anhtld commited on
Commit
beb423d
·
verified ·
1 Parent(s): 5a6240c

Auto-sync: 2026-06-28 21:50:30

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,336 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
3
+ "objective": "near_miss_policy_bc5",
4
+ "out_name": "policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.34956521739130436,
9
+ "std_success": 0.015457729421418423,
10
+ "mean_progress": 0.5655122530659223,
11
+ "mean_action_mse_to_best": 0.39662280325845317,
12
+ "gain_vs_h4": 0.05286521739130434,
13
+ "gain_vs_h16_rank_checkpoint": 0.05217391304347829,
14
+ "rows": [
15
+ {
16
+ "seed": 0,
17
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 4,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.2,
23
+ "prepend_policy_candidate": false,
24
+ "field_optim_steps": 0,
25
+ "field_optim_step_size": 0.0,
26
+ "field_optim_trust_radius": 0.0,
27
+ "field_optim_l2_penalty": 0.0,
28
+ "retrieval_neighbors": 4,
29
+ "retrieval_metric": "raw",
30
+ "retrieval_type_min_success": 0.0,
31
+ "retrieval_residual_scale": 1.0,
32
+ "retrieval_residual_scales": [
33
+ 0.35,
34
+ 0.4,
35
+ 0.45
36
+ ],
37
+ "retrieval_residual_anchor": "expert",
38
+ "retrieval_residual_reduce": "field_softmax",
39
+ "candidate_type_bonuses": {
40
+ "residual_no_op": 0.03
41
+ },
42
+ "selected_residual_scale_counts": {
43
+ "0.35": 4,
44
+ "0.4": 1,
45
+ "0.45": 13
46
+ },
47
+ "policy_rollout_success_rate": 0.3443478260869565,
48
+ "policy_rollout_progress": 0.5510196628251477,
49
+ "oracle_success_rate": 0.8573913043478261,
50
+ "action_mse_to_best": 0.3823184049016108,
51
+ "best_policy_val": {
52
+ "bc_loss": 0.13721593966086706,
53
+ "field_effect_loss": 0.009290305380192068,
54
+ "field_potential_loss": 0.2666468388504452,
55
+ "field_preference_loss": 0.5130573478009965,
56
+ "lattice_edges": 3833.3333333333335,
57
+ "progress_mae": 0.1933159919248687,
58
+ "rank_acc": 0.8265031774838766,
59
+ "rank_loss": 0.5130523675017886,
60
+ "regret_mae": 0.3756548762321472,
61
+ "success_accuracy": 0.8773836526605818,
62
+ "total_loss": 1.5581054819954767
63
+ },
64
+ "per_task": {
65
+ "LiftPegUpright-v1": {
66
+ "action_mse_to_best": 0.3214408095983654,
67
+ "expert_success_rate": 0.8865979381443299,
68
+ "num_groups": 97,
69
+ "oracle_success_rate": 0.9278350515463918,
70
+ "policy_expert_regret": 1.0917842305198158,
71
+ "policy_oracle_regret": 1.1181997082282586,
72
+ "policy_rollout_progress": 0.5683219414396384,
73
+ "policy_rollout_success_rate": 0.21649484536082475,
74
+ "restore_max_error": 2.384185791015625e-07
75
+ },
76
+ "PickCube-v1": {
77
+ "action_mse_to_best": 0.3050235868431628,
78
+ "expert_success_rate": 0.9375,
79
+ "num_groups": 208,
80
+ "oracle_success_rate": 0.9471153846153846,
81
+ "policy_expert_regret": 0.9613077554565209,
82
+ "policy_oracle_regret": 0.9747306228830264,
83
+ "policy_rollout_progress": 0.6159588883702571,
84
+ "policy_rollout_success_rate": 0.34134615384615385,
85
+ "restore_max_error": 2.384185791015625e-07
86
+ },
87
+ "PullCube-v1": {
88
+ "action_mse_to_best": 0.6012102336465538,
89
+ "expert_success_rate": 0.19480519480519481,
90
+ "num_groups": 77,
91
+ "oracle_success_rate": 0.36363636363636365,
92
+ "policy_expert_regret": 0.2478078695002024,
93
+ "policy_oracle_regret": 0.44347636305497257,
94
+ "policy_rollout_progress": 0.30005447082100556,
95
+ "policy_rollout_success_rate": 0.19480519480519481,
96
+ "restore_max_error": 2.3096799850463867e-07
97
+ },
98
+ "PushCube-v1": {
99
+ "action_mse_to_best": 0.3658033414406979,
100
+ "expert_success_rate": 0.8617021276595744,
101
+ "num_groups": 94,
102
+ "oracle_success_rate": 0.9893617021276596,
103
+ "policy_expert_regret": 0.5508309629686335,
104
+ "policy_oracle_regret": 0.628381609441118,
105
+ "policy_rollout_progress": 0.7014056246014352,
106
+ "policy_rollout_success_rate": 0.6702127659574468,
107
+ "restore_max_error": 2.384185791015625e-07
108
+ },
109
+ "StackCube-v1": {
110
+ "action_mse_to_best": 0.44979511250299636,
111
+ "expert_success_rate": 0.6767676767676768,
112
+ "num_groups": 99,
113
+ "oracle_success_rate": 0.8585858585858586,
114
+ "policy_expert_regret": 0.940562992685973,
115
+ "policy_oracle_regret": 1.0996043385580332,
116
+ "policy_rollout_progress": 0.45003339452574953,
117
+ "policy_rollout_success_rate": 0.2828282828282828,
118
+ "restore_max_error": 2.384185791015625e-07
119
+ }
120
+ }
121
+ },
122
+ {
123
+ "seed": 1,
124
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json",
125
+ "num_groups": 575,
126
+ "selection_mode": "retrieval_residual",
127
+ "num_candidates": 4,
128
+ "candidate_sigma": 0.0,
129
+ "selection_margin": 0.2,
130
+ "prepend_policy_candidate": false,
131
+ "field_optim_steps": 0,
132
+ "field_optim_step_size": 0.0,
133
+ "field_optim_trust_radius": 0.0,
134
+ "field_optim_l2_penalty": 0.0,
135
+ "retrieval_neighbors": 4,
136
+ "retrieval_metric": "raw",
137
+ "retrieval_type_min_success": 0.0,
138
+ "retrieval_residual_scale": 1.0,
139
+ "retrieval_residual_scales": [
140
+ 0.35,
141
+ 0.4,
142
+ 0.45
143
+ ],
144
+ "retrieval_residual_anchor": "expert",
145
+ "retrieval_residual_reduce": "field_softmax",
146
+ "candidate_type_bonuses": {
147
+ "residual_no_op": 0.03
148
+ },
149
+ "selected_residual_scale_counts": {
150
+ "0.35": 1,
151
+ "0.45": 7,
152
+ "0.4": 1
153
+ },
154
+ "policy_rollout_success_rate": 0.3373913043478261,
155
+ "policy_rollout_progress": 0.5639034366575272,
156
+ "oracle_success_rate": 0.8695652173913043,
157
+ "action_mse_to_best": 0.3904317512476574,
158
+ "best_policy_val": {
159
+ "bc_loss": 0.1315989200439718,
160
+ "field_effect_loss": 0.010110370814800262,
161
+ "field_potential_loss": 0.31546536915832096,
162
+ "field_preference_loss": 0.4960637440284093,
163
+ "lattice_edges": 3833.3333333333335,
164
+ "progress_mae": 0.19023226698239645,
165
+ "rank_acc": 0.8255469501018524,
166
+ "rank_loss": 0.4960531195004781,
167
+ "regret_mae": 0.42156532737943864,
168
+ "success_accuracy": 0.8807473679383596,
169
+ "total_loss": 1.561984618504842
170
+ },
171
+ "per_task": {
172
+ "LiftPegUpright-v1": {
173
+ "action_mse_to_best": 0.3255710795406351,
174
+ "expert_success_rate": 0.8584070796460177,
175
+ "num_groups": 113,
176
+ "oracle_success_rate": 0.9380530973451328,
177
+ "policy_expert_regret": 0.9873926745051831,
178
+ "policy_oracle_regret": 1.0665947045900126,
179
+ "policy_rollout_progress": 0.6355170211960784,
180
+ "policy_rollout_success_rate": 0.23008849557522124,
181
+ "restore_max_error": 2.384185791015625e-07
182
+ },
183
+ "PickCube-v1": {
184
+ "action_mse_to_best": 0.31958654039012996,
185
+ "expert_success_rate": 0.9402173913043478,
186
+ "num_groups": 184,
187
+ "oracle_success_rate": 0.9456521739130435,
188
+ "policy_expert_regret": 1.0786322231667682,
189
+ "policy_oracle_regret": 1.0786562276416984,
190
+ "policy_rollout_progress": 0.5720044086667021,
191
+ "policy_rollout_success_rate": 0.28804347826086957,
192
+ "restore_max_error": 2.384185791015625e-07
193
+ },
194
+ "PullCube-v1": {
195
+ "action_mse_to_best": 0.5726064107331791,
196
+ "expert_success_rate": 0.25,
197
+ "num_groups": 76,
198
+ "oracle_success_rate": 0.40789473684210525,
199
+ "policy_expert_regret": 0.3242254501796867,
200
+ "policy_oracle_regret": 0.5702821770682931,
201
+ "policy_rollout_progress": 0.26863707391251074,
202
+ "policy_rollout_success_rate": 0.15789473684210525,
203
+ "restore_max_error": 2.384185791015625e-07
204
+ },
205
+ "PushCube-v1": {
206
+ "action_mse_to_best": 0.35624687729386595,
207
+ "expert_success_rate": 0.8198198198198198,
208
+ "num_groups": 111,
209
+ "oracle_success_rate": 1.0,
210
+ "policy_expert_regret": 0.3303900519469837,
211
+ "policy_oracle_regret": 0.380037652345391,
212
+ "policy_rollout_progress": 0.8181605458528072,
213
+ "policy_rollout_success_rate": 0.8018018018018018,
214
+ "restore_max_error": 2.3795291781425476e-07
215
+ },
216
+ "StackCube-v1": {
217
+ "action_mse_to_best": 0.5037726478262262,
218
+ "expert_success_rate": 0.6923076923076923,
219
+ "num_groups": 91,
220
+ "oracle_success_rate": 0.8571428571428571,
221
+ "policy_expert_regret": 1.135888498890531,
222
+ "policy_oracle_regret": 1.2955503460470137,
223
+ "policy_rollout_progress": 0.39505498111248016,
224
+ "policy_rollout_success_rate": 0.15384615384615385,
225
+ "restore_max_error": 1.9371509552001953e-07
226
+ }
227
+ }
228
+ },
229
+ {
230
+ "seed": 2,
231
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json",
232
+ "num_groups": 575,
233
+ "selection_mode": "retrieval_residual",
234
+ "num_candidates": 4,
235
+ "candidate_sigma": 0.0,
236
+ "selection_margin": 0.2,
237
+ "prepend_policy_candidate": false,
238
+ "field_optim_steps": 0,
239
+ "field_optim_step_size": 0.0,
240
+ "field_optim_trust_radius": 0.0,
241
+ "field_optim_l2_penalty": 0.0,
242
+ "retrieval_neighbors": 4,
243
+ "retrieval_metric": "raw",
244
+ "retrieval_type_min_success": 0.0,
245
+ "retrieval_residual_scale": 1.0,
246
+ "retrieval_residual_scales": [
247
+ 0.35,
248
+ 0.4,
249
+ 0.45
250
+ ],
251
+ "retrieval_residual_anchor": "expert",
252
+ "retrieval_residual_reduce": "field_softmax",
253
+ "candidate_type_bonuses": {
254
+ "residual_no_op": 0.03
255
+ },
256
+ "selected_residual_scale_counts": {
257
+ "0.45": 11,
258
+ "0.35": 2
259
+ },
260
+ "policy_rollout_success_rate": 0.36695652173913046,
261
+ "policy_rollout_progress": 0.581613659715092,
262
+ "oracle_success_rate": 0.8765217391304347,
263
+ "action_mse_to_best": 0.41711825362609134,
264
+ "best_policy_val": {
265
+ "bc_loss": 0.11367896075050037,
266
+ "field_effect_loss": 0.009670218582161598,
267
+ "field_potential_loss": 0.2641640139950646,
268
+ "field_preference_loss": 0.5130490180518892,
269
+ "lattice_edges": 3833.3333333333335,
270
+ "progress_mae": 0.2021729110015763,
271
+ "rank_acc": 0.8333857821093665,
272
+ "rank_loss": 0.5130119257503085,
273
+ "regret_mae": 0.3958987047274907,
274
+ "success_accuracy": 0.8680730561415354,
275
+ "total_loss": 1.4394984311527677
276
+ },
277
+ "per_task": {
278
+ "LiftPegUpright-v1": {
279
+ "action_mse_to_best": 0.35361564661434386,
280
+ "expert_success_rate": 0.8229166666666666,
281
+ "num_groups": 96,
282
+ "oracle_success_rate": 0.9270833333333334,
283
+ "policy_expert_regret": 0.8476849501021206,
284
+ "policy_oracle_regret": 0.9539546399998168,
285
+ "policy_rollout_progress": 0.6463205499264101,
286
+ "policy_rollout_success_rate": 0.3333333333333333,
287
+ "restore_max_error": 1.955777406692505e-07
288
+ },
289
+ "PickCube-v1": {
290
+ "action_mse_to_best": 0.33572443467423774,
291
+ "expert_success_rate": 0.9444444444444444,
292
+ "num_groups": 198,
293
+ "oracle_success_rate": 0.9595959595959596,
294
+ "policy_expert_regret": 0.9952109518150488,
295
+ "policy_oracle_regret": 1.0022944023437572,
296
+ "policy_rollout_progress": 0.6313787720194368,
297
+ "policy_rollout_success_rate": 0.3282828282828283,
298
+ "restore_max_error": 2.384185791015625e-07
299
+ },
300
+ "PullCube-v1": {
301
+ "action_mse_to_best": 0.6243080761697557,
302
+ "expert_success_rate": 0.24444444444444444,
303
+ "num_groups": 90,
304
+ "oracle_success_rate": 0.4666666666666667,
305
+ "policy_expert_regret": 0.31340191627817454,
306
+ "policy_oracle_regret": 0.5218257787395891,
307
+ "policy_rollout_progress": 0.32862380725780693,
308
+ "policy_rollout_success_rate": 0.2222222222222222,
309
+ "restore_max_error": 2.384185791015625e-07
310
+ },
311
+ "PushCube-v1": {
312
+ "action_mse_to_best": 0.38675852031400887,
313
+ "expert_success_rate": 0.8514851485148515,
314
+ "num_groups": 101,
315
+ "oracle_success_rate": 1.0,
316
+ "policy_expert_regret": 0.36193514430877005,
317
+ "policy_oracle_regret": 0.41669123730447033,
318
+ "policy_rollout_progress": 0.8011305448737475,
319
+ "policy_rollout_success_rate": 0.7821782178217822,
320
+ "restore_max_error": 2.384185791015625e-07
321
+ },
322
+ "StackCube-v1": {
323
+ "action_mse_to_best": 0.490801314305928,
324
+ "expert_success_rate": 0.7666666666666667,
325
+ "num_groups": 90,
326
+ "oracle_success_rate": 0.9111111111111111,
327
+ "policy_expert_regret": 1.169327420492967,
328
+ "policy_oracle_regret": 1.3055954095390108,
329
+ "policy_rollout_progress": 0.40975285553269913,
330
+ "policy_rollout_success_rate": 0.16666666666666666,
331
+ "restore_max_error": 2.2351741790771484e-07
332
+ }
333
+ }
334
+ }
335
+ ]
336
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03_summary.md ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # h=16 Best-Policy Checkpoint Rollout
2
+
3
+ Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
4
+ Objective: `near_miss_policy_bc5`
5
+ Result file: `policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json`
6
+ Completed seeds: 3
7
+ Baseline h=4 policy success: 29.67%
8
+ Baseline h=16 rank-checkpoint success: 29.74%
9
+
10
+ Mean success: 34.96% +/- 1.55%
11
+ Gain vs h=16 rank checkpoint: +5.22%
12
+ Mean progress: 56.55%
13
+ Mean action MSE to best: 0.397
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 34.43% | 55.10% | 85.74% | 0.382 |
18
+ | 1 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 33.74% | 56.39% | 86.96% | 0.390 |
19
+ | 2 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 36.70% | 58.16% | 87.65% | 0.417 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_summary.json ADDED
@@ -0,0 +1,328 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
3
+ "objective": "near_miss_policy_bc5",
4
+ "out_name": "policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.34956521739130436,
9
+ "std_success": 0.015939393721585544,
10
+ "mean_progress": 0.5651957168639754,
11
+ "mean_action_mse_to_best": 0.3966233951844059,
12
+ "gain_vs_h4": 0.05286521739130434,
13
+ "gain_vs_h16_rank_checkpoint": 0.05217391304347829,
14
+ "rows": [
15
+ {
16
+ "seed": 0,
17
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 4,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.2,
23
+ "prepend_policy_candidate": false,
24
+ "field_optim_steps": 0,
25
+ "field_optim_step_size": 0.0,
26
+ "field_optim_trust_radius": 0.0,
27
+ "field_optim_l2_penalty": 0.0,
28
+ "retrieval_neighbors": 4,
29
+ "retrieval_metric": "raw",
30
+ "retrieval_type_min_success": 0.0,
31
+ "retrieval_residual_scale": 1.0,
32
+ "retrieval_residual_scales": [
33
+ 0.35,
34
+ 0.4,
35
+ 0.45
36
+ ],
37
+ "retrieval_residual_anchor": "expert",
38
+ "retrieval_residual_reduce": "field_softmax",
39
+ "candidate_type_bonuses": {},
40
+ "selected_residual_scale_counts": {
41
+ "0.35": 4,
42
+ "0.45": 12
43
+ },
44
+ "policy_rollout_success_rate": 0.34608695652173915,
45
+ "policy_rollout_progress": 0.5515618072636426,
46
+ "oracle_success_rate": 0.8573913043478261,
47
+ "action_mse_to_best": 0.3824618826628379,
48
+ "best_policy_val": {
49
+ "bc_loss": 0.13721593966086706,
50
+ "field_effect_loss": 0.009290305380192068,
51
+ "field_potential_loss": 0.2666468388504452,
52
+ "field_preference_loss": 0.5130573478009965,
53
+ "lattice_edges": 3833.3333333333335,
54
+ "progress_mae": 0.1933159919248687,
55
+ "rank_acc": 0.8265031774838766,
56
+ "rank_loss": 0.5130523675017886,
57
+ "regret_mae": 0.3756548762321472,
58
+ "success_accuracy": 0.8773836526605818,
59
+ "total_loss": 1.5581054819954767
60
+ },
61
+ "per_task": {
62
+ "LiftPegUpright-v1": {
63
+ "action_mse_to_best": 0.32150011145769014,
64
+ "expert_success_rate": 0.8865979381443299,
65
+ "num_groups": 97,
66
+ "oracle_success_rate": 0.9278350515463918,
67
+ "policy_expert_regret": 1.0783334703175063,
68
+ "policy_oracle_regret": 1.1047489480259491,
69
+ "policy_rollout_progress": 0.5714634232914325,
70
+ "policy_rollout_success_rate": 0.2268041237113402,
71
+ "restore_max_error": 2.384185791015625e-07
72
+ },
73
+ "PickCube-v1": {
74
+ "action_mse_to_best": 0.3043742849348256,
75
+ "expert_success_rate": 0.9375,
76
+ "num_groups": 208,
77
+ "oracle_success_rate": 0.9471153846153846,
78
+ "policy_expert_regret": 0.9613127920490044,
79
+ "policy_oracle_regret": 0.9747356594755099,
80
+ "policy_rollout_progress": 0.6159538517777736,
81
+ "policy_rollout_success_rate": 0.34134615384615385,
82
+ "restore_max_error": 2.384185791015625e-07
83
+ },
84
+ "PullCube-v1": {
85
+ "action_mse_to_best": 0.604050983275686,
86
+ "expert_success_rate": 0.19480519480519481,
87
+ "num_groups": 77,
88
+ "oracle_success_rate": 0.36363636363636365,
89
+ "policy_expert_regret": 0.24775487323013412,
90
+ "policy_oracle_regret": 0.4434233667849043,
91
+ "policy_rollout_progress": 0.3001125539991666,
92
+ "policy_rollout_success_rate": 0.19480519480519481,
93
+ "restore_max_error": 2.3096799850463867e-07
94
+ },
95
+ "PushCube-v1": {
96
+ "action_mse_to_best": 0.36579525946302616,
97
+ "expert_success_rate": 0.8617021276595744,
98
+ "num_groups": 94,
99
+ "oracle_success_rate": 0.9893617021276596,
100
+ "policy_expert_regret": 0.5508309629686335,
101
+ "policy_oracle_regret": 0.628381609441118,
102
+ "policy_rollout_progress": 0.7014056246014352,
103
+ "policy_rollout_success_rate": 0.6702127659574468,
104
+ "restore_max_error": 2.384185791015625e-07
105
+ },
106
+ "StackCube-v1": {
107
+ "action_mse_to_best": 0.44973273082363485,
108
+ "expert_success_rate": 0.6767676767676768,
109
+ "num_groups": 99,
110
+ "oracle_success_rate": 0.8585858585858586,
111
+ "policy_expert_regret": 0.9405284519448425,
112
+ "policy_oracle_regret": 1.0995681312951175,
113
+ "policy_rollout_progress": 0.4500696017886653,
114
+ "policy_rollout_success_rate": 0.2828282828282828,
115
+ "restore_max_error": 2.384185791015625e-07
116
+ }
117
+ }
118
+ },
119
+ {
120
+ "seed": 1,
121
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20.json",
122
+ "num_groups": 575,
123
+ "selection_mode": "retrieval_residual",
124
+ "num_candidates": 4,
125
+ "candidate_sigma": 0.0,
126
+ "selection_margin": 0.2,
127
+ "prepend_policy_candidate": false,
128
+ "field_optim_steps": 0,
129
+ "field_optim_step_size": 0.0,
130
+ "field_optim_trust_radius": 0.0,
131
+ "field_optim_l2_penalty": 0.0,
132
+ "retrieval_neighbors": 4,
133
+ "retrieval_metric": "raw",
134
+ "retrieval_type_min_success": 0.0,
135
+ "retrieval_residual_scale": 1.0,
136
+ "retrieval_residual_scales": [
137
+ 0.35,
138
+ 0.4,
139
+ 0.45
140
+ ],
141
+ "retrieval_residual_anchor": "expert",
142
+ "retrieval_residual_reduce": "field_softmax",
143
+ "candidate_type_bonuses": {},
144
+ "selected_residual_scale_counts": {
145
+ "0.35": 1,
146
+ "0.45": 8
147
+ },
148
+ "policy_rollout_success_rate": 0.3356521739130435,
149
+ "policy_rollout_progress": 0.56240815265995,
150
+ "oracle_success_rate": 0.8695652173913043,
151
+ "action_mse_to_best": 0.3904801991727689,
152
+ "best_policy_val": {
153
+ "bc_loss": 0.1315989200439718,
154
+ "field_effect_loss": 0.010110370814800262,
155
+ "field_potential_loss": 0.31546536915832096,
156
+ "field_preference_loss": 0.4960637440284093,
157
+ "lattice_edges": 3833.3333333333335,
158
+ "progress_mae": 0.19023226698239645,
159
+ "rank_acc": 0.8255469501018524,
160
+ "rank_loss": 0.4960531195004781,
161
+ "regret_mae": 0.42156532737943864,
162
+ "success_accuracy": 0.8807473679383596,
163
+ "total_loss": 1.561984618504842
164
+ },
165
+ "per_task": {
166
+ "LiftPegUpright-v1": {
167
+ "action_mse_to_best": 0.3255710795406351,
168
+ "expert_success_rate": 0.8584070796460177,
169
+ "num_groups": 113,
170
+ "oracle_success_rate": 0.9380530973451328,
171
+ "policy_expert_regret": 0.9873926745051831,
172
+ "policy_oracle_regret": 1.0665947045900126,
173
+ "policy_rollout_progress": 0.6355170211960784,
174
+ "policy_rollout_success_rate": 0.23008849557522124,
175
+ "restore_max_error": 2.384185791015625e-07
176
+ },
177
+ "PickCube-v1": {
178
+ "action_mse_to_best": 0.3195870343960174,
179
+ "expert_success_rate": 0.9402173913043478,
180
+ "num_groups": 184,
181
+ "oracle_success_rate": 0.9456521739130435,
182
+ "policy_expert_regret": 1.0786322231667682,
183
+ "policy_oracle_regret": 1.0786562276416984,
184
+ "policy_rollout_progress": 0.5720044086667021,
185
+ "policy_rollout_success_rate": 0.28804347826086957,
186
+ "restore_max_error": 2.384185791015625e-07
187
+ },
188
+ "PullCube-v1": {
189
+ "action_mse_to_best": 0.5725781589158272,
190
+ "expert_success_rate": 0.25,
191
+ "num_groups": 76,
192
+ "oracle_success_rate": 0.40789473684210525,
193
+ "policy_expert_regret": 0.3242254501796867,
194
+ "policy_oracle_regret": 0.5702528474656375,
195
+ "policy_rollout_progress": 0.26866640351516635,
196
+ "policy_rollout_success_rate": 0.15789473684210525,
197
+ "restore_max_error": 2.384185791015625e-07
198
+ },
199
+ "PushCube-v1": {
200
+ "action_mse_to_best": 0.3562556492450001,
201
+ "expert_success_rate": 0.8198198198198198,
202
+ "num_groups": 111,
203
+ "oracle_success_rate": 1.0,
204
+ "policy_expert_regret": 0.3303900519469837,
205
+ "policy_oracle_regret": 0.39677844619428787,
206
+ "policy_rollout_progress": 0.8104287610129193,
207
+ "policy_rollout_success_rate": 0.7927927927927928,
208
+ "restore_max_error": 2.3795291781425476e-07
209
+ },
210
+ "StackCube-v1": {
211
+ "action_mse_to_best": 0.5040906710284097,
212
+ "expert_success_rate": 0.6923076923076923,
213
+ "num_groups": 91,
214
+ "oracle_success_rate": 0.8571428571428571,
215
+ "policy_expert_regret": 1.1359301387936205,
216
+ "policy_oracle_regret": 1.2955919859501033,
217
+ "policy_rollout_progress": 0.39501334120939063,
218
+ "policy_rollout_success_rate": 0.15384615384615385,
219
+ "restore_max_error": 1.9371509552001953e-07
220
+ }
221
+ }
222
+ },
223
+ {
224
+ "seed": 2,
225
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20.json",
226
+ "num_groups": 575,
227
+ "selection_mode": "retrieval_residual",
228
+ "num_candidates": 4,
229
+ "candidate_sigma": 0.0,
230
+ "selection_margin": 0.2,
231
+ "prepend_policy_candidate": false,
232
+ "field_optim_steps": 0,
233
+ "field_optim_step_size": 0.0,
234
+ "field_optim_trust_radius": 0.0,
235
+ "field_optim_l2_penalty": 0.0,
236
+ "retrieval_neighbors": 4,
237
+ "retrieval_metric": "raw",
238
+ "retrieval_type_min_success": 0.0,
239
+ "retrieval_residual_scale": 1.0,
240
+ "retrieval_residual_scales": [
241
+ 0.35,
242
+ 0.4,
243
+ 0.45
244
+ ],
245
+ "retrieval_residual_anchor": "expert",
246
+ "retrieval_residual_reduce": "field_softmax",
247
+ "candidate_type_bonuses": {},
248
+ "selected_residual_scale_counts": {
249
+ "0.45": 10,
250
+ "0.35": 2
251
+ },
252
+ "policy_rollout_success_rate": 0.36695652173913046,
253
+ "policy_rollout_progress": 0.5816171906683335,
254
+ "oracle_success_rate": 0.8765217391304347,
255
+ "action_mse_to_best": 0.4169281037176109,
256
+ "best_policy_val": {
257
+ "bc_loss": 0.11367896075050037,
258
+ "field_effect_loss": 0.009670218582161598,
259
+ "field_potential_loss": 0.2641640139950646,
260
+ "field_preference_loss": 0.5130490180518892,
261
+ "lattice_edges": 3833.3333333333335,
262
+ "progress_mae": 0.2021729110015763,
263
+ "rank_acc": 0.8333857821093665,
264
+ "rank_loss": 0.5130119257503085,
265
+ "regret_mae": 0.3958987047274907,
266
+ "success_accuracy": 0.8680730561415354,
267
+ "total_loss": 1.4394984311527677
268
+ },
269
+ "per_task": {
270
+ "LiftPegUpright-v1": {
271
+ "action_mse_to_best": 0.35361564661434386,
272
+ "expert_success_rate": 0.8229166666666666,
273
+ "num_groups": 96,
274
+ "oracle_success_rate": 0.9270833333333334,
275
+ "policy_expert_regret": 0.8476849501021206,
276
+ "policy_oracle_regret": 0.9539546399998168,
277
+ "policy_rollout_progress": 0.6463205499264101,
278
+ "policy_rollout_success_rate": 0.3333333333333333,
279
+ "restore_max_error": 1.955777406692505e-07
280
+ },
281
+ "PickCube-v1": {
282
+ "action_mse_to_best": 0.3352602533431667,
283
+ "expert_success_rate": 0.9444444444444444,
284
+ "num_groups": 198,
285
+ "oracle_success_rate": 0.9595959595959596,
286
+ "policy_expert_regret": 0.9952109518150488,
287
+ "policy_oracle_regret": 1.0022944023437572,
288
+ "policy_rollout_progress": 0.6313787720194368,
289
+ "policy_rollout_success_rate": 0.3282828282828283,
290
+ "restore_max_error": 2.384185791015625e-07
291
+ },
292
+ "PullCube-v1": {
293
+ "action_mse_to_best": 0.6242523670196534,
294
+ "expert_success_rate": 0.24444444444444444,
295
+ "num_groups": 90,
296
+ "oracle_success_rate": 0.4666666666666667,
297
+ "policy_expert_regret": 0.31340191627817454,
298
+ "policy_oracle_regret": 0.5218257787395891,
299
+ "policy_rollout_progress": 0.32862380725780693,
300
+ "policy_rollout_success_rate": 0.2222222222222222,
301
+ "restore_max_error": 2.384185791015625e-07
302
+ },
303
+ "PushCube-v1": {
304
+ "action_mse_to_best": 0.38670731018675436,
305
+ "expert_success_rate": 0.8514851485148515,
306
+ "num_groups": 101,
307
+ "oracle_success_rate": 1.0,
308
+ "policy_expert_regret": 0.36193514430877005,
309
+ "policy_oracle_regret": 0.41669123730447033,
310
+ "policy_rollout_progress": 0.8011305448737475,
311
+ "policy_rollout_success_rate": 0.7821782178217822,
312
+ "restore_max_error": 2.384185791015625e-07
313
+ },
314
+ "StackCube-v1": {
315
+ "action_mse_to_best": 0.49072084488968054,
316
+ "expert_success_rate": 0.7666666666666667,
317
+ "num_groups": 90,
318
+ "oracle_success_rate": 0.9111111111111111,
319
+ "policy_expert_regret": 1.1693048616250357,
320
+ "policy_oracle_regret": 1.3055728506710795,
321
+ "policy_rollout_progress": 0.40977541440063053,
322
+ "policy_rollout_success_rate": 0.16666666666666666,
323
+ "restore_max_error": 2.2351741790771484e-07
324
+ }
325
+ }
326
+ }
327
+ ]
328
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_summary.md ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # h=16 Best-Policy Checkpoint Rollout
2
+
3
+ Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
4
+ Objective: `near_miss_policy_bc5`
5
+ Result file: `policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20.json`
6
+ Completed seeds: 3
7
+ Baseline h=4 policy success: 29.67%
8
+ Baseline h=16 rank-checkpoint success: 29.74%
9
+
10
+ Mean success: 34.96% +/- 1.59%
11
+ Gain vs h=16 rank checkpoint: +5.22%
12
+ Mean progress: 56.52%
13
+ Mean action MSE to best: 0.397
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 34.61% | 55.16% | 85.74% | 0.382 |
18
+ | 1 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 33.57% | 56.24% | 86.96% | 0.390 |
19
+ | 2 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 36.70% | 58.16% | 87.65% | 0.417 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,335 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
3
+ "objective": "near_miss_policy_bc5",
4
+ "out_name": "policy_rollout_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.34840579710144925,
9
+ "std_success": 0.013508614722006988,
10
+ "mean_progress": 0.5654974544787073,
11
+ "mean_action_mse_to_best": 0.3965390111870416,
12
+ "gain_vs_h4": 0.051705797101449236,
13
+ "gain_vs_h16_rank_checkpoint": 0.051014492753623186,
14
+ "rows": [
15
+ {
16
+ "seed": 0,
17
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 4,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.2,
23
+ "prepend_policy_candidate": false,
24
+ "field_optim_steps": 0,
25
+ "field_optim_step_size": 0.0,
26
+ "field_optim_trust_radius": 0.0,
27
+ "field_optim_l2_penalty": 0.0,
28
+ "retrieval_neighbors": 8,
29
+ "retrieval_metric": "raw",
30
+ "retrieval_type_min_success": 0.0,
31
+ "retrieval_residual_scale": 1.0,
32
+ "retrieval_residual_scales": [
33
+ 0.35,
34
+ 0.4,
35
+ 0.45
36
+ ],
37
+ "retrieval_residual_anchor": "expert",
38
+ "retrieval_residual_reduce": "field_softmax",
39
+ "candidate_type_bonuses": {
40
+ "residual_no_op": 0.03
41
+ },
42
+ "selected_residual_scale_counts": {
43
+ "0.35": 4,
44
+ "0.45": 11,
45
+ "0.4": 1
46
+ },
47
+ "policy_rollout_success_rate": 0.3443478260869565,
48
+ "policy_rollout_progress": 0.5507538553222042,
49
+ "oracle_success_rate": 0.8573913043478261,
50
+ "action_mse_to_best": 0.3818266779330114,
51
+ "best_policy_val": {
52
+ "bc_loss": 0.13721593966086706,
53
+ "field_effect_loss": 0.009290305380192068,
54
+ "field_potential_loss": 0.2666468388504452,
55
+ "field_preference_loss": 0.5130573478009965,
56
+ "lattice_edges": 3833.3333333333335,
57
+ "progress_mae": 0.1933159919248687,
58
+ "rank_acc": 0.8265031774838766,
59
+ "rank_loss": 0.5130523675017886,
60
+ "regret_mae": 0.3756548762321472,
61
+ "success_accuracy": 0.8773836526605818,
62
+ "total_loss": 1.5581054819954767
63
+ },
64
+ "per_task": {
65
+ "LiftPegUpright-v1": {
66
+ "action_mse_to_best": 0.3214190201065743,
67
+ "expert_success_rate": 0.8865979381443299,
68
+ "num_groups": 97,
69
+ "oracle_success_rate": 0.9278350515463918,
70
+ "policy_expert_regret": 1.0783334703175063,
71
+ "policy_oracle_regret": 1.1047489480259491,
72
+ "policy_rollout_progress": 0.5714634232914325,
73
+ "policy_rollout_success_rate": 0.2268041237113402,
74
+ "restore_max_error": 2.384185791015625e-07
75
+ },
76
+ "PickCube-v1": {
77
+ "action_mse_to_best": 0.30417163487380516,
78
+ "expert_success_rate": 0.9375,
79
+ "num_groups": 208,
80
+ "oracle_success_rate": 0.9471153846153846,
81
+ "policy_expert_regret": 0.9681773735926702,
82
+ "policy_oracle_regret": 0.9816002410191756,
83
+ "policy_rollout_progress": 0.6138969625418003,
84
+ "policy_rollout_success_rate": 0.33653846153846156,
85
+ "restore_max_error": 2.384185791015625e-07
86
+ },
87
+ "PullCube-v1": {
88
+ "action_mse_to_best": 0.6003020009050122,
89
+ "expert_success_rate": 0.19480519480519481,
90
+ "num_groups": 77,
91
+ "oracle_success_rate": 0.36363636363636365,
92
+ "policy_expert_regret": 0.2474593260664154,
93
+ "policy_oracle_regret": 0.4431278196211856,
94
+ "policy_rollout_progress": 0.30044171508619344,
95
+ "policy_rollout_success_rate": 0.19480519480519481,
96
+ "restore_max_error": 2.3096799850463867e-07
97
+ },
98
+ "PushCube-v1": {
99
+ "action_mse_to_best": 0.36545782457006737,
100
+ "expert_success_rate": 0.8617021276595744,
101
+ "num_groups": 94,
102
+ "oracle_success_rate": 0.9893617021276596,
103
+ "policy_expert_regret": 0.5508309629686335,
104
+ "policy_oracle_regret": 0.628381609441118,
105
+ "policy_rollout_progress": 0.7014056246014352,
106
+ "policy_rollout_success_rate": 0.6702127659574468,
107
+ "restore_max_error": 2.384185791015625e-07
108
+ },
109
+ "StackCube-v1": {
110
+ "action_mse_to_best": 0.44978490129414234,
111
+ "expert_success_rate": 0.6767676767676768,
112
+ "num_groups": 99,
113
+ "oracle_success_rate": 0.8585858585858586,
114
+ "policy_expert_regret": 0.9405433550627544,
115
+ "policy_oracle_regret": 1.1001952505773969,
116
+ "policy_rollout_progress": 0.44944248250638597,
117
+ "policy_rollout_success_rate": 0.2828282828282828,
118
+ "restore_max_error": 2.384185791015625e-07
119
+ }
120
+ }
121
+ },
122
+ {
123
+ "seed": 1,
124
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json",
125
+ "num_groups": 575,
126
+ "selection_mode": "retrieval_residual",
127
+ "num_candidates": 4,
128
+ "candidate_sigma": 0.0,
129
+ "selection_margin": 0.2,
130
+ "prepend_policy_candidate": false,
131
+ "field_optim_steps": 0,
132
+ "field_optim_step_size": 0.0,
133
+ "field_optim_trust_radius": 0.0,
134
+ "field_optim_l2_penalty": 0.0,
135
+ "retrieval_neighbors": 8,
136
+ "retrieval_metric": "raw",
137
+ "retrieval_type_min_success": 0.0,
138
+ "retrieval_residual_scale": 1.0,
139
+ "retrieval_residual_scales": [
140
+ 0.35,
141
+ 0.4,
142
+ 0.45
143
+ ],
144
+ "retrieval_residual_anchor": "expert",
145
+ "retrieval_residual_reduce": "field_softmax",
146
+ "candidate_type_bonuses": {
147
+ "residual_no_op": 0.03
148
+ },
149
+ "selected_residual_scale_counts": {
150
+ "0.45": 7,
151
+ "0.35": 1
152
+ },
153
+ "policy_rollout_success_rate": 0.3373913043478261,
154
+ "policy_rollout_progress": 0.5646436969838713,
155
+ "oracle_success_rate": 0.8695652173913043,
156
+ "action_mse_to_best": 0.39020738080953776,
157
+ "best_policy_val": {
158
+ "bc_loss": 0.1315989200439718,
159
+ "field_effect_loss": 0.010110370814800262,
160
+ "field_potential_loss": 0.31546536915832096,
161
+ "field_preference_loss": 0.4960637440284093,
162
+ "lattice_edges": 3833.3333333333335,
163
+ "progress_mae": 0.19023226698239645,
164
+ "rank_acc": 0.8255469501018524,
165
+ "rank_loss": 0.4960531195004781,
166
+ "regret_mae": 0.42156532737943864,
167
+ "success_accuracy": 0.8807473679383596,
168
+ "total_loss": 1.561984618504842
169
+ },
170
+ "per_task": {
171
+ "LiftPegUpright-v1": {
172
+ "action_mse_to_best": 0.3255710795406351,
173
+ "expert_success_rate": 0.8584070796460177,
174
+ "num_groups": 113,
175
+ "oracle_success_rate": 0.9380530973451328,
176
+ "policy_expert_regret": 0.9873926745051831,
177
+ "policy_oracle_regret": 1.0665947045900126,
178
+ "policy_rollout_progress": 0.6355170211960784,
179
+ "policy_rollout_success_rate": 0.23008849557522124,
180
+ "restore_max_error": 2.384185791015625e-07
181
+ },
182
+ "PickCube-v1": {
183
+ "action_mse_to_best": 0.319557454943171,
184
+ "expert_success_rate": 0.9402173913043478,
185
+ "num_groups": 184,
186
+ "oracle_success_rate": 0.9456521739130435,
187
+ "policy_expert_regret": 1.086398953602285,
188
+ "policy_oracle_regret": 1.086422958077215,
189
+ "policy_rollout_progress": 0.5696724608398812,
190
+ "policy_rollout_success_rate": 0.2826086956521739,
191
+ "restore_max_error": 2.384185791015625e-07
192
+ },
193
+ "PullCube-v1": {
194
+ "action_mse_to_best": 0.5727081786056882,
195
+ "expert_success_rate": 0.25,
196
+ "num_groups": 76,
197
+ "oracle_success_rate": 0.40789473684210525,
198
+ "policy_expert_regret": 0.3242254501796867,
199
+ "policy_oracle_regret": 0.5696249035138049,
200
+ "policy_rollout_progress": 0.269294347466999,
201
+ "policy_rollout_success_rate": 0.15789473684210525,
202
+ "restore_max_error": 2.384185791015625e-07
203
+ },
204
+ "PushCube-v1": {
205
+ "action_mse_to_best": 0.3546674934190673,
206
+ "expert_success_rate": 0.8198198198198198,
207
+ "num_groups": 111,
208
+ "oracle_success_rate": 1.0,
209
+ "policy_expert_regret": 0.3304452524260358,
210
+ "policy_oracle_regret": 0.36389342286028303,
211
+ "policy_rollout_progress": 0.8252957663289061,
212
+ "policy_rollout_success_rate": 0.8108108108108109,
213
+ "restore_max_error": 2.3795291781425476e-07
214
+ },
215
+ "StackCube-v1": {
216
+ "action_mse_to_best": 0.5042552409263757,
217
+ "expert_success_rate": 0.6923076923076923,
218
+ "num_groups": 91,
219
+ "oracle_success_rate": 0.8571428571428571,
220
+ "policy_expert_regret": 1.1357482151015774,
221
+ "policy_oracle_regret": 1.2954100622580602,
222
+ "policy_rollout_progress": 0.3951952649014337,
223
+ "policy_rollout_success_rate": 0.15384615384615385,
224
+ "restore_max_error": 1.9371509552001953e-07
225
+ }
226
+ }
227
+ },
228
+ {
229
+ "seed": 2,
230
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json",
231
+ "num_groups": 575,
232
+ "selection_mode": "retrieval_residual",
233
+ "num_candidates": 4,
234
+ "candidate_sigma": 0.0,
235
+ "selection_margin": 0.2,
236
+ "prepend_policy_candidate": false,
237
+ "field_optim_steps": 0,
238
+ "field_optim_step_size": 0.0,
239
+ "field_optim_trust_radius": 0.0,
240
+ "field_optim_l2_penalty": 0.0,
241
+ "retrieval_neighbors": 8,
242
+ "retrieval_metric": "raw",
243
+ "retrieval_type_min_success": 0.0,
244
+ "retrieval_residual_scale": 1.0,
245
+ "retrieval_residual_scales": [
246
+ 0.35,
247
+ 0.4,
248
+ 0.45
249
+ ],
250
+ "retrieval_residual_anchor": "expert",
251
+ "retrieval_residual_reduce": "field_softmax",
252
+ "candidate_type_bonuses": {
253
+ "residual_no_op": 0.03
254
+ },
255
+ "selected_residual_scale_counts": {
256
+ "0.45": 11,
257
+ "0.35": 2
258
+ },
259
+ "policy_rollout_success_rate": 0.3634782608695652,
260
+ "policy_rollout_progress": 0.5810948111300461,
261
+ "oracle_success_rate": 0.8765217391304347,
262
+ "action_mse_to_best": 0.41758297481857565,
263
+ "best_policy_val": {
264
+ "bc_loss": 0.11367896075050037,
265
+ "field_effect_loss": 0.009670218582161598,
266
+ "field_potential_loss": 0.2641640139950646,
267
+ "field_preference_loss": 0.5130490180518892,
268
+ "lattice_edges": 3833.3333333333335,
269
+ "progress_mae": 0.2021729110015763,
270
+ "rank_acc": 0.8333857821093665,
271
+ "rank_loss": 0.5130119257503085,
272
+ "regret_mae": 0.3958987047274907,
273
+ "success_accuracy": 0.8680730561415354,
274
+ "total_loss": 1.4394984311527677
275
+ },
276
+ "per_task": {
277
+ "LiftPegUpright-v1": {
278
+ "action_mse_to_best": 0.35361564661434386,
279
+ "expert_success_rate": 0.8229166666666666,
280
+ "num_groups": 96,
281
+ "oracle_success_rate": 0.9270833333333334,
282
+ "policy_expert_regret": 0.8476849501021206,
283
+ "policy_oracle_regret": 0.9539546399998168,
284
+ "policy_rollout_progress": 0.6463205499264101,
285
+ "policy_rollout_success_rate": 0.3333333333333333,
286
+ "restore_max_error": 1.955777406692505e-07
287
+ },
288
+ "PickCube-v1": {
289
+ "action_mse_to_best": 0.3352602533431667,
290
+ "expert_success_rate": 0.9444444444444444,
291
+ "num_groups": 198,
292
+ "oracle_success_rate": 0.9595959595959596,
293
+ "policy_expert_regret": 0.9952109518150488,
294
+ "policy_oracle_regret": 1.0022944023437572,
295
+ "policy_rollout_progress": 0.6313787720194368,
296
+ "policy_rollout_success_rate": 0.3282828282828283,
297
+ "restore_max_error": 2.384185791015625e-07
298
+ },
299
+ "PullCube-v1": {
300
+ "action_mse_to_best": 0.6275552757912212,
301
+ "expert_success_rate": 0.24444444444444444,
302
+ "num_groups": 90,
303
+ "oracle_success_rate": 0.4666666666666667,
304
+ "policy_expert_regret": 0.31338860672656466,
305
+ "policy_oracle_regret": 0.5218124691879792,
306
+ "policy_rollout_progress": 0.3286371168094168,
307
+ "policy_rollout_success_rate": 0.2222222222222222,
308
+ "restore_max_error": 2.384185791015625e-07
309
+ },
310
+ "PushCube-v1": {
311
+ "action_mse_to_best": 0.3847439618098854,
312
+ "expert_success_rate": 0.8514851485148515,
313
+ "num_groups": 101,
314
+ "oracle_success_rate": 1.0,
315
+ "policy_expert_regret": 0.36193514430877005,
316
+ "policy_oracle_regret": 0.41669123730447033,
317
+ "policy_rollout_progress": 0.8011305448737475,
318
+ "policy_rollout_success_rate": 0.7821782178217822,
319
+ "restore_max_error": 2.384185791015625e-07
320
+ },
321
+ "StackCube-v1": {
322
+ "action_mse_to_best": 0.4938051479972071,
323
+ "expert_success_rate": 0.7666666666666667,
324
+ "num_groups": 90,
325
+ "oracle_success_rate": 0.9111111111111111,
326
+ "policy_expert_regret": 1.1820782697863048,
327
+ "policy_oracle_regret": 1.3311458072728581,
328
+ "policy_rollout_progress": 0.40642468002107407,
329
+ "policy_rollout_success_rate": 0.14444444444444443,
330
+ "restore_max_error": 2.2351741790771484e-07
331
+ }
332
+ }
333
+ }
334
+ ]
335
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03_summary.md ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # h=16 Best-Policy Checkpoint Rollout
2
+
3
+ Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
4
+ Objective: `near_miss_policy_bc5`
5
+ Result file: `policy_rollout_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03.json`
6
+ Completed seeds: 3
7
+ Baseline h=4 policy success: 29.67%
8
+ Baseline h=16 rank-checkpoint success: 29.74%
9
+
10
+ Mean success: 34.84% +/- 1.35%
11
+ Gain vs h=16 rank checkpoint: +5.10%
12
+ Mean progress: 56.55%
13
+ Mean action MSE to best: 0.397
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 4 | no | 8 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 34.43% | 55.08% | 85.74% | 0.382 |
18
+ | 1 | retrieval_residual | 4 | no | 8 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 33.74% | 56.46% | 86.96% | 0.390 |
19
+ | 2 | retrieval_residual | 4 | no | 8 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 36.35% | 58.11% | 87.65% | 0.418 |
results/paper_analysis.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "best_clean_key": "residual_k4_consensus_noopbonus003",
3
- "generated_utc": "2026-06-29T01:43:07+00:00",
4
  "mechanism_gap": {
5
  "best_clean_vs_direct_same_ckpt": 0.06956521739130428,
6
  "best_clean_vs_h16": 0.05507246376811592,
@@ -1284,9 +1284,77 @@
1284
  "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03_summary.json"
1285
  },
1286
  "residual_k4_fieldsoftmax_grid_noopbonus003": {
 
1287
  "label": "K4 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03",
1288
- "missing": true,
1289
- "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03_summary.json"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1290
  },
1291
  "residual_k4_kernel_consensus": {
1292
  "ci95_success": 0.028331991434178094,
@@ -1597,9 +1665,77 @@
1597
  "std_success": 0.010190374394925756
1598
  },
1599
  "residual_k8_fieldsoftmax_grid_noopbonus003": {
 
1600
  "label": "K8 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03",
1601
- "missing": true,
1602
- "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03_summary.json"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1603
  },
1604
  "same_state_full": {
1605
  "ci95_success": 0.0885808391099616,
 
1
  {
2
  "best_clean_key": "residual_k4_consensus_noopbonus003",
3
+ "generated_utc": "2026-06-29T01:49:02+00:00",
4
  "mechanism_gap": {
5
  "best_clean_vs_direct_same_ckpt": 0.06956521739130428,
6
  "best_clean_vs_h16": 0.05507246376811592,
 
1284
  "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03_summary.json"
1285
  },
1286
  "residual_k4_fieldsoftmax_grid_noopbonus003": {
1287
+ "ci95_success": 0.038402227815547745,
1288
  "label": "K4 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03",
1289
+ "mean_action_mse_to_best": 0.39662280325845317,
1290
+ "mean_progress": 0.5655122530659223,
1291
+ "mean_success": 0.34956521739130436,
1292
+ "num_completed": 3,
1293
+ "per_task_success": {
1294
+ "LiftPegUpright-v1": {
1295
+ "mean_num_groups": 102.0,
1296
+ "mean_success": 0.25997222475645976,
1297
+ "std_success": 0.06389511734286676
1298
+ },
1299
+ "PickCube-v1": {
1300
+ "mean_num_groups": 196.66666666666666,
1301
+ "mean_success": 0.31922415346328387,
1302
+ "std_success": 0.027781981539339504
1303
+ },
1304
+ "PullCube-v1": {
1305
+ "mean_num_groups": 81.0,
1306
+ "mean_success": 0.19164071795650742,
1307
+ "std_success": 0.03228028468162416
1308
+ },
1309
+ "PushCube-v1": {
1310
+ "mean_num_groups": 102.0,
1311
+ "mean_success": 0.751397595193677,
1312
+ "std_success": 0.07098946144071037
1313
+ },
1314
+ "StackCube-v1": {
1315
+ "mean_num_groups": 93.33333333333333,
1316
+ "mean_success": 0.2011137011137011,
1317
+ "std_success": 0.07105663963231741
1318
+ }
1319
+ },
1320
+ "seed_action_mse_to_best": {
1321
+ "0": 0.3823184049016108,
1322
+ "1": 0.3904317512476574,
1323
+ "2": 0.41711825362609134
1324
+ },
1325
+ "seed_progress": {
1326
+ "0": 0.5510196628251477,
1327
+ "1": 0.5639034366575272,
1328
+ "2": 0.581613659715092
1329
+ },
1330
+ "seed_success": {
1331
+ "0": 0.3443478260869565,
1332
+ "1": 0.3373913043478261,
1333
+ "2": 0.36695652173913046
1334
+ },
1335
+ "selected_candidate_type_counts": {
1336
+ "retrieval_residual_field_softmax": 40,
1337
+ "retrieval_residual_policy_residual": 1685
1338
+ },
1339
+ "selected_residual_scale_counts": {
1340
+ "0.35": 7,
1341
+ "0.4": 2,
1342
+ "0.45": 31
1343
+ },
1344
+ "selected_type_outcomes": {
1345
+ "retrieval_residual_field_softmax": {
1346
+ "count": 40.0,
1347
+ "mean_progress": 0.7493742050603032,
1348
+ "success_rate": 0.6
1349
+ },
1350
+ "retrieval_residual_policy_residual": {
1351
+ "count": 1685.0,
1352
+ "mean_progress": 0.5611475776476581,
1353
+ "success_rate": 0.34362017804154305
1354
+ }
1355
+ },
1356
+ "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03_summary.json",
1357
+ "std_success": 0.015457729421418421
1358
  },
1359
  "residual_k4_kernel_consensus": {
1360
  "ci95_success": 0.028331991434178094,
 
1665
  "std_success": 0.010190374394925756
1666
  },
1667
  "residual_k8_fieldsoftmax_grid_noopbonus003": {
1668
+ "ci95_success": 0.033559967695396,
1669
  "label": "K8 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03",
1670
+ "mean_action_mse_to_best": 0.39653901118704155,
1671
+ "mean_progress": 0.5654974544787073,
1672
+ "mean_success": 0.3484057971014493,
1673
+ "num_completed": 3,
1674
+ "per_task_success": {
1675
+ "LiftPegUpright-v1": {
1676
+ "mean_num_groups": 102.0,
1677
+ "mean_success": 0.26340865087329823,
1678
+ "std_success": 0.06057881385012102
1679
+ },
1680
+ "PickCube-v1": {
1681
+ "mean_num_groups": 196.66666666666666,
1682
+ "mean_success": 0.31580999515782127,
1683
+ "std_success": 0.029047953228423067
1684
+ },
1685
+ "PullCube-v1": {
1686
+ "mean_num_groups": 81.0,
1687
+ "mean_success": 0.19164071795650742,
1688
+ "std_success": 0.03228028468162416
1689
+ },
1690
+ "PushCube-v1": {
1691
+ "mean_num_groups": 102.0,
1692
+ "mean_success": 0.7544005981966798,
1693
+ "std_success": 0.0743010744752375
1694
+ },
1695
+ "StackCube-v1": {
1696
+ "mean_num_groups": 93.33333333333333,
1697
+ "mean_success": 0.1937062937062937,
1698
+ "std_success": 0.07732492963924059
1699
+ }
1700
+ },
1701
+ "seed_action_mse_to_best": {
1702
+ "0": 0.3818266779330114,
1703
+ "1": 0.39020738080953776,
1704
+ "2": 0.41758297481857565
1705
+ },
1706
+ "seed_progress": {
1707
+ "0": 0.5507538553222042,
1708
+ "1": 0.5646436969838713,
1709
+ "2": 0.5810948111300461
1710
+ },
1711
+ "seed_success": {
1712
+ "0": 0.3443478260869565,
1713
+ "1": 0.3373913043478261,
1714
+ "2": 0.3634782608695652
1715
+ },
1716
+ "selected_candidate_type_counts": {
1717
+ "retrieval_residual_field_softmax": 37,
1718
+ "retrieval_residual_policy_residual": 1688
1719
+ },
1720
+ "selected_residual_scale_counts": {
1721
+ "0.35": 7,
1722
+ "0.4": 1,
1723
+ "0.45": 29
1724
+ },
1725
+ "selected_type_outcomes": {
1726
+ "retrieval_residual_field_softmax": {
1727
+ "count": 37.0,
1728
+ "mean_progress": 0.7771831553530049,
1729
+ "success_rate": 0.5945945945945946
1730
+ },
1731
+ "retrieval_residual_policy_residual": {
1732
+ "count": 1688.0,
1733
+ "mean_progress": 0.5608574243055147,
1734
+ "success_rate": 0.34300947867298576
1735
+ }
1736
+ },
1737
+ "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k8_fieldsoftmax_grid_safe_margin0p20_noopbonus0p03_summary.json",
1738
+ "std_success": 0.013508614722006988
1739
  },
1740
  "same_state_full": {
1741
  "ci95_success": 0.0885808391099616,
results/paper_analysis.md CHANGED
@@ -1,6 +1,6 @@
1
  # Paper Analysis
2
 
3
- Generated: `2026-06-29T01:43:07+00:00`
4
 
5
  ## Main Seed Statistics
6
 
@@ -16,11 +16,11 @@ Generated: `2026-06-29T01:43:07+00:00`
16
  | residual_k4_kernel_consensus_s035_noopbonus003 | K4 kernel-weighted tangent consensus, scale 0.35, no-op bonus 0.03 | 3 | 35.13% +/- 1.14 | +/- 2.83 | 56.56% | 0.395 | +5.39 pp |
17
  | residual_k4_kernel_consensus_s045_noopbonus003 | K4 kernel-weighted tangent consensus, scale 0.45, no-op bonus 0.03 | 3 | 35.19% +/- 1.02 | +/- 2.53 | 56.71% | 0.397 | +5.45 pp |
18
  | residual_k4_fieldsoftmax_grid | K4 field-softmax tangent transport, scales 0.35/0.40/0.45 | 3 | 34.96% +/- 1.59 | +/- 3.96 | 56.52% | 0.397 | +5.22 pp |
19
- | residual_k4_fieldsoftmax_grid_noopbonus003 | K4 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
20
  | residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | K4 field-softmax tangent transport, margin 0.10, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
21
  | residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | K4 field-softmax tangent transport, margin 0.05, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
22
  | residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | K4 field-softmax tangent transport, margin 0.00, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
23
- | residual_k8_fieldsoftmax_grid_noopbonus003 | K8 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
24
  | residual_k4_consensus_noopbonus003 | K4 mean-by-type tangent consensus, no-op bonus 0.03 | 3 | 35.25% +/- 1.28 | +/- 3.18 | 56.68% | 0.395 | +5.51 pp |
25
  | residual_k4_consensus_noopbonus001 | K4 mean-by-type tangent consensus, no-op bonus 0.01 | 3 | 35.19% +/- 1.32 | +/- 3.27 | 56.63% | 0.395 | +5.45 pp |
26
  | residual_k4_consensus_noopbonus002 | K4 mean-by-type tangent consensus, no-op bonus 0.02 | 3 | 35.19% +/- 1.32 | +/- 3.27 | 56.64% | 0.395 | +5.45 pp |
@@ -99,6 +99,10 @@ These rows are measured from raw rollout rows. In residual retrieval, `policy_re
99
  | residual_k4_kernel_consensus_s045_noopbonus003 | retrieval_residual_residual_wrong_gripper | 26 | 46.15% | 63.14% |
100
  | residual_k4_fieldsoftmax_grid | retrieval_residual_policy_residual | 1688 | 34.42% | 56.17% |
101
  | residual_k4_fieldsoftmax_grid | retrieval_residual_field_softmax | 37 | 59.46% | 72.64% |
 
 
 
 
102
  | residual_k4_consensus_noopbonus003 | retrieval_residual_policy_residual | 1649 | 34.57% | 56.12% |
103
  | residual_k4_consensus_noopbonus003 | retrieval_residual_residual_no_op | 53 | 52.83% | 73.47% |
104
  | residual_k4_consensus_noopbonus003 | retrieval_residual_residual_wrong_gripper | 23 | 43.48% | 58.07% |
 
1
  # Paper Analysis
2
 
3
+ Generated: `2026-06-29T01:49:02+00:00`
4
 
5
  ## Main Seed Statistics
6
 
 
16
  | residual_k4_kernel_consensus_s035_noopbonus003 | K4 kernel-weighted tangent consensus, scale 0.35, no-op bonus 0.03 | 3 | 35.13% +/- 1.14 | +/- 2.83 | 56.56% | 0.395 | +5.39 pp |
17
  | residual_k4_kernel_consensus_s045_noopbonus003 | K4 kernel-weighted tangent consensus, scale 0.45, no-op bonus 0.03 | 3 | 35.19% +/- 1.02 | +/- 2.53 | 56.71% | 0.397 | +5.45 pp |
18
  | residual_k4_fieldsoftmax_grid | K4 field-softmax tangent transport, scales 0.35/0.40/0.45 | 3 | 34.96% +/- 1.59 | +/- 3.96 | 56.52% | 0.397 | +5.22 pp |
19
+ | residual_k4_fieldsoftmax_grid_noopbonus003 | K4 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 34.96% +/- 1.55 | +/- 3.84 | 56.55% | 0.397 | +5.22 pp |
20
  | residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | K4 field-softmax tangent transport, margin 0.10, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
21
  | residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | K4 field-softmax tangent transport, margin 0.05, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
22
  | residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | K4 field-softmax tangent transport, margin 0.00, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
23
+ | residual_k8_fieldsoftmax_grid_noopbonus003 | K8 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 34.84% +/- 1.35 | +/- 3.36 | 56.55% | 0.397 | +5.10 pp |
24
  | residual_k4_consensus_noopbonus003 | K4 mean-by-type tangent consensus, no-op bonus 0.03 | 3 | 35.25% +/- 1.28 | +/- 3.18 | 56.68% | 0.395 | +5.51 pp |
25
  | residual_k4_consensus_noopbonus001 | K4 mean-by-type tangent consensus, no-op bonus 0.01 | 3 | 35.19% +/- 1.32 | +/- 3.27 | 56.63% | 0.395 | +5.45 pp |
26
  | residual_k4_consensus_noopbonus002 | K4 mean-by-type tangent consensus, no-op bonus 0.02 | 3 | 35.19% +/- 1.32 | +/- 3.27 | 56.64% | 0.395 | +5.45 pp |
 
99
  | residual_k4_kernel_consensus_s045_noopbonus003 | retrieval_residual_residual_wrong_gripper | 26 | 46.15% | 63.14% |
100
  | residual_k4_fieldsoftmax_grid | retrieval_residual_policy_residual | 1688 | 34.42% | 56.17% |
101
  | residual_k4_fieldsoftmax_grid | retrieval_residual_field_softmax | 37 | 59.46% | 72.64% |
102
+ | residual_k4_fieldsoftmax_grid_noopbonus003 | retrieval_residual_policy_residual | 1685 | 34.36% | 56.11% |
103
+ | residual_k4_fieldsoftmax_grid_noopbonus003 | retrieval_residual_field_softmax | 40 | 60.00% | 74.94% |
104
+ | residual_k8_fieldsoftmax_grid_noopbonus003 | retrieval_residual_policy_residual | 1688 | 34.30% | 56.09% |
105
+ | residual_k8_fieldsoftmax_grid_noopbonus003 | retrieval_residual_field_softmax | 37 | 59.46% | 77.72% |
106
  | residual_k4_consensus_noopbonus003 | retrieval_residual_policy_residual | 1649 | 34.57% | 56.12% |
107
  | residual_k4_consensus_noopbonus003 | retrieval_residual_residual_no_op | 53 | 52.83% | 73.47% |
108
  | residual_k4_consensus_noopbonus003 | retrieval_residual_residual_wrong_gripper | 23 | 43.48% | 58.07% |
results/paper_table_status.json CHANGED
@@ -717,14 +717,14 @@
717
  "story_role": "field-conditioned tangent transport with sparse-action prior",
718
  "fallback_success": null,
719
  "pending_job": "14891902/14891946",
720
- "path_exists": false,
721
- "status": "pending",
722
- "success": null,
723
- "std_success": null,
724
  "completed_seeds": null,
725
- "num_completed": null,
726
  "best_config": null,
727
- "gain_vs_h16_policy": null
728
  },
729
  {
730
  "key": "retrieval_residual_k4_fieldsoftmax_grid_margin010_noopbonus003",
@@ -735,7 +735,7 @@
735
  "expert_proposal": "no",
736
  "story_role": "field-conditioned tangent transport abstention sweep",
737
  "fallback_success": null,
738
- "pending_job": "",
739
  "path_exists": false,
740
  "status": "pending",
741
  "success": null,
@@ -754,7 +754,7 @@
754
  "expert_proposal": "no",
755
  "story_role": "field-conditioned tangent transport abstention sweep",
756
  "fallback_success": null,
757
- "pending_job": "",
758
  "path_exists": false,
759
  "status": "pending",
760
  "success": null,
@@ -773,7 +773,7 @@
773
  "expert_proposal": "no",
774
  "story_role": "field-conditioned tangent transport no-abstention diagnostic",
775
  "fallback_success": null,
776
- "pending_job": "",
777
  "path_exists": false,
778
  "status": "pending",
779
  "success": null,
@@ -793,14 +793,14 @@
793
  "story_role": "field-conditioned tangent transport neighborhood scaling",
794
  "fallback_success": null,
795
  "pending_job": "14891923/14891960",
796
- "path_exists": false,
797
- "status": "pending",
798
- "success": null,
799
- "std_success": null,
800
  "completed_seeds": null,
801
- "num_completed": null,
802
  "best_config": null,
803
- "gain_vs_h16_policy": null
804
  },
805
  {
806
  "key": "retrieval_residual_k4_mean_noopbonus003",
 
717
  "story_role": "field-conditioned tangent transport with sparse-action prior",
718
  "fallback_success": null,
719
  "pending_job": "14891902/14891946",
720
+ "path_exists": true,
721
+ "status": "complete",
722
+ "success": 0.34956521739130436,
723
+ "std_success": 0.015457729421418423,
724
  "completed_seeds": null,
725
+ "num_completed": 3,
726
  "best_config": null,
727
+ "gain_vs_h16_policy": 0.05217391304347829
728
  },
729
  {
730
  "key": "retrieval_residual_k4_fieldsoftmax_grid_margin010_noopbonus003",
 
735
  "expert_proposal": "no",
736
  "story_role": "field-conditioned tangent transport abstention sweep",
737
  "fallback_success": null,
738
+ "pending_job": "14892958/14893002",
739
  "path_exists": false,
740
  "status": "pending",
741
  "success": null,
 
754
  "expert_proposal": "no",
755
  "story_role": "field-conditioned tangent transport abstention sweep",
756
  "fallback_success": null,
757
+ "pending_job": "14892975/14893016",
758
  "path_exists": false,
759
  "status": "pending",
760
  "success": null,
 
773
  "expert_proposal": "no",
774
  "story_role": "field-conditioned tangent transport no-abstention diagnostic",
775
  "fallback_success": null,
776
+ "pending_job": "14892990/14893028",
777
  "path_exists": false,
778
  "status": "pending",
779
  "success": null,
 
793
  "story_role": "field-conditioned tangent transport neighborhood scaling",
794
  "fallback_success": null,
795
  "pending_job": "14891923/14891960",
796
+ "path_exists": true,
797
+ "status": "complete",
798
+ "success": 0.34840579710144925,
799
+ "std_success": 0.013508614722006988,
800
  "completed_seeds": null,
801
+ "num_completed": 3,
802
  "best_config": null,
803
+ "gain_vs_h16_policy": 0.051014492753623186
804
  },
805
  {
806
  "key": "retrieval_residual_k4_mean_noopbonus003",
results/paper_table_status.md CHANGED
@@ -40,11 +40,11 @@ Baseline h=16 policy: 29.74%
40
  | retrieval_residual_k4_kernel_mean_s035_noopbonus003 | K4 kernel-weighted residual retrieval, scale 0.35, margin 0.20, no-op residual bonus 0.03 | complete | 35.13% | +5.39 pp | yes | no | no | local counterfactual tangent-field interpolation scale check |
41
  | retrieval_residual_k4_kernel_mean_s045_noopbonus003 | K4 kernel-weighted residual retrieval, scale 0.45, margin 0.20, no-op residual bonus 0.03 | complete | 35.19% | +5.45 pp | yes | no | no | local counterfactual tangent-field interpolation scale check |
42
  | retrieval_residual_k4_fieldsoftmax_grid | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20 | complete | 34.96% | +5.22 pp | yes | no | no | field-conditioned counterfactual tangent transport |
43
- | retrieval_residual_k4_fieldsoftmax_grid_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20, no-op residual bonus 0.03 | pending 14891902/14891946 | pending | pending | yes | no | no | field-conditioned tangent transport with sparse-action prior |
44
- | retrieval_residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.10, no-op residual bonus 0.03 | pending | pending | pending | yes | no | no | field-conditioned tangent transport abstention sweep |
45
- | retrieval_residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.05, no-op residual bonus 0.03 | pending | pending | pending | yes | no | no | field-conditioned tangent transport abstention sweep |
46
- | retrieval_residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.00, no-op residual bonus 0.03 | pending | pending | pending | yes | no | no | field-conditioned tangent transport no-abstention diagnostic |
47
- | retrieval_residual_k8_fieldsoftmax_grid_noopbonus003 | K8 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20, no-op residual bonus 0.03 | pending 14891923/14891960 | pending | pending | yes | no | no | field-conditioned tangent transport neighborhood scaling |
48
  | retrieval_residual_k4_mean_noopbonus003 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03 | complete | 35.25% | +5.51 pp | yes | no | no | current best clean typed sparse-intervention prior |
49
  | retrieval_residual_k4_mean_noopbonus001 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.01 | complete | 35.19% | +5.45 pp | yes | no | no | typed sparse-intervention prior fine sweep |
50
  | retrieval_residual_k4_mean_noopbonus002 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.02 | complete | 35.19% | +5.45 pp | yes | no | no | typed sparse-intervention prior fine sweep |
 
40
  | retrieval_residual_k4_kernel_mean_s035_noopbonus003 | K4 kernel-weighted residual retrieval, scale 0.35, margin 0.20, no-op residual bonus 0.03 | complete | 35.13% | +5.39 pp | yes | no | no | local counterfactual tangent-field interpolation scale check |
41
  | retrieval_residual_k4_kernel_mean_s045_noopbonus003 | K4 kernel-weighted residual retrieval, scale 0.45, margin 0.20, no-op residual bonus 0.03 | complete | 35.19% | +5.45 pp | yes | no | no | local counterfactual tangent-field interpolation scale check |
42
  | retrieval_residual_k4_fieldsoftmax_grid | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20 | complete | 34.96% | +5.22 pp | yes | no | no | field-conditioned counterfactual tangent transport |
43
+ | retrieval_residual_k4_fieldsoftmax_grid_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20, no-op residual bonus 0.03 | complete | 34.96% | +5.22 pp | yes | no | no | field-conditioned tangent transport with sparse-action prior |
44
+ | retrieval_residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.10, no-op residual bonus 0.03 | pending 14892958/14893002 | pending | pending | yes | no | no | field-conditioned tangent transport abstention sweep |
45
+ | retrieval_residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.05, no-op residual bonus 0.03 | pending 14892975/14893016 | pending | pending | yes | no | no | field-conditioned tangent transport abstention sweep |
46
+ | retrieval_residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.00, no-op residual bonus 0.03 | pending 14892990/14893028 | pending | pending | yes | no | no | field-conditioned tangent transport no-abstention diagnostic |
47
+ | retrieval_residual_k8_fieldsoftmax_grid_noopbonus003 | K8 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20, no-op residual bonus 0.03 | complete | 34.84% | +5.10 pp | yes | no | no | field-conditioned tangent transport neighborhood scaling |
48
  | retrieval_residual_k4_mean_noopbonus003 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03 | complete | 35.25% | +5.51 pp | yes | no | no | current best clean typed sparse-intervention prior |
49
  | retrieval_residual_k4_mean_noopbonus001 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.01 | complete | 35.19% | +5.45 pp | yes | no | no | typed sparse-intervention prior fine sweep |
50
  | retrieval_residual_k4_mean_noopbonus002 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.02 | complete | 35.19% | +5.45 pp | yes | no | no | typed sparse-intervention prior fine sweep |