.5" height="18" rx="3"></rect>
|
| 1884 |
+
<text class="ts" x="161.5" y="179" text-anchor="middle">10%</text>
|
| 1885 |
+
<rect class="sg-act" x="186.2" y="166" width="96.8" height="18" rx="3"></rect>
|
| 1886 |
+
<text class="ts" x="235.7" y="179" text-anchor="middle">19%</text>
|
| 1887 |
+
<rect class="sg-st" x="285.1" y="166" width="0.5" height="18" rx="3"></rect>
|
| 1888 |
+
<rect class="sg-trkp" x="286.3" y="166" width="298.6" height="18" rx="3"></rect>
|
| 1889 |
+
<text class="ts" x="436.6" y="179" text-anchor="middle">58%</text>
|
| 1890 |
+
<rect class="sg-trkf" x="586.9" y="166" width="39.1" height="18" rx="3"></rect>
|
| 1891 |
+
<text class="ts" x="607.4" y="179" text-anchor="middle">8%</text>
|
| 1892 |
+
<text class="t" x="102" y="207" text-anchor="end">D · layer 0</text>
|
| 1893 |
+
<rect class="sg-cond" x="110.0" y="194" width="12.0" height="18" rx="3"></rect>
|
| 1894 |
+
<rect class="sg-obs" x="124.0" y="194" width="50.1" height="18" rx="3"></rect>
|
| 1895 |
+
<text class="ts" x="150.0" y="207" text-anchor="middle">10%</text>
|
| 1896 |
+
<rect class="sg-act" x="176.1" y="194" width="33.0" height="18" rx="3"></rect>
|
| 1897 |
+
<text class="ts" x="193.6" y="207" text-anchor="middle">7%</text>
|
| 1898 |
+
<rect class="sg-st" x="211.1" y="194" width="0.5" height="18" rx="3"></rect>
|
| 1899 |
+
<rect class="sg-trkp" x="211.8" y="194" width="233.0" height="18" rx="3"></rect>
|
| 1900 |
+
<text class="ts" x="329.3" y="207" text-anchor="middle">45%</text>
|
| 1901 |
+
<rect class="sg-trkf" x="446.8" y="194" width="179.2" height="18" rx="3"></rect>
|
| 1902 |
+
<text class="ts" x="537.4" y="207" text-anchor="middle">35%</text>
|
| 1903 |
+
<text class="t" x="102" y="235" text-anchor="end">D · layer 29</text>
|
| 1904 |
+
<rect class="sg-cond" x="110.0" y="222" width="25.2" height="18" rx="3"></rect>
|
| 1905 |
+
<rect class="sg-obs" x="137.2" y="222" width="48.6" height="18" rx="3"></rect>
|
| 1906 |
+
<text class="ts" x="162.5" y="235" text-anchor="middle">10%</text>
|
| 1907 |
+
<rect class="sg-act" x="187.7" y="222" width="89.7" height="18" rx="3"></rect>
|
| 1908 |
+
<text class="ts" x="233.6" y="235" text-anchor="middle">18%</text>
|
| 1909 |
+
<rect class="sg-st" x="279.4" y="222" width="0.5" height="18" rx="3"></rect>
|
| 1910 |
+
<rect class="sg-trkp" x="280.9" y="222" width="301.5" height="18" rx="3"></rect>
|
| 1911 |
+
<text class="ts" x="432.7" y="235" text-anchor="middle">59%</text>
|
| 1912 |
+
<rect class="sg-trkf" x="584.4" y="222" width="41.6" height="18" rx="3"></rect>
|
| 1913 |
+
<text class="ts" x="606.2" y="235" text-anchor="middle">8%</text>
|
| 1914 |
</svg><figcaption><b>Figure A3 — Stage-1 attention routing.</b> Share of self-attention mass from action queries onto each token group at τ=0.4 (mean over heads/blocks).</figcaption></figure>
|
| 1915 |
+
<figure>
|
| 1916 |
+
<div class="legend" role="list" aria-label="Stage-2 key token groups">
|
| 1917 |
+
<span class="k"><span class="sw" style="background:var(--c-act);border-color:var(--c-act)"></span>action</span>
|
| 1918 |
+
<span class="k"><span class="sw" style="background:var(--c-frc);border-color:var(--c-frc)"></span>F6 now</span>
|
| 1919 |
+
<span class="k"><span class="sw" style="background:var(--c-frc);border-color:var(--c-frc);opacity:.45"></span>VQ history</span>
|
| 1920 |
+
<span class="k"><span class="sw" style="background:var(--c-frc);border-color:var(--c-frc);opacity:.7"></span>deform</span>
|
| 1921 |
+
<span class="k"><span class="sw" style="background:var(--c-mem);border-color:var(--c-mem)"></span>coarse memory</span>
|
| 1922 |
+
</div>
|
| 1923 |
+
<svg viewBox="0 0 640 118" role="img" aria-label="Stage-2: where action queries attend (mass share)">
|
| 1924 |
<text class="th" x="0" y="14">Stage-2: where action queries attend (mass share)</text>
|
| 1925 |
<text class="t" x="102" y="39" text-anchor="end">A · L0–L5 mean</text>
|
| 1926 |
<rect class="sg-act" x="110.0" y="26" width="12.3" height="18" rx="3"></rect>
|
|
|
|
| 1937 |
<text class="ts" x="295.1" y="67" text-anchor="middle">10%</text>
|
| 1938 |
<rect class="sg-mem" x="321.7" y="54" width="304.3" height="18" rx="3"></rect>
|
| 1939 |
<text class="ts" x="474.8" y="67" text-anchor="middle">59%</text>
|
| 1940 |
+
<text class="t" x="102" y="95" text-anchor="end">D · L0–L5 mean</text>
|
| 1941 |
+
<rect class="sg-act" x="110.0" y="82" width="11.4" height="18" rx="3"></rect>
|
| 1942 |
+
<rect class="sg-f6" x="123.4" y="82" width="5.4" height="18" rx="3"></rect>
|
| 1943 |
+
<rect class="sg-vq" x="130.8" y="82" width="6.2" height="18" rx="3"></rect>
|
| 1944 |
+
<rect class="sg-deform" x="139.0" y="82" width="6.4" height="18" rx="3"></rect>
|
| 1945 |
+
<rect class="sg-mem" x="147.4" y="82" width="478.6" height="18" rx="3"></rect>
|
| 1946 |
+
<text class="ts" x="387.7" y="95" text-anchor="middle">93%</text>
|
| 1947 |
</svg><figcaption><b>Figure A4 — Stage-2 attention routing.</b> Action-query mass over F6 / VQ-history / deform / coarse-memory tokens during real cascaded refinement.</figcaption></figure>
|
| 1948 |
|
| 1949 |
<div class="tbl-wrap">
|
|
|
|
| 1951 |
<thead><tr><th>Tactile sensitivity (normalized units)</th>
|
| 1952 |
<th>mean |Δa| tactile vs masked</th><th>max |Δa|</th>
|
| 1953 |
<th>mean |refinement|</th></tr></thead>
|
| 1954 |
+
<tbody><tr><td><span class="chip"><span class="vs vs-full"></span> A · full (track+force)</span></td><td class="shape">0.0141</td><td class="shape">0.0664</td><td class="shape">0.2452</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span> B · no track</span></td><td class="shape">0.0256</td><td class="shape">0.1309</td><td class="shape">0.2472</td></tr><tr><td><span class="chip"><span class="vs vs-deform"></span> D · deform tactile + force</span></td><td class="shape">0.0131</td><td class="shape">0.0547</td><td class="shape">0.2472</td></tr></tbody>
|
| 1955 |
</table>
|
| 1956 |
</div>
|
| 1957 |
+
<h3>What the three questions get answered</h3>
|
| 1958 |
+
<div class="prose">
|
| 1959 |
+
<p><b>1 · Do the track tokens help?</b> Yes — and the mechanism is
|
| 1960 |
+
visible. With identical budgets, removing the track pathway leaves the
|
| 1961 |
+
single-step action flow MSE almost unchanged (0.745 vs 0.752 at τ=0.2)
|
| 1962 |
+
but costs <b>+12% open-loop cascade error</b> (0.230 vs 0.205, paired
|
| 1963 |
+
+0.0245 ± 0.0033 over the same 16 anchors and seeds), with the
|
| 1964 |
+
largest hit on the hand dimensions (+19%). The routing explains why: in the
|
| 1965 |
+
full model the action queries put <b>~80% of their self-attention mass on
|
| 1966 |
+
track tokens</b> (past + future), so the plan flows through the geometric
|
| 1967 |
+
pathway; without it they fall back onto raw observation patches and
|
| 1968 |
+
themselves. The loss also propagates downstream — the force stage reads a
|
| 1969 |
+
poorer coarse memory and its flow MSE degrades from 0.643 to 0.778, while
|
| 1970 |
+
its reliance on tactile tokens roughly triples (attention mass 13–15% vs
|
| 1971 |
+
4–5%; tactile sensitivity 0.026 vs 0.014) to compensate.</p>
|
| 1972 |
+
<p><b>2 · Does the force stage earn its place as an output?</b> This is
|
| 1973 |
+
the largest effect in the study. Running the tactile refinement cuts
|
| 1974 |
+
open-loop action error <b>almost in half</b>: 0.386 → 0.205 (paired
|
| 1975 |
+
−0.1805 ± 0.0077), and on the hand dimensions from 0.371 to
|
| 1976 |
+
0.170 (−54%). Critically, variant C’s coarse model — trained without
|
| 1977 |
+
any force stage — performs the same as A’s coarse readout
|
| 1978 |
+
(0.379 vs 0.386), so the gain is produced by the refinement stage itself,
|
| 1979 |
+
not by a training side-effect on the world model. The low-noise segment
|
| 1980 |
+
τ ∈ (0, 0.4] is simply better served by a small,
|
| 1981 |
+
tactile-conditioned expert than by six more Euler steps of the 5B coarse
|
| 1982 |
+
model.</p>
|
| 1983 |
+
<p><b>3 · F6 alone vs deformation tactile + F6?</b> Adding ten per-finger
|
| 1984 |
+
deformation-image tokens gives a small but consistent further gain:
|
| 1985 |
+
open-loop cascade 0.2016 vs 0.2054 (paired −0.0038 ± 0.0015),
|
| 1986 |
+
with the force-stage flow MSE edging from 0.643 to 0.639 and training force
|
| 1987 |
+
loss from 0.64 to 0.61. The refiner does read the new tokens — they
|
| 1988 |
+
receive the same attention mass as each existing tactile stream
|
| 1989 |
+
(≈1.6%, vs F6 1.4% and VQ 1.6%) — but coarse memory still dominates
|
| 1990 |
+
at ~93%. On a single-episode overfit benchmark the headroom above F6 is
|
| 1991 |
+
inherently small; the direction of every metric favors deformation input,
|
| 1992 |
+
and the honest read is “cheap to add, consistently non-negative,
|
| 1993 |
+
needs multi-episode training to show its real value.”</p>
|
| 1994 |
+
<p style="font-size:13.5px;color:var(--muted);">Scope note: all runs overfit
|
| 1995 |
+
one 902-frame episode (~537 anchors) for 3,000 steps; the comparisons
|
| 1996 |
+
isolate representation and architecture effects under identical budgets,
|
| 1997 |
+
not generalization. The 4-GPU 20k-step production run of the full model is
|
| 1998 |
+
reported separately above.</p>
|
| 1999 |
+
</div>
|
| 2000 |
|
| 2001 |
</section>
|
| 2002 |
</div>
|