anhtld commited on
Commit
a6cc9e8
·
verified ·
1 Parent(s): ce7242d

Auto-sync: 2026-06-30 23:51:37

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_typesuccessbonus0p02_nmchallenger0p01_summary.json ADDED
@@ -0,0 +1,466 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_typesuccessbonus0p02_nmchallenger0p01.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3571014492753623,
9
+ "std_success": 0.017506862458598824,
10
+ "mean_progress": 0.5719172206129608,
11
+ "mean_action_mse_to_best": 0.4072915274819926,
12
+ "gain_vs_h4": 0.060401449275362284,
13
+ "gain_vs_h16_rank_checkpoint": 0.059710144927536235,
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_typesuccessbonus0p02_nmchallenger0p01.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 48,
21
+ "candidate_sigma": 0.0,
22
+ "proposal_types": [],
23
+ "proposal_lattice_types": [],
24
+ "selection_margin": 0.2,
25
+ "prepend_policy_candidate": false,
26
+ "field_optim_steps": 0,
27
+ "field_optim_step_size": 0.0,
28
+ "field_optim_trust_radius": 0.0,
29
+ "field_optim_l2_penalty": 0.0,
30
+ "retrieval_neighbors": 4,
31
+ "retrieval_metric": "raw",
32
+ "retrieval_type_min_success": 0.0,
33
+ "retrieval_type_success_bonus_scale": 0.02,
34
+ "retrieval_residual_consensus_penalty_scale": 0.0,
35
+ "retrieval_residual_min_source_progress": 0.0,
36
+ "retrieval_residual_min_source_advantage": -1000000000.0,
37
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
38
+ "retrieval_residual_source_score_bonus_scale": 0.0,
39
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
40
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
41
+ "retrieval_residual_scale": 1.0,
42
+ "retrieval_residual_scales": [
43
+ 0.35,
44
+ 0.4,
45
+ 0.45
46
+ ],
47
+ "retrieval_residual_anchor": "expert",
48
+ "retrieval_residual_direction": "candidate_minus_anchor",
49
+ "retrieval_residual_reduce": "compose_mean_by_type",
50
+ "retrieval_residual_challenger_types": [
51
+ "residual_near_miss"
52
+ ],
53
+ "retrieval_residual_challenger_scales": [],
54
+ "retrieval_residual_challenger_margin": 0.01,
55
+ "retrieval_residual_challenger_type_margins": {},
56
+ "retrieval_residual_challenger_tasks": [],
57
+ "retrieval_residual_challenger_type_tasks": {},
58
+ "lattice_exclude_types": [
59
+ "residual_random_negative",
60
+ "residual_wrong_direction",
61
+ "residual_near_miss+residual_no_op"
62
+ ],
63
+ "lattice_exclude_type_tasks": {},
64
+ "candidate_type_bonuses": {
65
+ "retrieval_residual_no_op": 0.03
66
+ },
67
+ "candidate_type_bonus_components": false,
68
+ "candidate_oracle_rollouts": 0,
69
+ "candidate_oracle_unique_tolerance": 1e-06,
70
+ "candidate_oracle_success_rate": null,
71
+ "candidate_oracle_progress": null,
72
+ "candidate_oracle_score": null,
73
+ "candidate_oracle_regret": null,
74
+ "candidate_oracle_score_gain_over_selected": null,
75
+ "candidate_oracle_unique_count": null,
76
+ "candidate_oracle_improvement_rate": null,
77
+ "candidate_oracle_selected_branch_success_rate": null,
78
+ "candidate_oracle_selected_branch_progress": null,
79
+ "candidate_oracle_type_counts": {},
80
+ "candidate_oracle_best_branch_rank": null,
81
+ "candidate_oracle_best_branch_rank_counts": {},
82
+ "candidate_oracle_branch_success_rates": [],
83
+ "candidate_oracle_branch_progress": [],
84
+ "candidate_oracle_branch_score_gains_over_selected": [],
85
+ "selected_residual_scale_counts": {
86
+ "0.35": 517,
87
+ "0.45": 53,
88
+ "0.4": 5
89
+ },
90
+ "policy_rollout_success_rate": 0.35478260869565215,
91
+ "policy_rollout_progress": 0.5571618246667734,
92
+ "oracle_success_rate": 0.8573913043478261,
93
+ "action_mse_to_best": 0.39444961267482975,
94
+ "best_policy_val": {
95
+ "bc_loss": 0.13721593966086706,
96
+ "field_effect_loss": 0.009290305380192068,
97
+ "field_potential_loss": 0.2666468388504452,
98
+ "field_preference_loss": 0.5130573478009965,
99
+ "lattice_edges": 3833.3333333333335,
100
+ "progress_mae": 0.1933159919248687,
101
+ "rank_acc": 0.8265031774838766,
102
+ "rank_loss": 0.5130523675017886,
103
+ "regret_mae": 0.3756548762321472,
104
+ "success_accuracy": 0.8773836526605818,
105
+ "total_loss": 1.5581054819954767
106
+ },
107
+ "per_task": {
108
+ "LiftPegUpright-v1": {
109
+ "action_mse_to_best": 0.32300319151050344,
110
+ "expert_success_rate": 0.8865979381443299,
111
+ "num_groups": 97,
112
+ "oracle_success_rate": 0.9278350515463918,
113
+ "policy_expert_regret": 1.0783334703175063,
114
+ "policy_oracle_regret": 1.1047489480259491,
115
+ "policy_rollout_progress": 0.5714634232914325,
116
+ "policy_rollout_success_rate": 0.2268041237113402,
117
+ "restore_max_error": 2.384185791015625e-07
118
+ },
119
+ "PickCube-v1": {
120
+ "action_mse_to_best": 0.30278675958442575,
121
+ "expert_success_rate": 0.9375,
122
+ "num_groups": 208,
123
+ "oracle_success_rate": 0.9471153846153846,
124
+ "policy_expert_regret": 0.9510086904494808,
125
+ "policy_oracle_regret": 0.9644315578759863,
126
+ "policy_rollout_progress": 0.6166425687619127,
127
+ "policy_rollout_success_rate": 0.35096153846153844,
128
+ "restore_max_error": 2.384185791015625e-07
129
+ },
130
+ "PullCube-v1": {
131
+ "action_mse_to_best": 0.5978217993850832,
132
+ "expert_success_rate": 0.19480519480519481,
133
+ "num_groups": 77,
134
+ "oracle_success_rate": 0.36363636363636365,
135
+ "policy_expert_regret": 0.2471449282322708,
136
+ "policy_oracle_regret": 0.4208110002740347,
137
+ "policy_rollout_progress": 0.31884596189000197,
138
+ "policy_rollout_success_rate": 0.22077922077922077,
139
+ "restore_max_error": 2.3096799850463867e-07
140
+ },
141
+ "PushCube-v1": {
142
+ "action_mse_to_best": 0.3688891567289829,
143
+ "expert_success_rate": 0.8617021276595744,
144
+ "num_groups": 94,
145
+ "oracle_success_rate": 0.9893617021276596,
146
+ "policy_expert_regret": 0.5508309629686335,
147
+ "policy_oracle_regret": 0.628381609441118,
148
+ "policy_rollout_progress": 0.7014056246014352,
149
+ "policy_rollout_success_rate": 0.6702127659574468,
150
+ "restore_max_error": 2.384185791015625e-07
151
+ },
152
+ "StackCube-v1": {
153
+ "action_mse_to_best": 0.5231284084118376,
154
+ "expert_success_rate": 0.6767676767676768,
155
+ "num_groups": 99,
156
+ "oracle_success_rate": 0.8585858585858586,
157
+ "policy_expert_regret": 0.9252789386294105,
158
+ "policy_oracle_regret": 1.0729591972900159,
159
+ "policy_rollout_progress": 0.46657752569275673,
160
+ "policy_rollout_success_rate": 0.29292929292929293,
161
+ "restore_max_error": 2.384185791015625e-07
162
+ }
163
+ }
164
+ },
165
+ {
166
+ "seed": 1,
167
+ "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_typesuccessbonus0p02_nmchallenger0p01.json",
168
+ "num_groups": 575,
169
+ "selection_mode": "retrieval_residual",
170
+ "num_candidates": 48,
171
+ "candidate_sigma": 0.0,
172
+ "proposal_types": [],
173
+ "proposal_lattice_types": [],
174
+ "selection_margin": 0.2,
175
+ "prepend_policy_candidate": false,
176
+ "field_optim_steps": 0,
177
+ "field_optim_step_size": 0.0,
178
+ "field_optim_trust_radius": 0.0,
179
+ "field_optim_l2_penalty": 0.0,
180
+ "retrieval_neighbors": 4,
181
+ "retrieval_metric": "raw",
182
+ "retrieval_type_min_success": 0.0,
183
+ "retrieval_type_success_bonus_scale": 0.02,
184
+ "retrieval_residual_consensus_penalty_scale": 0.0,
185
+ "retrieval_residual_min_source_progress": 0.0,
186
+ "retrieval_residual_min_source_advantage": -1000000000.0,
187
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
188
+ "retrieval_residual_source_score_bonus_scale": 0.0,
189
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
190
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
191
+ "retrieval_residual_scale": 1.0,
192
+ "retrieval_residual_scales": [
193
+ 0.35,
194
+ 0.4,
195
+ 0.45
196
+ ],
197
+ "retrieval_residual_anchor": "expert",
198
+ "retrieval_residual_direction": "candidate_minus_anchor",
199
+ "retrieval_residual_reduce": "compose_mean_by_type",
200
+ "retrieval_residual_challenger_types": [
201
+ "residual_near_miss"
202
+ ],
203
+ "retrieval_residual_challenger_scales": [],
204
+ "retrieval_residual_challenger_margin": 0.01,
205
+ "retrieval_residual_challenger_type_margins": {},
206
+ "retrieval_residual_challenger_tasks": [],
207
+ "retrieval_residual_challenger_type_tasks": {},
208
+ "lattice_exclude_types": [
209
+ "residual_random_negative",
210
+ "residual_wrong_direction",
211
+ "residual_near_miss+residual_no_op"
212
+ ],
213
+ "lattice_exclude_type_tasks": {},
214
+ "candidate_type_bonuses": {
215
+ "retrieval_residual_no_op": 0.03
216
+ },
217
+ "candidate_type_bonus_components": false,
218
+ "candidate_oracle_rollouts": 0,
219
+ "candidate_oracle_unique_tolerance": 1e-06,
220
+ "candidate_oracle_success_rate": null,
221
+ "candidate_oracle_progress": null,
222
+ "candidate_oracle_score": null,
223
+ "candidate_oracle_regret": null,
224
+ "candidate_oracle_score_gain_over_selected": null,
225
+ "candidate_oracle_unique_count": null,
226
+ "candidate_oracle_improvement_rate": null,
227
+ "candidate_oracle_selected_branch_success_rate": null,
228
+ "candidate_oracle_selected_branch_progress": null,
229
+ "candidate_oracle_type_counts": {},
230
+ "candidate_oracle_best_branch_rank": null,
231
+ "candidate_oracle_best_branch_rank_counts": {},
232
+ "candidate_oracle_branch_success_rates": [],
233
+ "candidate_oracle_branch_progress": [],
234
+ "candidate_oracle_branch_score_gains_over_selected": [],
235
+ "selected_residual_scale_counts": {
236
+ "0.35": 521,
237
+ "0.45": 49,
238
+ "0.4": 5
239
+ },
240
+ "policy_rollout_success_rate": 0.3408695652173913,
241
+ "policy_rollout_progress": 0.5687438898404007,
242
+ "oracle_success_rate": 0.8695652173913043,
243
+ "action_mse_to_best": 0.3999333341209137,
244
+ "best_policy_val": {
245
+ "bc_loss": 0.1315989200439718,
246
+ "field_effect_loss": 0.010110370814800262,
247
+ "field_potential_loss": 0.31546536915832096,
248
+ "field_preference_loss": 0.4960637440284093,
249
+ "lattice_edges": 3833.3333333333335,
250
+ "progress_mae": 0.19023226698239645,
251
+ "rank_acc": 0.8255469501018524,
252
+ "rank_loss": 0.4960531195004781,
253
+ "regret_mae": 0.42156532737943864,
254
+ "success_accuracy": 0.8807473679383596,
255
+ "total_loss": 1.561984618504842
256
+ },
257
+ "per_task": {
258
+ "LiftPegUpright-v1": {
259
+ "action_mse_to_best": 0.3281630301287612,
260
+ "expert_success_rate": 0.8584070796460177,
261
+ "num_groups": 113,
262
+ "oracle_success_rate": 0.9380530973451328,
263
+ "policy_expert_regret": 0.987445328615408,
264
+ "policy_oracle_regret": 1.0666473587002374,
265
+ "policy_rollout_progress": 0.6354643670858535,
266
+ "policy_rollout_success_rate": 0.23008849557522124,
267
+ "restore_max_error": 2.384185791015625e-07
268
+ },
269
+ "PickCube-v1": {
270
+ "action_mse_to_best": 0.3249870876619673,
271
+ "expert_success_rate": 0.9402173913043478,
272
+ "num_groups": 184,
273
+ "oracle_success_rate": 0.9456521739130435,
274
+ "policy_expert_regret": 1.0982505323102132,
275
+ "policy_oracle_regret": 1.0982745367851432,
276
+ "policy_rollout_progress": 0.5632556647406486,
277
+ "policy_rollout_success_rate": 0.27717391304347827,
278
+ "restore_max_error": 2.384185791015625e-07
279
+ },
280
+ "PullCube-v1": {
281
+ "action_mse_to_best": 0.5930280594253227,
282
+ "expert_success_rate": 0.25,
283
+ "num_groups": 76,
284
+ "oracle_success_rate": 0.40789473684210525,
285
+ "policy_expert_regret": 0.2989812125696948,
286
+ "policy_oracle_regret": 0.5224439605678383,
287
+ "policy_rollout_progress": 0.2907984899847131,
288
+ "policy_rollout_success_rate": 0.18421052631578946,
289
+ "restore_max_error": 2.384185791015625e-07
290
+ },
291
+ "PushCube-v1": {
292
+ "action_mse_to_best": 0.3595503710948669,
293
+ "expert_success_rate": 0.8198198198198198,
294
+ "num_groups": 111,
295
+ "oracle_success_rate": 1.0,
296
+ "policy_expert_regret": 0.3128753544779511,
297
+ "policy_oracle_regret": 0.33004292161078064,
298
+ "policy_rollout_progress": 0.8411282495603906,
299
+ "policy_rollout_success_rate": 0.8288288288288288,
300
+ "restore_max_error": 2.3795291781425476e-07
301
+ },
302
+ "StackCube-v1": {
303
+ "action_mse_to_best": 0.5285867788716332,
304
+ "expert_success_rate": 0.6923076923076923,
305
+ "num_groups": 91,
306
+ "oracle_success_rate": 0.8571428571428571,
307
+ "policy_expert_regret": 1.1450610632424827,
308
+ "policy_oracle_regret": 1.3047229103989653,
309
+ "policy_rollout_progress": 0.3968714277495395,
310
+ "policy_rollout_success_rate": 0.14285714285714285,
311
+ "restore_max_error": 1.9371509552001953e-07
312
+ }
313
+ }
314
+ },
315
+ {
316
+ "seed": 2,
317
+ "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_typesuccessbonus0p02_nmchallenger0p01.json",
318
+ "num_groups": 575,
319
+ "selection_mode": "retrieval_residual",
320
+ "num_candidates": 48,
321
+ "candidate_sigma": 0.0,
322
+ "proposal_types": [],
323
+ "proposal_lattice_types": [],
324
+ "selection_margin": 0.2,
325
+ "prepend_policy_candidate": false,
326
+ "field_optim_steps": 0,
327
+ "field_optim_step_size": 0.0,
328
+ "field_optim_trust_radius": 0.0,
329
+ "field_optim_l2_penalty": 0.0,
330
+ "retrieval_neighbors": 4,
331
+ "retrieval_metric": "raw",
332
+ "retrieval_type_min_success": 0.0,
333
+ "retrieval_type_success_bonus_scale": 0.02,
334
+ "retrieval_residual_consensus_penalty_scale": 0.0,
335
+ "retrieval_residual_min_source_progress": 0.0,
336
+ "retrieval_residual_min_source_advantage": -1000000000.0,
337
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
338
+ "retrieval_residual_source_score_bonus_scale": 0.0,
339
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
340
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
341
+ "retrieval_residual_scale": 1.0,
342
+ "retrieval_residual_scales": [
343
+ 0.35,
344
+ 0.4,
345
+ 0.45
346
+ ],
347
+ "retrieval_residual_anchor": "expert",
348
+ "retrieval_residual_direction": "candidate_minus_anchor",
349
+ "retrieval_residual_reduce": "compose_mean_by_type",
350
+ "retrieval_residual_challenger_types": [
351
+ "residual_near_miss"
352
+ ],
353
+ "retrieval_residual_challenger_scales": [],
354
+ "retrieval_residual_challenger_margin": 0.01,
355
+ "retrieval_residual_challenger_type_margins": {},
356
+ "retrieval_residual_challenger_tasks": [],
357
+ "retrieval_residual_challenger_type_tasks": {},
358
+ "lattice_exclude_types": [
359
+ "residual_random_negative",
360
+ "residual_wrong_direction",
361
+ "residual_near_miss+residual_no_op"
362
+ ],
363
+ "lattice_exclude_type_tasks": {},
364
+ "candidate_type_bonuses": {
365
+ "retrieval_residual_no_op": 0.03
366
+ },
367
+ "candidate_type_bonus_components": false,
368
+ "candidate_oracle_rollouts": 0,
369
+ "candidate_oracle_unique_tolerance": 1e-06,
370
+ "candidate_oracle_success_rate": null,
371
+ "candidate_oracle_progress": null,
372
+ "candidate_oracle_score": null,
373
+ "candidate_oracle_regret": null,
374
+ "candidate_oracle_score_gain_over_selected": null,
375
+ "candidate_oracle_unique_count": null,
376
+ "candidate_oracle_improvement_rate": null,
377
+ "candidate_oracle_selected_branch_success_rate": null,
378
+ "candidate_oracle_selected_branch_progress": null,
379
+ "candidate_oracle_type_counts": {},
380
+ "candidate_oracle_best_branch_rank": null,
381
+ "candidate_oracle_best_branch_rank_counts": {},
382
+ "candidate_oracle_branch_success_rates": [],
383
+ "candidate_oracle_branch_progress": [],
384
+ "candidate_oracle_branch_score_gains_over_selected": [],
385
+ "selected_residual_scale_counts": {
386
+ "0.35": 519,
387
+ "0.45": 46,
388
+ "0.4": 10
389
+ },
390
+ "policy_rollout_success_rate": 0.37565217391304345,
391
+ "policy_rollout_progress": 0.5898459473317084,
392
+ "oracle_success_rate": 0.8765217391304347,
393
+ "action_mse_to_best": 0.4274916356502344,
394
+ "best_policy_val": {
395
+ "bc_loss": 0.11367896075050037,
396
+ "field_effect_loss": 0.009670218582161598,
397
+ "field_potential_loss": 0.2641640139950646,
398
+ "field_preference_loss": 0.5130490180518892,
399
+ "lattice_edges": 3833.3333333333335,
400
+ "progress_mae": 0.2021729110015763,
401
+ "rank_acc": 0.8333857821093665,
402
+ "rank_loss": 0.5130119257503085,
403
+ "regret_mae": 0.3958987047274907,
404
+ "success_accuracy": 0.8680730561415354,
405
+ "total_loss": 1.4394984311527677
406
+ },
407
+ "per_task": {
408
+ "LiftPegUpright-v1": {
409
+ "action_mse_to_best": 0.35602021238203935,
410
+ "expert_success_rate": 0.8229166666666666,
411
+ "num_groups": 96,
412
+ "oracle_success_rate": 0.9270833333333334,
413
+ "policy_expert_regret": 0.8476849501021206,
414
+ "policy_oracle_regret": 0.9539546399998168,
415
+ "policy_rollout_progress": 0.6463205499264101,
416
+ "policy_rollout_success_rate": 0.3333333333333333,
417
+ "restore_max_error": 1.955777406692505e-07
418
+ },
419
+ "PickCube-v1": {
420
+ "action_mse_to_best": 0.3382689572231035,
421
+ "expert_success_rate": 0.9444444444444444,
422
+ "num_groups": 198,
423
+ "oracle_success_rate": 0.9595959595959596,
424
+ "policy_expert_regret": 0.9734908278015527,
425
+ "policy_oracle_regret": 0.980574278330261,
426
+ "policy_rollout_progress": 0.6379473808814179,
427
+ "policy_rollout_success_rate": 0.3434343434343434,
428
+ "restore_max_error": 2.384185791015625e-07
429
+ },
430
+ "PullCube-v1": {
431
+ "action_mse_to_best": 0.6390281232694784,
432
+ "expert_success_rate": 0.24444444444444444,
433
+ "num_groups": 90,
434
+ "oracle_success_rate": 0.4666666666666667,
435
+ "policy_expert_regret": 0.26827634293586017,
436
+ "policy_oracle_regret": 0.45376074814962014,
437
+ "policy_rollout_progress": 0.37135277808540396,
438
+ "policy_rollout_success_rate": 0.26666666666666666,
439
+ "restore_max_error": 2.384185791015625e-07
440
+ },
441
+ "PushCube-v1": {
442
+ "action_mse_to_best": 0.397049247997232,
443
+ "expert_success_rate": 0.8514851485148515,
444
+ "num_groups": 101,
445
+ "oracle_success_rate": 1.0,
446
+ "policy_expert_regret": 0.3614020819711213,
447
+ "policy_oracle_regret": 0.41630433337523204,
448
+ "policy_rollout_progress": 0.8015174488029858,
449
+ "policy_rollout_success_rate": 0.7821782178217822,
450
+ "restore_max_error": 2.384185791015625e-07
451
+ },
452
+ "StackCube-v1": {
453
+ "action_mse_to_best": 0.5226443493117888,
454
+ "expert_success_rate": 0.7666666666666667,
455
+ "num_groups": 90,
456
+ "oracle_success_rate": 0.9111111111111111,
457
+ "policy_expert_regret": 1.1748778588242001,
458
+ "policy_oracle_regret": 1.3369984101918009,
459
+ "policy_rollout_progress": 0.4047339239054256,
460
+ "policy_rollout_success_rate": 0.14444444444444443,
461
+ "restore_max_error": 2.2351741790771484e-07
462
+ }
463
+ }
464
+ }
465
+ ]
466
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_typesuccessbonus0p02_nmchallenger0p01_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_typesuccessbonus0p02_nmchallenger0p01.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: 35.71% +/- 1.75%
11
+ Gain vs h=16 rank checkpoint: +5.97%
12
+ Mean progress: 57.19%
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@0.01 | 0.00 | 0.020 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 35.48% | 55.72% | 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@0.01 | 0.00 | 0.020 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 34.09% | 56.87% | 86.96% | n/a | n/a | 0.400 |
19
+ | 2 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss@0.01 | 0.00 | 0.020 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 37.57% | 58.98% | 87.65% | n/a | n/a | 0.427 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_typesuccessbonus0p05_nmchallenger0p01_summary.json ADDED
@@ -0,0 +1,466 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_typesuccessbonus0p05_nmchallenger0p01.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.35304347826086957,
9
+ "std_success": 0.017650246200160362,
10
+ "mean_progress": 0.5707256394622011,
11
+ "mean_action_mse_to_best": 0.40862887538510606,
12
+ "gain_vs_h4": 0.05634347826086955,
13
+ "gain_vs_h16_rank_checkpoint": 0.0556521739130435,
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_typesuccessbonus0p05_nmchallenger0p01.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 48,
21
+ "candidate_sigma": 0.0,
22
+ "proposal_types": [],
23
+ "proposal_lattice_types": [],
24
+ "selection_margin": 0.2,
25
+ "prepend_policy_candidate": false,
26
+ "field_optim_steps": 0,
27
+ "field_optim_step_size": 0.0,
28
+ "field_optim_trust_radius": 0.0,
29
+ "field_optim_l2_penalty": 0.0,
30
+ "retrieval_neighbors": 4,
31
+ "retrieval_metric": "raw",
32
+ "retrieval_type_min_success": 0.0,
33
+ "retrieval_type_success_bonus_scale": 0.05,
34
+ "retrieval_residual_consensus_penalty_scale": 0.0,
35
+ "retrieval_residual_min_source_progress": 0.0,
36
+ "retrieval_residual_min_source_advantage": -1000000000.0,
37
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
38
+ "retrieval_residual_source_score_bonus_scale": 0.0,
39
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
40
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
41
+ "retrieval_residual_scale": 1.0,
42
+ "retrieval_residual_scales": [
43
+ 0.35,
44
+ 0.4,
45
+ 0.45
46
+ ],
47
+ "retrieval_residual_anchor": "expert",
48
+ "retrieval_residual_direction": "candidate_minus_anchor",
49
+ "retrieval_residual_reduce": "compose_mean_by_type",
50
+ "retrieval_residual_challenger_types": [
51
+ "residual_near_miss"
52
+ ],
53
+ "retrieval_residual_challenger_scales": [],
54
+ "retrieval_residual_challenger_margin": 0.01,
55
+ "retrieval_residual_challenger_type_margins": {},
56
+ "retrieval_residual_challenger_tasks": [],
57
+ "retrieval_residual_challenger_type_tasks": {},
58
+ "lattice_exclude_types": [
59
+ "residual_random_negative",
60
+ "residual_wrong_direction",
61
+ "residual_near_miss+residual_no_op"
62
+ ],
63
+ "lattice_exclude_type_tasks": {},
64
+ "candidate_type_bonuses": {
65
+ "retrieval_residual_no_op": 0.03
66
+ },
67
+ "candidate_type_bonus_components": false,
68
+ "candidate_oracle_rollouts": 0,
69
+ "candidate_oracle_unique_tolerance": 1e-06,
70
+ "candidate_oracle_success_rate": null,
71
+ "candidate_oracle_progress": null,
72
+ "candidate_oracle_score": null,
73
+ "candidate_oracle_regret": null,
74
+ "candidate_oracle_score_gain_over_selected": null,
75
+ "candidate_oracle_unique_count": null,
76
+ "candidate_oracle_improvement_rate": null,
77
+ "candidate_oracle_selected_branch_success_rate": null,
78
+ "candidate_oracle_selected_branch_progress": null,
79
+ "candidate_oracle_type_counts": {},
80
+ "candidate_oracle_best_branch_rank": null,
81
+ "candidate_oracle_best_branch_rank_counts": {},
82
+ "candidate_oracle_branch_success_rates": [],
83
+ "candidate_oracle_branch_progress": [],
84
+ "candidate_oracle_branch_score_gains_over_selected": [],
85
+ "selected_residual_scale_counts": {
86
+ "0.35": 512,
87
+ "0.45": 57,
88
+ "0.4": 6
89
+ },
90
+ "policy_rollout_success_rate": 0.34956521739130436,
91
+ "policy_rollout_progress": 0.5553485355777261,
92
+ "oracle_success_rate": 0.8573913043478261,
93
+ "action_mse_to_best": 0.3962073888224752,
94
+ "best_policy_val": {
95
+ "bc_loss": 0.13721593966086706,
96
+ "field_effect_loss": 0.009290305380192068,
97
+ "field_potential_loss": 0.2666468388504452,
98
+ "field_preference_loss": 0.5130573478009965,
99
+ "lattice_edges": 3833.3333333333335,
100
+ "progress_mae": 0.1933159919248687,
101
+ "rank_acc": 0.8265031774838766,
102
+ "rank_loss": 0.5130523675017886,
103
+ "regret_mae": 0.3756548762321472,
104
+ "success_accuracy": 0.8773836526605818,
105
+ "total_loss": 1.5581054819954767
106
+ },
107
+ "per_task": {
108
+ "LiftPegUpright-v1": {
109
+ "action_mse_to_best": 0.3253316712306486,
110
+ "expert_success_rate": 0.8865979381443299,
111
+ "num_groups": 97,
112
+ "oracle_success_rate": 0.9278350515463918,
113
+ "policy_expert_regret": 1.0783287166934652,
114
+ "policy_oracle_regret": 1.104744194401908,
115
+ "policy_rollout_progress": 0.5714681769154736,
116
+ "policy_rollout_success_rate": 0.2268041237113402,
117
+ "restore_max_error": 2.384185791015625e-07
118
+ },
119
+ "PickCube-v1": {
120
+ "action_mse_to_best": 0.30287658573629767,
121
+ "expert_success_rate": 0.9375,
122
+ "num_groups": 208,
123
+ "oracle_success_rate": 0.9471153846153846,
124
+ "policy_expert_regret": 0.9715208464588684,
125
+ "policy_oracle_regret": 0.9849437138853738,
126
+ "policy_rollout_progress": 0.6105534896756021,
127
+ "policy_rollout_success_rate": 0.33653846153846156,
128
+ "restore_max_error": 2.384185791015625e-07
129
+ },
130
+ "PullCube-v1": {
131
+ "action_mse_to_best": 0.6057444390925494,
132
+ "expert_success_rate": 0.19480519480519481,
133
+ "num_groups": 77,
134
+ "oracle_success_rate": 0.36363636363636365,
135
+ "policy_expert_regret": 0.24417919438791932,
136
+ "policy_oracle_regret": 0.41790936232314674,
137
+ "policy_rollout_progress": 0.3217475998408899,
138
+ "policy_rollout_success_rate": 0.22077922077922077,
139
+ "restore_max_error": 2.3096799850463867e-07
140
+ },
141
+ "PushCube-v1": {
142
+ "action_mse_to_best": 0.37055013178193824,
143
+ "expert_success_rate": 0.8617021276595744,
144
+ "num_groups": 94,
145
+ "oracle_success_rate": 0.9893617021276596,
146
+ "policy_expert_regret": 0.5508309629686335,
147
+ "policy_oracle_regret": 0.628381609441118,
148
+ "policy_rollout_progress": 0.7014056246014352,
149
+ "policy_rollout_success_rate": 0.6702127659574468,
150
+ "restore_max_error": 2.384185791015625e-07
151
+ },
152
+ "StackCube-v1": {
153
+ "action_mse_to_best": 0.5231284084118376,
154
+ "expert_success_rate": 0.6767676767676768,
155
+ "num_groups": 99,
156
+ "oracle_success_rate": 0.8585858585858586,
157
+ "policy_expert_regret": 0.9252789386294105,
158
+ "policy_oracle_regret": 1.0729591972900159,
159
+ "policy_rollout_progress": 0.46657752569275673,
160
+ "policy_rollout_success_rate": 0.29292929292929293,
161
+ "restore_max_error": 2.384185791015625e-07
162
+ }
163
+ }
164
+ },
165
+ {
166
+ "seed": 1,
167
+ "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_typesuccessbonus0p05_nmchallenger0p01.json",
168
+ "num_groups": 575,
169
+ "selection_mode": "retrieval_residual",
170
+ "num_candidates": 48,
171
+ "candidate_sigma": 0.0,
172
+ "proposal_types": [],
173
+ "proposal_lattice_types": [],
174
+ "selection_margin": 0.2,
175
+ "prepend_policy_candidate": false,
176
+ "field_optim_steps": 0,
177
+ "field_optim_step_size": 0.0,
178
+ "field_optim_trust_radius": 0.0,
179
+ "field_optim_l2_penalty": 0.0,
180
+ "retrieval_neighbors": 4,
181
+ "retrieval_metric": "raw",
182
+ "retrieval_type_min_success": 0.0,
183
+ "retrieval_type_success_bonus_scale": 0.05,
184
+ "retrieval_residual_consensus_penalty_scale": 0.0,
185
+ "retrieval_residual_min_source_progress": 0.0,
186
+ "retrieval_residual_min_source_advantage": -1000000000.0,
187
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
188
+ "retrieval_residual_source_score_bonus_scale": 0.0,
189
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
190
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
191
+ "retrieval_residual_scale": 1.0,
192
+ "retrieval_residual_scales": [
193
+ 0.35,
194
+ 0.4,
195
+ 0.45
196
+ ],
197
+ "retrieval_residual_anchor": "expert",
198
+ "retrieval_residual_direction": "candidate_minus_anchor",
199
+ "retrieval_residual_reduce": "compose_mean_by_type",
200
+ "retrieval_residual_challenger_types": [
201
+ "residual_near_miss"
202
+ ],
203
+ "retrieval_residual_challenger_scales": [],
204
+ "retrieval_residual_challenger_margin": 0.01,
205
+ "retrieval_residual_challenger_type_margins": {},
206
+ "retrieval_residual_challenger_tasks": [],
207
+ "retrieval_residual_challenger_type_tasks": {},
208
+ "lattice_exclude_types": [
209
+ "residual_random_negative",
210
+ "residual_wrong_direction",
211
+ "residual_near_miss+residual_no_op"
212
+ ],
213
+ "lattice_exclude_type_tasks": {},
214
+ "candidate_type_bonuses": {
215
+ "retrieval_residual_no_op": 0.03
216
+ },
217
+ "candidate_type_bonus_components": false,
218
+ "candidate_oracle_rollouts": 0,
219
+ "candidate_oracle_unique_tolerance": 1e-06,
220
+ "candidate_oracle_success_rate": null,
221
+ "candidate_oracle_progress": null,
222
+ "candidate_oracle_score": null,
223
+ "candidate_oracle_regret": null,
224
+ "candidate_oracle_score_gain_over_selected": null,
225
+ "candidate_oracle_unique_count": null,
226
+ "candidate_oracle_improvement_rate": null,
227
+ "candidate_oracle_selected_branch_success_rate": null,
228
+ "candidate_oracle_selected_branch_progress": null,
229
+ "candidate_oracle_type_counts": {},
230
+ "candidate_oracle_best_branch_rank": null,
231
+ "candidate_oracle_best_branch_rank_counts": {},
232
+ "candidate_oracle_branch_success_rates": [],
233
+ "candidate_oracle_branch_progress": [],
234
+ "candidate_oracle_branch_score_gains_over_selected": [],
235
+ "selected_residual_scale_counts": {
236
+ "0.35": 505,
237
+ "0.45": 65,
238
+ "0.4": 5
239
+ },
240
+ "policy_rollout_success_rate": 0.3373913043478261,
241
+ "policy_rollout_progress": 0.5674013672185981,
242
+ "oracle_success_rate": 0.8695652173913043,
243
+ "action_mse_to_best": 0.4005508915480712,
244
+ "best_policy_val": {
245
+ "bc_loss": 0.1315989200439718,
246
+ "field_effect_loss": 0.010110370814800262,
247
+ "field_potential_loss": 0.31546536915832096,
248
+ "field_preference_loss": 0.4960637440284093,
249
+ "lattice_edges": 3833.3333333333335,
250
+ "progress_mae": 0.19023226698239645,
251
+ "rank_acc": 0.8255469501018524,
252
+ "rank_loss": 0.4960531195004781,
253
+ "regret_mae": 0.42156532737943864,
254
+ "success_accuracy": 0.8807473679383596,
255
+ "total_loss": 1.561984618504842
256
+ },
257
+ "per_task": {
258
+ "LiftPegUpright-v1": {
259
+ "action_mse_to_best": 0.32962725897509176,
260
+ "expert_success_rate": 0.8584070796460177,
261
+ "num_groups": 113,
262
+ "oracle_success_rate": 0.9380530973451328,
263
+ "policy_expert_regret": 0.987445328615408,
264
+ "policy_oracle_regret": 1.0666473587002374,
265
+ "policy_rollout_progress": 0.6354643670858535,
266
+ "policy_rollout_success_rate": 0.23008849557522124,
267
+ "restore_max_error": 2.384185791015625e-07
268
+ },
269
+ "PickCube-v1": {
270
+ "action_mse_to_best": 0.325074275408912,
271
+ "expert_success_rate": 0.9402173913043478,
272
+ "num_groups": 184,
273
+ "oracle_success_rate": 0.9456521739130435,
274
+ "policy_expert_regret": 1.1134865133334761,
275
+ "policy_oracle_regret": 1.1135105178084062,
276
+ "policy_rollout_progress": 0.5588892489347769,
277
+ "policy_rollout_success_rate": 0.266304347826087,
278
+ "restore_max_error": 2.384185791015625e-07
279
+ },
280
+ "PullCube-v1": {
281
+ "action_mse_to_best": 0.5935916217338097,
282
+ "expert_success_rate": 0.25,
283
+ "num_groups": 76,
284
+ "oracle_success_rate": 0.40789473684210525,
285
+ "policy_expert_regret": 0.2985671336125386,
286
+ "policy_oracle_regret": 0.5220298816106821,
287
+ "policy_rollout_progress": 0.29121256894186925,
288
+ "policy_rollout_success_rate": 0.18421052631578946,
289
+ "restore_max_error": 2.384185791015625e-07
290
+ },
291
+ "PushCube-v1": {
292
+ "action_mse_to_best": 0.3607284285732218,
293
+ "expert_success_rate": 0.8198198198198198,
294
+ "num_groups": 111,
295
+ "oracle_success_rate": 1.0,
296
+ "policy_expert_regret": 0.3128753544779511,
297
+ "policy_oracle_regret": 0.33004292161078064,
298
+ "policy_rollout_progress": 0.8411282495603906,
299
+ "policy_rollout_success_rate": 0.8288288288288288,
300
+ "restore_max_error": 2.3795291781425476e-07
301
+ },
302
+ "StackCube-v1": {
303
+ "action_mse_to_best": 0.5285867788716332,
304
+ "expert_success_rate": 0.6923076923076923,
305
+ "num_groups": 91,
306
+ "oracle_success_rate": 0.8571428571428571,
307
+ "policy_expert_regret": 1.1450610632424827,
308
+ "policy_oracle_regret": 1.3047229103989653,
309
+ "policy_rollout_progress": 0.3968714277495395,
310
+ "policy_rollout_success_rate": 0.14285714285714285,
311
+ "restore_max_error": 1.9371509552001953e-07
312
+ }
313
+ }
314
+ },
315
+ {
316
+ "seed": 2,
317
+ "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_typesuccessbonus0p05_nmchallenger0p01.json",
318
+ "num_groups": 575,
319
+ "selection_mode": "retrieval_residual",
320
+ "num_candidates": 48,
321
+ "candidate_sigma": 0.0,
322
+ "proposal_types": [],
323
+ "proposal_lattice_types": [],
324
+ "selection_margin": 0.2,
325
+ "prepend_policy_candidate": false,
326
+ "field_optim_steps": 0,
327
+ "field_optim_step_size": 0.0,
328
+ "field_optim_trust_radius": 0.0,
329
+ "field_optim_l2_penalty": 0.0,
330
+ "retrieval_neighbors": 4,
331
+ "retrieval_metric": "raw",
332
+ "retrieval_type_min_success": 0.0,
333
+ "retrieval_type_success_bonus_scale": 0.05,
334
+ "retrieval_residual_consensus_penalty_scale": 0.0,
335
+ "retrieval_residual_min_source_progress": 0.0,
336
+ "retrieval_residual_min_source_advantage": -1000000000.0,
337
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
338
+ "retrieval_residual_source_score_bonus_scale": 0.0,
339
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
340
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
341
+ "retrieval_residual_scale": 1.0,
342
+ "retrieval_residual_scales": [
343
+ 0.35,
344
+ 0.4,
345
+ 0.45
346
+ ],
347
+ "retrieval_residual_anchor": "expert",
348
+ "retrieval_residual_direction": "candidate_minus_anchor",
349
+ "retrieval_residual_reduce": "compose_mean_by_type",
350
+ "retrieval_residual_challenger_types": [
351
+ "residual_near_miss"
352
+ ],
353
+ "retrieval_residual_challenger_scales": [],
354
+ "retrieval_residual_challenger_margin": 0.01,
355
+ "retrieval_residual_challenger_type_margins": {},
356
+ "retrieval_residual_challenger_tasks": [],
357
+ "retrieval_residual_challenger_type_tasks": {},
358
+ "lattice_exclude_types": [
359
+ "residual_random_negative",
360
+ "residual_wrong_direction",
361
+ "residual_near_miss+residual_no_op"
362
+ ],
363
+ "lattice_exclude_type_tasks": {},
364
+ "candidate_type_bonuses": {
365
+ "retrieval_residual_no_op": 0.03
366
+ },
367
+ "candidate_type_bonus_components": false,
368
+ "candidate_oracle_rollouts": 0,
369
+ "candidate_oracle_unique_tolerance": 1e-06,
370
+ "candidate_oracle_success_rate": null,
371
+ "candidate_oracle_progress": null,
372
+ "candidate_oracle_score": null,
373
+ "candidate_oracle_regret": null,
374
+ "candidate_oracle_score_gain_over_selected": null,
375
+ "candidate_oracle_unique_count": null,
376
+ "candidate_oracle_improvement_rate": null,
377
+ "candidate_oracle_selected_branch_success_rate": null,
378
+ "candidate_oracle_selected_branch_progress": null,
379
+ "candidate_oracle_type_counts": {},
380
+ "candidate_oracle_best_branch_rank": null,
381
+ "candidate_oracle_best_branch_rank_counts": {},
382
+ "candidate_oracle_branch_success_rates": [],
383
+ "candidate_oracle_branch_progress": [],
384
+ "candidate_oracle_branch_score_gains_over_selected": [],
385
+ "selected_residual_scale_counts": {
386
+ "0.35": 509,
387
+ "0.45": 56,
388
+ "0.4": 10
389
+ },
390
+ "policy_rollout_success_rate": 0.37217391304347824,
391
+ "policy_rollout_progress": 0.589427015590279,
392
+ "oracle_success_rate": 0.8765217391304347,
393
+ "action_mse_to_best": 0.4291283457847717,
394
+ "best_policy_val": {
395
+ "bc_loss": 0.11367896075050037,
396
+ "field_effect_loss": 0.009670218582161598,
397
+ "field_potential_loss": 0.2641640139950646,
398
+ "field_preference_loss": 0.5130490180518892,
399
+ "lattice_edges": 3833.3333333333335,
400
+ "progress_mae": 0.2021729110015763,
401
+ "rank_acc": 0.8333857821093665,
402
+ "rank_loss": 0.5130119257503085,
403
+ "regret_mae": 0.3958987047274907,
404
+ "success_accuracy": 0.8680730561415354,
405
+ "total_loss": 1.4394984311527677
406
+ },
407
+ "per_task": {
408
+ "LiftPegUpright-v1": {
409
+ "action_mse_to_best": 0.35602021238203935,
410
+ "expert_success_rate": 0.8229166666666666,
411
+ "num_groups": 96,
412
+ "oracle_success_rate": 0.9270833333333334,
413
+ "policy_expert_regret": 0.8476849501021206,
414
+ "policy_oracle_regret": 0.9539546399998168,
415
+ "policy_rollout_progress": 0.6463205499264101,
416
+ "policy_rollout_success_rate": 0.3333333333333333,
417
+ "restore_max_error": 1.955777406692505e-07
418
+ },
419
+ "PickCube-v1": {
420
+ "action_mse_to_best": 0.34098692140494935,
421
+ "expert_success_rate": 0.9444444444444444,
422
+ "num_groups": 198,
423
+ "oracle_success_rate": 0.9595959595959596,
424
+ "policy_expert_regret": 0.9879075713229902,
425
+ "policy_oracle_regret": 0.9949910218516985,
426
+ "policy_rollout_progress": 0.6336316474609904,
427
+ "policy_rollout_success_rate": 0.3333333333333333,
428
+ "restore_max_error": 2.384185791015625e-07
429
+ },
430
+ "PullCube-v1": {
431
+ "action_mse_to_best": 0.6388417792816957,
432
+ "expert_success_rate": 0.24444444444444444,
433
+ "num_groups": 90,
434
+ "oracle_success_rate": 0.4666666666666667,
435
+ "policy_expert_regret": 0.26818666950696046,
436
+ "policy_oracle_regret": 0.4536436658559574,
437
+ "policy_rollout_progress": 0.3714698603790667,
438
+ "policy_rollout_success_rate": 0.26666666666666666,
439
+ "restore_max_error": 2.384185791015625e-07
440
+ },
441
+ "PushCube-v1": {
442
+ "action_mse_to_best": 0.3985377462341054,
443
+ "expert_success_rate": 0.8514851485148515,
444
+ "num_groups": 101,
445
+ "oracle_success_rate": 1.0,
446
+ "policy_expert_regret": 0.3614020819711213,
447
+ "policy_oracle_regret": 0.41630433337523204,
448
+ "policy_rollout_progress": 0.8015174488029858,
449
+ "policy_rollout_success_rate": 0.7821782178217822,
450
+ "restore_max_error": 2.384185791015625e-07
451
+ },
452
+ "StackCube-v1": {
453
+ "action_mse_to_best": 0.5256375054932303,
454
+ "expert_success_rate": 0.7666666666666667,
455
+ "num_groups": 90,
456
+ "oracle_success_rate": 0.9111111111111111,
457
+ "policy_expert_regret": 1.168176835940944,
458
+ "policy_oracle_regret": 1.3302973873085446,
459
+ "policy_rollout_progress": 0.41143494678868187,
460
+ "policy_rollout_success_rate": 0.14444444444444443,
461
+ "restore_max_error": 2.2351741790771484e-07
462
+ }
463
+ }
464
+ }
465
+ ]
466
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_typesuccessbonus0p05_nmchallenger0p01_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_typesuccessbonus0p05_nmchallenger0p01.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: 35.30% +/- 1.77%
11
+ Gain vs h=16 rank checkpoint: +5.57%
12
+ Mean progress: 57.07%
13
+ Mean action MSE to best: 0.409
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@0.01 | 0.00 | 0.050 | 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% | 55.53% | 85.74% | n/a | n/a | 0.396 |
18
+ | 1 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss@0.01 | 0.00 | 0.050 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 33.74% | 56.74% | 86.96% | n/a | n/a | 0.401 |
19
+ | 2 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type + challenger residual_near_miss@0.01 | 0.00 | 0.050 | 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% | 58.94% | 87.65% | n/a | n/a | 0.429 |