anhtld commited on
Commit
ff3db2e
·
verified ·
1 Parent(s): 7b42cd2

Auto-sync: 2026-06-30 05:12:29 (part 2)

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035_summary.json ADDED
@@ -0,0 +1,454 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.36,
9
+ "std_success": 0.011404240911829549,
10
+ "mean_progress": 0.5732656509485567,
11
+ "mean_action_mse_to_best": 0.40711666752759745,
12
+ "gain_vs_h4": 0.06329999999999997,
13
+ "gain_vs_h16_rank_checkpoint": 0.06260869565217392,
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_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 48,
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_type_success_bonus_scale": 0.0,
32
+ "retrieval_residual_consensus_penalty_scale": 0.0,
33
+ "retrieval_residual_min_source_progress": 0.0,
34
+ "retrieval_residual_min_source_advantage": -1000000000.0,
35
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
36
+ "retrieval_residual_source_score_bonus_scale": 0.0,
37
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
38
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
39
+ "retrieval_residual_scale": 1.0,
40
+ "retrieval_residual_scales": [
41
+ 0.35,
42
+ 0.4,
43
+ 0.45
44
+ ],
45
+ "retrieval_residual_anchor": "expert",
46
+ "retrieval_residual_direction": "candidate_minus_anchor",
47
+ "retrieval_residual_reduce": "compose_mean_by_type",
48
+ "retrieval_residual_challenger_types": [
49
+ "residual_near_miss"
50
+ ],
51
+ "retrieval_residual_challenger_scales": [
52
+ 0.35
53
+ ],
54
+ "retrieval_residual_challenger_margin": 0.01,
55
+ "lattice_exclude_types": [
56
+ "residual_random_negative",
57
+ "residual_wrong_direction",
58
+ "residual_near_miss+residual_no_op"
59
+ ],
60
+ "candidate_type_bonuses": {
61
+ "residual_no_op": 0.03
62
+ },
63
+ "candidate_type_bonus_components": false,
64
+ "candidate_oracle_rollouts": 0,
65
+ "candidate_oracle_unique_tolerance": 1e-06,
66
+ "candidate_oracle_success_rate": null,
67
+ "candidate_oracle_progress": null,
68
+ "candidate_oracle_score": null,
69
+ "candidate_oracle_regret": null,
70
+ "candidate_oracle_score_gain_over_selected": null,
71
+ "candidate_oracle_unique_count": null,
72
+ "candidate_oracle_improvement_rate": null,
73
+ "candidate_oracle_selected_branch_success_rate": null,
74
+ "candidate_oracle_selected_branch_progress": null,
75
+ "candidate_oracle_type_counts": {},
76
+ "candidate_oracle_best_branch_rank": null,
77
+ "candidate_oracle_best_branch_rank_counts": {},
78
+ "candidate_oracle_branch_success_rates": [],
79
+ "candidate_oracle_branch_progress": [],
80
+ "candidate_oracle_branch_score_gains_over_selected": [],
81
+ "selected_residual_scale_counts": {
82
+ "0.35": 536,
83
+ "0.45": 35,
84
+ "0.4": 4
85
+ },
86
+ "policy_rollout_success_rate": 0.3582608695652174,
87
+ "policy_rollout_progress": 0.5603290924210222,
88
+ "oracle_success_rate": 0.8573913043478261,
89
+ "action_mse_to_best": 0.39446062079104394,
90
+ "best_policy_val": {
91
+ "bc_loss": 0.13721593966086706,
92
+ "field_effect_loss": 0.009290305380192068,
93
+ "field_potential_loss": 0.2666468388504452,
94
+ "field_preference_loss": 0.5130573478009965,
95
+ "lattice_edges": 3833.3333333333335,
96
+ "progress_mae": 0.1933159919248687,
97
+ "rank_acc": 0.8265031774838766,
98
+ "rank_loss": 0.5130523675017886,
99
+ "regret_mae": 0.3756548762321472,
100
+ "success_accuracy": 0.8773836526605818,
101
+ "total_loss": 1.5581054819954767
102
+ },
103
+ "per_task": {
104
+ "LiftPegUpright-v1": {
105
+ "action_mse_to_best": 0.32254583662202985,
106
+ "expert_success_rate": 0.8865979381443299,
107
+ "num_groups": 97,
108
+ "oracle_success_rate": 0.9278350515463918,
109
+ "policy_expert_regret": 1.078290365866779,
110
+ "policy_oracle_regret": 1.1046864339678557,
111
+ "policy_rollout_progress": 0.571525937349526,
112
+ "policy_rollout_success_rate": 0.2268041237113402,
113
+ "restore_max_error": 2.384185791015625e-07
114
+ },
115
+ "PickCube-v1": {
116
+ "action_mse_to_best": 0.30278678062873393,
117
+ "expert_success_rate": 0.9375,
118
+ "num_groups": 208,
119
+ "oracle_success_rate": 0.9471153846153846,
120
+ "policy_expert_regret": 0.9425014915267149,
121
+ "policy_oracle_regret": 0.9559243589532204,
122
+ "policy_rollout_progress": 0.6203420753769863,
123
+ "policy_rollout_success_rate": 0.3557692307692308,
124
+ "restore_max_error": 2.2910535335540771e-07
125
+ },
126
+ "PullCube-v1": {
127
+ "action_mse_to_best": 0.6012081153594054,
128
+ "expert_success_rate": 0.19480519480519481,
129
+ "num_groups": 77,
130
+ "oracle_success_rate": 0.36363636363636365,
131
+ "policy_expert_regret": 0.2426418623797801,
132
+ "policy_oracle_regret": 0.4158791025152563,
133
+ "policy_rollout_progress": 0.3329333368644977,
134
+ "policy_rollout_success_rate": 0.23376623376623376,
135
+ "restore_max_error": 2.3096799850463867e-07
136
+ },
137
+ "PushCube-v1": {
138
+ "action_mse_to_best": 0.3675119048578942,
139
+ "expert_success_rate": 0.8617021276595744,
140
+ "num_groups": 94,
141
+ "oracle_success_rate": 0.9893617021276596,
142
+ "policy_expert_regret": 0.5723979790159996,
143
+ "policy_oracle_regret": 0.6499486254884842,
144
+ "policy_rollout_progress": 0.6904769064264095,
145
+ "policy_rollout_success_rate": 0.6595744680851063,
146
+ "restore_max_error": 1.1920928955078125e-07
147
+ },
148
+ "StackCube-v1": {
149
+ "action_mse_to_best": 0.5223143080042468,
150
+ "expert_success_rate": 0.6767676767676768,
151
+ "num_groups": 99,
152
+ "oracle_success_rate": 0.8585858585858586,
153
+ "policy_expert_regret": 0.9052192963132955,
154
+ "policy_oracle_regret": 1.0528768245318922,
155
+ "policy_rollout_progress": 0.47655923634466496,
156
+ "policy_rollout_success_rate": 0.30303030303030304,
157
+ "restore_max_error": 2.1606683731079102e-07
158
+ }
159
+ }
160
+ },
161
+ {
162
+ "seed": 1,
163
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035.json",
164
+ "num_groups": 575,
165
+ "selection_mode": "retrieval_residual",
166
+ "num_candidates": 48,
167
+ "candidate_sigma": 0.0,
168
+ "selection_margin": 0.2,
169
+ "prepend_policy_candidate": false,
170
+ "field_optim_steps": 0,
171
+ "field_optim_step_size": 0.0,
172
+ "field_optim_trust_radius": 0.0,
173
+ "field_optim_l2_penalty": 0.0,
174
+ "retrieval_neighbors": 4,
175
+ "retrieval_metric": "raw",
176
+ "retrieval_type_min_success": 0.0,
177
+ "retrieval_type_success_bonus_scale": 0.0,
178
+ "retrieval_residual_consensus_penalty_scale": 0.0,
179
+ "retrieval_residual_min_source_progress": 0.0,
180
+ "retrieval_residual_min_source_advantage": -1000000000.0,
181
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
182
+ "retrieval_residual_source_score_bonus_scale": 0.0,
183
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
184
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
185
+ "retrieval_residual_scale": 1.0,
186
+ "retrieval_residual_scales": [
187
+ 0.35,
188
+ 0.4,
189
+ 0.45
190
+ ],
191
+ "retrieval_residual_anchor": "expert",
192
+ "retrieval_residual_direction": "candidate_minus_anchor",
193
+ "retrieval_residual_reduce": "compose_mean_by_type",
194
+ "retrieval_residual_challenger_types": [
195
+ "residual_near_miss"
196
+ ],
197
+ "retrieval_residual_challenger_scales": [
198
+ 0.35
199
+ ],
200
+ "retrieval_residual_challenger_margin": 0.01,
201
+ "lattice_exclude_types": [
202
+ "residual_random_negative",
203
+ "residual_wrong_direction",
204
+ "residual_near_miss+residual_no_op"
205
+ ],
206
+ "candidate_type_bonuses": {
207
+ "residual_no_op": 0.03
208
+ },
209
+ "candidate_type_bonus_components": false,
210
+ "candidate_oracle_rollouts": 0,
211
+ "candidate_oracle_unique_tolerance": 1e-06,
212
+ "candidate_oracle_success_rate": null,
213
+ "candidate_oracle_progress": null,
214
+ "candidate_oracle_score": null,
215
+ "candidate_oracle_regret": null,
216
+ "candidate_oracle_score_gain_over_selected": null,
217
+ "candidate_oracle_unique_count": null,
218
+ "candidate_oracle_improvement_rate": null,
219
+ "candidate_oracle_selected_branch_success_rate": null,
220
+ "candidate_oracle_selected_branch_progress": null,
221
+ "candidate_oracle_type_counts": {},
222
+ "candidate_oracle_best_branch_rank": null,
223
+ "candidate_oracle_best_branch_rank_counts": {},
224
+ "candidate_oracle_branch_success_rates": [],
225
+ "candidate_oracle_branch_progress": [],
226
+ "candidate_oracle_branch_score_gains_over_selected": [],
227
+ "selected_residual_scale_counts": {
228
+ "0.35": 549,
229
+ "0.45": 23,
230
+ "0.4": 3
231
+ },
232
+ "policy_rollout_success_rate": 0.34956521739130436,
233
+ "policy_rollout_progress": 0.5727110972935738,
234
+ "oracle_success_rate": 0.8695652173913043,
235
+ "action_mse_to_best": 0.3992000118400092,
236
+ "best_policy_val": {
237
+ "bc_loss": 0.1315989200439718,
238
+ "field_effect_loss": 0.010110370814800262,
239
+ "field_potential_loss": 0.31546536915832096,
240
+ "field_preference_loss": 0.4960637440284093,
241
+ "lattice_edges": 3833.3333333333335,
242
+ "progress_mae": 0.19023226698239645,
243
+ "rank_acc": 0.8255469501018524,
244
+ "rank_loss": 0.4960531195004781,
245
+ "regret_mae": 0.42156532737943864,
246
+ "success_accuracy": 0.8807473679383596,
247
+ "total_loss": 1.561984618504842
248
+ },
249
+ "per_task": {
250
+ "LiftPegUpright-v1": {
251
+ "action_mse_to_best": 0.3285531557816188,
252
+ "expert_success_rate": 0.8584070796460177,
253
+ "num_groups": 113,
254
+ "oracle_success_rate": 0.9380530973451328,
255
+ "policy_expert_regret": 0.9874511470573138,
256
+ "policy_oracle_regret": 1.0666533279999169,
257
+ "policy_rollout_progress": 0.6354583977861742,
258
+ "policy_rollout_success_rate": 0.23008849557522124,
259
+ "restore_max_error": 2.384185791015625e-07
260
+ },
261
+ "PickCube-v1": {
262
+ "action_mse_to_best": 0.32251623995683115,
263
+ "expert_success_rate": 0.9402173913043478,
264
+ "num_groups": 184,
265
+ "oracle_success_rate": 0.9456521739130435,
266
+ "policy_expert_regret": 1.0648970120363748,
267
+ "policy_oracle_regret": 1.0649210165113048,
268
+ "policy_rollout_progress": 0.5748700545797043,
269
+ "policy_rollout_success_rate": 0.29891304347826086,
270
+ "restore_max_error": 2.384185791015625e-07
271
+ },
272
+ "PullCube-v1": {
273
+ "action_mse_to_best": 0.59552585637491,
274
+ "expert_success_rate": 0.25,
275
+ "num_groups": 76,
276
+ "oracle_success_rate": 0.40789473684210525,
277
+ "policy_expert_regret": 0.2754331403411925,
278
+ "policy_oracle_regret": 0.4985335908192945,
279
+ "policy_rollout_progress": 0.30170775403415684,
280
+ "policy_rollout_success_rate": 0.19736842105263158,
281
+ "restore_max_error": 2.3096799850463867e-07
282
+ },
283
+ "PushCube-v1": {
284
+ "action_mse_to_best": 0.35950080446294835,
285
+ "expert_success_rate": 0.8198198198198198,
286
+ "num_groups": 111,
287
+ "oracle_success_rate": 1.0,
288
+ "policy_expert_regret": 0.32904101532321794,
289
+ "policy_oracle_regret": 0.34620858245604746,
290
+ "policy_rollout_progress": 0.8339715977241328,
291
+ "policy_rollout_success_rate": 0.8198198198198198,
292
+ "restore_max_error": 1.7881393432617188e-07
293
+ },
294
+ "StackCube-v1": {
295
+ "action_mse_to_best": 0.52643909530489,
296
+ "expert_success_rate": 0.6923076923076923,
297
+ "num_groups": 91,
298
+ "oracle_success_rate": 0.8571428571428571,
299
+ "policy_expert_regret": 1.1328637084135642,
300
+ "policy_oracle_regret": 1.2925244997163394,
301
+ "policy_rollout_progress": 0.3980808274431543,
302
+ "policy_rollout_success_rate": 0.15384615384615385,
303
+ "restore_max_error": 1.9371509552001953e-07
304
+ }
305
+ }
306
+ },
307
+ {
308
+ "seed": 2,
309
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035.json",
310
+ "num_groups": 575,
311
+ "selection_mode": "retrieval_residual",
312
+ "num_candidates": 48,
313
+ "candidate_sigma": 0.0,
314
+ "selection_margin": 0.2,
315
+ "prepend_policy_candidate": false,
316
+ "field_optim_steps": 0,
317
+ "field_optim_step_size": 0.0,
318
+ "field_optim_trust_radius": 0.0,
319
+ "field_optim_l2_penalty": 0.0,
320
+ "retrieval_neighbors": 4,
321
+ "retrieval_metric": "raw",
322
+ "retrieval_type_min_success": 0.0,
323
+ "retrieval_type_success_bonus_scale": 0.0,
324
+ "retrieval_residual_consensus_penalty_scale": 0.0,
325
+ "retrieval_residual_min_source_progress": 0.0,
326
+ "retrieval_residual_min_source_advantage": -1000000000.0,
327
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
328
+ "retrieval_residual_source_score_bonus_scale": 0.0,
329
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
330
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
331
+ "retrieval_residual_scale": 1.0,
332
+ "retrieval_residual_scales": [
333
+ 0.35,
334
+ 0.4,
335
+ 0.45
336
+ ],
337
+ "retrieval_residual_anchor": "expert",
338
+ "retrieval_residual_direction": "candidate_minus_anchor",
339
+ "retrieval_residual_reduce": "compose_mean_by_type",
340
+ "retrieval_residual_challenger_types": [
341
+ "residual_near_miss"
342
+ ],
343
+ "retrieval_residual_challenger_scales": [
344
+ 0.35
345
+ ],
346
+ "retrieval_residual_challenger_margin": 0.01,
347
+ "lattice_exclude_types": [
348
+ "residual_random_negative",
349
+ "residual_wrong_direction",
350
+ "residual_near_miss+residual_no_op"
351
+ ],
352
+ "candidate_type_bonuses": {
353
+ "residual_no_op": 0.03
354
+ },
355
+ "candidate_type_bonus_components": false,
356
+ "candidate_oracle_rollouts": 0,
357
+ "candidate_oracle_unique_tolerance": 1e-06,
358
+ "candidate_oracle_success_rate": null,
359
+ "candidate_oracle_progress": null,
360
+ "candidate_oracle_score": null,
361
+ "candidate_oracle_regret": null,
362
+ "candidate_oracle_score_gain_over_selected": null,
363
+ "candidate_oracle_unique_count": null,
364
+ "candidate_oracle_improvement_rate": null,
365
+ "candidate_oracle_selected_branch_success_rate": null,
366
+ "candidate_oracle_selected_branch_progress": null,
367
+ "candidate_oracle_type_counts": {},
368
+ "candidate_oracle_best_branch_rank": null,
369
+ "candidate_oracle_best_branch_rank_counts": {},
370
+ "candidate_oracle_branch_success_rates": [],
371
+ "candidate_oracle_branch_progress": [],
372
+ "candidate_oracle_branch_score_gains_over_selected": [],
373
+ "selected_residual_scale_counts": {
374
+ "0.35": 546,
375
+ "0.45": 25,
376
+ "0.4": 4
377
+ },
378
+ "policy_rollout_success_rate": 0.37217391304347824,
379
+ "policy_rollout_progress": 0.5867567631310743,
380
+ "oracle_success_rate": 0.8765217391304347,
381
+ "action_mse_to_best": 0.4276893699517393,
382
+ "best_policy_val": {
383
+ "bc_loss": 0.11367896075050037,
384
+ "field_effect_loss": 0.009670218582161598,
385
+ "field_potential_loss": 0.2641640139950646,
386
+ "field_preference_loss": 0.5130490180518892,
387
+ "lattice_edges": 3833.3333333333335,
388
+ "progress_mae": 0.2021729110015763,
389
+ "rank_acc": 0.8333857821093665,
390
+ "rank_loss": 0.5130119257503085,
391
+ "regret_mae": 0.3958987047274907,
392
+ "success_accuracy": 0.8680730561415354,
393
+ "total_loss": 1.4394984311527677
394
+ },
395
+ "per_task": {
396
+ "LiftPegUpright-v1": {
397
+ "action_mse_to_best": 0.35602021238203935,
398
+ "expert_success_rate": 0.8229166666666666,
399
+ "num_groups": 96,
400
+ "oracle_success_rate": 0.9270833333333334,
401
+ "policy_expert_regret": 0.8475987246880928,
402
+ "policy_oracle_regret": 0.9538775502393643,
403
+ "policy_rollout_progress": 0.6463976396868626,
404
+ "policy_rollout_success_rate": 0.3333333333333333,
405
+ "restore_max_error": 2.384185791015625e-07
406
+ },
407
+ "PickCube-v1": {
408
+ "action_mse_to_best": 0.3390811845246289,
409
+ "expert_success_rate": 0.9444444444444444,
410
+ "num_groups": 198,
411
+ "oracle_success_rate": 0.9595959595959596,
412
+ "policy_expert_regret": 0.9734921298281412,
413
+ "policy_oracle_regret": 0.9805755803568496,
414
+ "policy_rollout_progress": 0.6379460788548293,
415
+ "policy_rollout_success_rate": 0.3434343434343434,
416
+ "restore_max_error": 2.384185791015625e-07
417
+ },
418
+ "PullCube-v1": {
419
+ "action_mse_to_best": 0.6385045246945487,
420
+ "expert_success_rate": 0.24444444444444444,
421
+ "num_groups": 90,
422
+ "oracle_success_rate": 0.4666666666666667,
423
+ "policy_expert_regret": 0.29265180681314734,
424
+ "policy_oracle_regret": 0.49763310268107386,
425
+ "policy_rollout_progress": 0.3416753923934367,
426
+ "policy_rollout_success_rate": 0.23333333333333334,
427
+ "restore_max_error": 2.384185791015625e-07
428
+ },
429
+ "PushCube-v1": {
430
+ "action_mse_to_best": 0.3970492488824495,
431
+ "expert_success_rate": 0.8514851485148515,
432
+ "num_groups": 101,
433
+ "oracle_success_rate": 1.0,
434
+ "policy_expert_regret": 0.3425388766987489,
435
+ "policy_oracle_regret": 0.39744112810285964,
436
+ "policy_rollout_progress": 0.8104796639763483,
437
+ "policy_rollout_success_rate": 0.7920792079207921,
438
+ "restore_max_error": 2.384185791015625e-07
439
+ },
440
+ "StackCube-v1": {
441
+ "action_mse_to_best": 0.5226443493117888,
442
+ "expert_success_rate": 0.7666666666666667,
443
+ "num_groups": 90,
444
+ "oracle_success_rate": 0.9111111111111111,
445
+ "policy_expert_regret": 1.1750739713509877,
446
+ "policy_oracle_regret": 1.3371944083107843,
447
+ "policy_rollout_progress": 0.4045378933350245,
448
+ "policy_rollout_success_rate": 0.14444444444444443,
449
+ "restore_max_error": 2.086162567138672e-07
450
+ }
451
+ }
452
+ }
453
+ ]
454
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035_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_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035.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: 36.00% +/- 1.14%
11
+ Gain vs h=16 rank checkpoint: +6.26%
12
+ Mean progress: 57.33%
13
+ Mean action MSE to best: 0.407
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual direction | residual reduce | min type success | type success bonus | consensus penalty | min source progress | source progress bonus | source score bonus | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | candidate oracle | oracle gain | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---|---:|---:|---:|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss[0.35]@0.01 | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 35.83% | 56.03% | 85.74% | n/a | n/a | 0.394 |
18
+ | 1 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss[0.35]@0.01 | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 34.96% | 57.27% | 86.96% | n/a | n/a | 0.399 |
19
+ | 2 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss[0.35]@0.01 | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 37.22% | 58.68% | 87.65% | n/a | n/a | 0.428 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040_summary.json ADDED
@@ -0,0 +1,457 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.36,
9
+ "std_success": 0.013130146843949137,
10
+ "mean_progress": 0.5736821288862468,
11
+ "mean_action_mse_to_best": 0.4071669423609864,
12
+ "gain_vs_h4": 0.06329999999999997,
13
+ "gain_vs_h16_rank_checkpoint": 0.06260869565217392,
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_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 48,
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_type_success_bonus_scale": 0.0,
32
+ "retrieval_residual_consensus_penalty_scale": 0.0,
33
+ "retrieval_residual_min_source_progress": 0.0,
34
+ "retrieval_residual_min_source_advantage": -1000000000.0,
35
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
36
+ "retrieval_residual_source_score_bonus_scale": 0.0,
37
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
38
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
39
+ "retrieval_residual_scale": 1.0,
40
+ "retrieval_residual_scales": [
41
+ 0.35,
42
+ 0.4,
43
+ 0.45
44
+ ],
45
+ "retrieval_residual_anchor": "expert",
46
+ "retrieval_residual_direction": "candidate_minus_anchor",
47
+ "retrieval_residual_reduce": "compose_mean_by_type",
48
+ "retrieval_residual_challenger_types": [
49
+ "residual_near_miss"
50
+ ],
51
+ "retrieval_residual_challenger_scales": [
52
+ 0.35,
53
+ 0.4
54
+ ],
55
+ "retrieval_residual_challenger_margin": 0.01,
56
+ "lattice_exclude_types": [
57
+ "residual_random_negative",
58
+ "residual_wrong_direction",
59
+ "residual_near_miss+residual_no_op"
60
+ ],
61
+ "candidate_type_bonuses": {
62
+ "residual_no_op": 0.03
63
+ },
64
+ "candidate_type_bonus_components": false,
65
+ "candidate_oracle_rollouts": 0,
66
+ "candidate_oracle_unique_tolerance": 1e-06,
67
+ "candidate_oracle_success_rate": null,
68
+ "candidate_oracle_progress": null,
69
+ "candidate_oracle_score": null,
70
+ "candidate_oracle_regret": null,
71
+ "candidate_oracle_score_gain_over_selected": null,
72
+ "candidate_oracle_unique_count": null,
73
+ "candidate_oracle_improvement_rate": null,
74
+ "candidate_oracle_selected_branch_success_rate": null,
75
+ "candidate_oracle_selected_branch_progress": null,
76
+ "candidate_oracle_type_counts": {},
77
+ "candidate_oracle_best_branch_rank": null,
78
+ "candidate_oracle_best_branch_rank_counts": {},
79
+ "candidate_oracle_branch_success_rates": [],
80
+ "candidate_oracle_branch_progress": [],
81
+ "candidate_oracle_branch_score_gains_over_selected": [],
82
+ "selected_residual_scale_counts": {
83
+ "0.35": 516,
84
+ "0.4": 24,
85
+ "0.45": 35
86
+ },
87
+ "policy_rollout_success_rate": 0.3582608695652174,
88
+ "policy_rollout_progress": 0.5607444118218415,
89
+ "oracle_success_rate": 0.8573913043478261,
90
+ "action_mse_to_best": 0.39465496900450925,
91
+ "best_policy_val": {
92
+ "bc_loss": 0.13721593966086706,
93
+ "field_effect_loss": 0.009290305380192068,
94
+ "field_potential_loss": 0.2666468388504452,
95
+ "field_preference_loss": 0.5130573478009965,
96
+ "lattice_edges": 3833.3333333333335,
97
+ "progress_mae": 0.1933159919248687,
98
+ "rank_acc": 0.8265031774838766,
99
+ "rank_loss": 0.5130523675017886,
100
+ "regret_mae": 0.3756548762321472,
101
+ "success_accuracy": 0.8773836526605818,
102
+ "total_loss": 1.5581054819954767
103
+ },
104
+ "per_task": {
105
+ "LiftPegUpright-v1": {
106
+ "action_mse_to_best": 0.32265400968146385,
107
+ "expert_success_rate": 0.8865979381443299,
108
+ "num_groups": 97,
109
+ "oracle_success_rate": 0.9278350515463918,
110
+ "policy_expert_regret": 1.078290365866779,
111
+ "policy_oracle_regret": 1.1046864339678557,
112
+ "policy_rollout_progress": 0.571525937349526,
113
+ "policy_rollout_success_rate": 0.2268041237113402,
114
+ "restore_max_error": 2.384185791015625e-07
115
+ },
116
+ "PickCube-v1": {
117
+ "action_mse_to_best": 0.3027997398748994,
118
+ "expert_success_rate": 0.9375,
119
+ "num_groups": 208,
120
+ "oracle_success_rate": 0.9471153846153846,
121
+ "policy_expert_regret": 0.9423819873661089,
122
+ "policy_oracle_regret": 0.9558048547926143,
123
+ "policy_rollout_progress": 0.6204615795375923,
124
+ "policy_rollout_success_rate": 0.3557692307692308,
125
+ "restore_max_error": 2.2910535335540771e-07
126
+ },
127
+ "PullCube-v1": {
128
+ "action_mse_to_best": 0.6011350315499615,
129
+ "expert_success_rate": 0.19480519480519481,
130
+ "num_groups": 77,
131
+ "oracle_success_rate": 0.36363636363636365,
132
+ "policy_expert_regret": 0.24029884475597799,
133
+ "policy_oracle_regret": 0.413100507839087,
134
+ "policy_rollout_progress": 0.33571193154066703,
135
+ "policy_rollout_success_rate": 0.23376623376623376,
136
+ "restore_max_error": 2.3096799850463867e-07
137
+ },
138
+ "PushCube-v1": {
139
+ "action_mse_to_best": 0.3675119048578942,
140
+ "expert_success_rate": 0.8617021276595744,
141
+ "num_groups": 94,
142
+ "oracle_success_rate": 0.9893617021276596,
143
+ "policy_expert_regret": 0.5723979790159996,
144
+ "policy_oracle_regret": 0.6499486254884842,
145
+ "policy_rollout_progress": 0.6904769064264095,
146
+ "policy_rollout_success_rate": 0.6595744680851063,
147
+ "restore_max_error": 1.1920928955078125e-07
148
+ },
149
+ "StackCube-v1": {
150
+ "action_mse_to_best": 0.5233667258436632,
151
+ "expert_success_rate": 0.6767676767676768,
152
+ "num_groups": 99,
153
+ "oracle_success_rate": 0.8585858585858586,
154
+ "policy_expert_regret": 0.9052192963132955,
155
+ "policy_oracle_regret": 1.0528768245318922,
156
+ "policy_rollout_progress": 0.47655923634466496,
157
+ "policy_rollout_success_rate": 0.30303030303030304,
158
+ "restore_max_error": 2.1606683731079102e-07
159
+ }
160
+ }
161
+ },
162
+ {
163
+ "seed": 1,
164
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040.json",
165
+ "num_groups": 575,
166
+ "selection_mode": "retrieval_residual",
167
+ "num_candidates": 48,
168
+ "candidate_sigma": 0.0,
169
+ "selection_margin": 0.2,
170
+ "prepend_policy_candidate": false,
171
+ "field_optim_steps": 0,
172
+ "field_optim_step_size": 0.0,
173
+ "field_optim_trust_radius": 0.0,
174
+ "field_optim_l2_penalty": 0.0,
175
+ "retrieval_neighbors": 4,
176
+ "retrieval_metric": "raw",
177
+ "retrieval_type_min_success": 0.0,
178
+ "retrieval_type_success_bonus_scale": 0.0,
179
+ "retrieval_residual_consensus_penalty_scale": 0.0,
180
+ "retrieval_residual_min_source_progress": 0.0,
181
+ "retrieval_residual_min_source_advantage": -1000000000.0,
182
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
183
+ "retrieval_residual_source_score_bonus_scale": 0.0,
184
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
185
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
186
+ "retrieval_residual_scale": 1.0,
187
+ "retrieval_residual_scales": [
188
+ 0.35,
189
+ 0.4,
190
+ 0.45
191
+ ],
192
+ "retrieval_residual_anchor": "expert",
193
+ "retrieval_residual_direction": "candidate_minus_anchor",
194
+ "retrieval_residual_reduce": "compose_mean_by_type",
195
+ "retrieval_residual_challenger_types": [
196
+ "residual_near_miss"
197
+ ],
198
+ "retrieval_residual_challenger_scales": [
199
+ 0.35,
200
+ 0.4
201
+ ],
202
+ "retrieval_residual_challenger_margin": 0.01,
203
+ "lattice_exclude_types": [
204
+ "residual_random_negative",
205
+ "residual_wrong_direction",
206
+ "residual_near_miss+residual_no_op"
207
+ ],
208
+ "candidate_type_bonuses": {
209
+ "residual_no_op": 0.03
210
+ },
211
+ "candidate_type_bonus_components": false,
212
+ "candidate_oracle_rollouts": 0,
213
+ "candidate_oracle_unique_tolerance": 1e-06,
214
+ "candidate_oracle_success_rate": null,
215
+ "candidate_oracle_progress": null,
216
+ "candidate_oracle_score": null,
217
+ "candidate_oracle_regret": null,
218
+ "candidate_oracle_score_gain_over_selected": null,
219
+ "candidate_oracle_unique_count": null,
220
+ "candidate_oracle_improvement_rate": null,
221
+ "candidate_oracle_selected_branch_success_rate": null,
222
+ "candidate_oracle_selected_branch_progress": null,
223
+ "candidate_oracle_type_counts": {},
224
+ "candidate_oracle_best_branch_rank": null,
225
+ "candidate_oracle_best_branch_rank_counts": {},
226
+ "candidate_oracle_branch_success_rates": [],
227
+ "candidate_oracle_branch_progress": [],
228
+ "candidate_oracle_branch_score_gains_over_selected": [],
229
+ "selected_residual_scale_counts": {
230
+ "0.35": 532,
231
+ "0.45": 23,
232
+ "0.4": 20
233
+ },
234
+ "policy_rollout_success_rate": 0.34782608695652173,
235
+ "policy_rollout_progress": 0.5718394296454347,
236
+ "oracle_success_rate": 0.8695652173913043,
237
+ "action_mse_to_best": 0.3990837439878479,
238
+ "best_policy_val": {
239
+ "bc_loss": 0.1315989200439718,
240
+ "field_effect_loss": 0.010110370814800262,
241
+ "field_potential_loss": 0.31546536915832096,
242
+ "field_preference_loss": 0.4960637440284093,
243
+ "lattice_edges": 3833.3333333333335,
244
+ "progress_mae": 0.19023226698239645,
245
+ "rank_acc": 0.8255469501018524,
246
+ "rank_loss": 0.4960531195004781,
247
+ "regret_mae": 0.42156532737943864,
248
+ "success_accuracy": 0.8807473679383596,
249
+ "total_loss": 1.561984618504842
250
+ },
251
+ "per_task": {
252
+ "LiftPegUpright-v1": {
253
+ "action_mse_to_best": 0.3285531557816188,
254
+ "expert_success_rate": 0.8584070796460177,
255
+ "num_groups": 113,
256
+ "oracle_success_rate": 0.9380530973451328,
257
+ "policy_expert_regret": 0.9874511470573138,
258
+ "policy_oracle_regret": 1.0666533279999169,
259
+ "policy_rollout_progress": 0.6354583977861742,
260
+ "policy_rollout_success_rate": 0.23008849557522124,
261
+ "restore_max_error": 2.384185791015625e-07
262
+ },
263
+ "PickCube-v1": {
264
+ "action_mse_to_best": 0.3225348035642958,
265
+ "expert_success_rate": 0.9402173913043478,
266
+ "num_groups": 184,
267
+ "oracle_success_rate": 0.9456521739130435,
268
+ "policy_expert_regret": 1.0725293160193479,
269
+ "policy_oracle_regret": 1.0725533204942779,
270
+ "policy_rollout_progress": 0.5726725332054269,
271
+ "policy_rollout_success_rate": 0.29347826086956524,
272
+ "restore_max_error": 2.384185791015625e-07
273
+ },
274
+ "PullCube-v1": {
275
+ "action_mse_to_best": 0.5943113193896256,
276
+ "expert_success_rate": 0.25,
277
+ "num_groups": 76,
278
+ "oracle_success_rate": 0.40789473684210525,
279
+ "policy_expert_regret": 0.2753879312474869,
280
+ "policy_oracle_regret": 0.4997317737311517,
281
+ "policy_rollout_progress": 0.30041614848826276,
282
+ "policy_rollout_success_rate": 0.19736842105263158,
283
+ "restore_max_error": 2.3096799850463867e-07
284
+ },
285
+ "PushCube-v1": {
286
+ "action_mse_to_best": 0.35952484795639106,
287
+ "expert_success_rate": 0.8198198198198198,
288
+ "num_groups": 111,
289
+ "oracle_success_rate": 1.0,
290
+ "policy_expert_regret": 0.32904101532321794,
291
+ "policy_oracle_regret": 0.34620858245604746,
292
+ "policy_rollout_progress": 0.8339715977241328,
293
+ "policy_rollout_success_rate": 0.8198198198198198,
294
+ "restore_max_error": 1.7881393432617188e-07
295
+ },
296
+ "StackCube-v1": {
297
+ "action_mse_to_best": 0.5266519113965742,
298
+ "expert_success_rate": 0.6923076923076923,
299
+ "num_groups": 91,
300
+ "oracle_success_rate": 0.8571428571428571,
301
+ "policy_expert_regret": 1.1328494550435098,
302
+ "policy_oracle_regret": 1.292510246346285,
303
+ "policy_rollout_progress": 0.39809508081320877,
304
+ "policy_rollout_success_rate": 0.15384615384615385,
305
+ "restore_max_error": 1.9371509552001953e-07
306
+ }
307
+ }
308
+ },
309
+ {
310
+ "seed": 2,
311
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040.json",
312
+ "num_groups": 575,
313
+ "selection_mode": "retrieval_residual",
314
+ "num_candidates": 48,
315
+ "candidate_sigma": 0.0,
316
+ "selection_margin": 0.2,
317
+ "prepend_policy_candidate": false,
318
+ "field_optim_steps": 0,
319
+ "field_optim_step_size": 0.0,
320
+ "field_optim_trust_radius": 0.0,
321
+ "field_optim_l2_penalty": 0.0,
322
+ "retrieval_neighbors": 4,
323
+ "retrieval_metric": "raw",
324
+ "retrieval_type_min_success": 0.0,
325
+ "retrieval_type_success_bonus_scale": 0.0,
326
+ "retrieval_residual_consensus_penalty_scale": 0.0,
327
+ "retrieval_residual_min_source_progress": 0.0,
328
+ "retrieval_residual_min_source_advantage": -1000000000.0,
329
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
330
+ "retrieval_residual_source_score_bonus_scale": 0.0,
331
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
332
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
333
+ "retrieval_residual_scale": 1.0,
334
+ "retrieval_residual_scales": [
335
+ 0.35,
336
+ 0.4,
337
+ 0.45
338
+ ],
339
+ "retrieval_residual_anchor": "expert",
340
+ "retrieval_residual_direction": "candidate_minus_anchor",
341
+ "retrieval_residual_reduce": "compose_mean_by_type",
342
+ "retrieval_residual_challenger_types": [
343
+ "residual_near_miss"
344
+ ],
345
+ "retrieval_residual_challenger_scales": [
346
+ 0.35,
347
+ 0.4
348
+ ],
349
+ "retrieval_residual_challenger_margin": 0.01,
350
+ "lattice_exclude_types": [
351
+ "residual_random_negative",
352
+ "residual_wrong_direction",
353
+ "residual_near_miss+residual_no_op"
354
+ ],
355
+ "candidate_type_bonuses": {
356
+ "residual_no_op": 0.03
357
+ },
358
+ "candidate_type_bonus_components": false,
359
+ "candidate_oracle_rollouts": 0,
360
+ "candidate_oracle_unique_tolerance": 1e-06,
361
+ "candidate_oracle_success_rate": null,
362
+ "candidate_oracle_progress": null,
363
+ "candidate_oracle_score": null,
364
+ "candidate_oracle_regret": null,
365
+ "candidate_oracle_score_gain_over_selected": null,
366
+ "candidate_oracle_unique_count": null,
367
+ "candidate_oracle_improvement_rate": null,
368
+ "candidate_oracle_selected_branch_success_rate": null,
369
+ "candidate_oracle_selected_branch_progress": null,
370
+ "candidate_oracle_type_counts": {},
371
+ "candidate_oracle_best_branch_rank": null,
372
+ "candidate_oracle_best_branch_rank_counts": {},
373
+ "candidate_oracle_branch_success_rates": [],
374
+ "candidate_oracle_branch_progress": [],
375
+ "candidate_oracle_branch_score_gains_over_selected": [],
376
+ "selected_residual_scale_counts": {
377
+ "0.35": 519,
378
+ "0.45": 25,
379
+ "0.4": 31
380
+ },
381
+ "policy_rollout_success_rate": 0.3739130434782609,
382
+ "policy_rollout_progress": 0.5884625451914642,
383
+ "oracle_success_rate": 0.8765217391304347,
384
+ "action_mse_to_best": 0.42776211409060205,
385
+ "best_policy_val": {
386
+ "bc_loss": 0.11367896075050037,
387
+ "field_effect_loss": 0.009670218582161598,
388
+ "field_potential_loss": 0.2641640139950646,
389
+ "field_preference_loss": 0.5130490180518892,
390
+ "lattice_edges": 3833.3333333333335,
391
+ "progress_mae": 0.2021729110015763,
392
+ "rank_acc": 0.8333857821093665,
393
+ "rank_loss": 0.5130119257503085,
394
+ "regret_mae": 0.3958987047274907,
395
+ "success_accuracy": 0.8680730561415354,
396
+ "total_loss": 1.4394984311527677
397
+ },
398
+ "per_task": {
399
+ "LiftPegUpright-v1": {
400
+ "action_mse_to_best": 0.35602021238203935,
401
+ "expert_success_rate": 0.8229166666666666,
402
+ "num_groups": 96,
403
+ "oracle_success_rate": 0.9270833333333334,
404
+ "policy_expert_regret": 0.8475987246880928,
405
+ "policy_oracle_regret": 0.9538775502393643,
406
+ "policy_rollout_progress": 0.6463976396868626,
407
+ "policy_rollout_success_rate": 0.3333333333333333,
408
+ "restore_max_error": 2.384185791015625e-07
409
+ },
410
+ "PickCube-v1": {
411
+ "action_mse_to_best": 0.3390803976229044,
412
+ "expert_success_rate": 0.9444444444444444,
413
+ "num_groups": 198,
414
+ "oracle_success_rate": 0.9595959595959596,
415
+ "policy_expert_regret": 0.9734896586436514,
416
+ "policy_oracle_regret": 0.9805731091723598,
417
+ "policy_rollout_progress": 0.637948550039319,
418
+ "policy_rollout_success_rate": 0.3434343434343434,
419
+ "restore_max_error": 2.384185791015625e-07
420
+ },
421
+ "PullCube-v1": {
422
+ "action_mse_to_best": 0.6389710100988547,
423
+ "expert_success_rate": 0.24444444444444444,
424
+ "num_groups": 90,
425
+ "oracle_success_rate": 0.4666666666666667,
426
+ "policy_expert_regret": 0.2902786168580254,
427
+ "policy_oracle_regret": 0.4949937036467923,
428
+ "policy_rollout_progress": 0.3525680078400506,
429
+ "policy_rollout_success_rate": 0.24444444444444444,
430
+ "restore_max_error": 2.384185791015625e-07
431
+ },
432
+ "PushCube-v1": {
433
+ "action_mse_to_best": 0.3970492488824495,
434
+ "expert_success_rate": 0.8514851485148515,
435
+ "num_groups": 101,
436
+ "oracle_success_rate": 1.0,
437
+ "policy_expert_regret": 0.3425388766987489,
438
+ "policy_oracle_regret": 0.39744112810285964,
439
+ "policy_rollout_progress": 0.8104796639763483,
440
+ "policy_rollout_success_rate": 0.7920792079207921,
441
+ "restore_max_error": 2.384185791015625e-07
442
+ },
443
+ "StackCube-v1": {
444
+ "action_mse_to_best": 0.5226443493117888,
445
+ "expert_success_rate": 0.7666666666666667,
446
+ "num_groups": 90,
447
+ "oracle_success_rate": 0.9111111111111111,
448
+ "policy_expert_regret": 1.1750739713509877,
449
+ "policy_oracle_regret": 1.3371944083107843,
450
+ "policy_rollout_progress": 0.4045378933350245,
451
+ "policy_rollout_success_rate": 0.14444444444444443,
452
+ "restore_max_error": 2.086162567138672e-07
453
+ }
454
+ }
455
+ }
456
+ ]
457
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040_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_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040.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: 36.00% +/- 1.31%
11
+ Gain vs h=16 rank checkpoint: +6.26%
12
+ Mean progress: 57.37%
13
+ Mean action MSE to best: 0.407
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual direction | residual reduce | min type success | type success bonus | consensus penalty | min source progress | source progress bonus | source score bonus | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | candidate oracle | oracle gain | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---|---:|---:|---:|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss[0.35,0.4]@0.01 | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 35.83% | 56.07% | 85.74% | n/a | n/a | 0.395 |
18
+ | 1 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss[0.35,0.4]@0.01 | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 34.78% | 57.18% | 86.96% | n/a | n/a | 0.399 |
19
+ | 2 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss[0.35,0.4]@0.01 | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 37.39% | 58.85% | 87.65% | n/a | n/a | 0.428 |