anhtld commited on
Commit
b329db5
·
verified ·
1 Parent(s): 70cfc10

Auto-sync: 2026-06-29 13:04:05

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,382 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.35594202898550725,
9
+ "std_success": 0.009889114266221456,
10
+ "mean_progress": 0.5706848979810626,
11
+ "mean_action_mse_to_best": 0.4063519324413568,
12
+ "gain_vs_h4": 0.059242028985507234,
13
+ "gain_vs_h16_rank_checkpoint": 0.058550724637681184,
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.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
+ "lattice_exclude_types": [
49
+ "residual_random_negative",
50
+ "residual_wrong_direction",
51
+ "residual_near_miss+residual_no_op"
52
+ ],
53
+ "candidate_type_bonuses": {
54
+ "residual_no_op": 0.03
55
+ },
56
+ "candidate_type_bonus_components": false,
57
+ "selected_residual_scale_counts": {
58
+ "0.35": 536,
59
+ "0.45": 35,
60
+ "0.4": 4
61
+ },
62
+ "policy_rollout_success_rate": 0.35304347826086957,
63
+ "policy_rollout_progress": 0.5583201415569562,
64
+ "oracle_success_rate": 0.8573913043478261,
65
+ "action_mse_to_best": 0.39301921862784933,
66
+ "best_policy_val": {
67
+ "bc_loss": 0.13721593966086706,
68
+ "field_effect_loss": 0.009290305380192068,
69
+ "field_potential_loss": 0.2666468388504452,
70
+ "field_preference_loss": 0.5130573478009965,
71
+ "lattice_edges": 3833.3333333333335,
72
+ "progress_mae": 0.1933159919248687,
73
+ "rank_acc": 0.8265031774838766,
74
+ "rank_loss": 0.5130523675017886,
75
+ "regret_mae": 0.3756548762321472,
76
+ "success_accuracy": 0.8773836526605818,
77
+ "total_loss": 1.5581054819954767
78
+ },
79
+ "per_task": {
80
+ "LiftPegUpright-v1": {
81
+ "action_mse_to_best": 0.3206862742465335,
82
+ "expert_success_rate": 0.8865979381443299,
83
+ "num_groups": 97,
84
+ "oracle_success_rate": 0.9278350515463918,
85
+ "policy_expert_regret": 1.0783334703175063,
86
+ "policy_oracle_regret": 1.1047489480259491,
87
+ "policy_rollout_progress": 0.5714634232914325,
88
+ "policy_rollout_success_rate": 0.2268041237113402,
89
+ "restore_max_error": 2.384185791015625e-07
90
+ },
91
+ "PickCube-v1": {
92
+ "action_mse_to_best": 0.30273104519941485,
93
+ "expert_success_rate": 0.9375,
94
+ "num_groups": 208,
95
+ "oracle_success_rate": 0.9471153846153846,
96
+ "policy_expert_regret": 0.9543058642974267,
97
+ "policy_oracle_regret": 0.9677287317239321,
98
+ "policy_rollout_progress": 0.6181530872216592,
99
+ "policy_rollout_success_rate": 0.34615384615384615,
100
+ "restore_max_error": 2.384185791015625e-07
101
+ },
102
+ "PullCube-v1": {
103
+ "action_mse_to_best": 0.6033564091889889,
104
+ "expert_success_rate": 0.19480519480519481,
105
+ "num_groups": 77,
106
+ "oracle_success_rate": 0.36363636363636365,
107
+ "policy_expert_regret": 0.24368244630296704,
108
+ "policy_oracle_regret": 0.41723848439680483,
109
+ "policy_rollout_progress": 0.31240937069614794,
110
+ "policy_rollout_success_rate": 0.2077922077922078,
111
+ "restore_max_error": 2.3096799850463867e-07
112
+ },
113
+ "PushCube-v1": {
114
+ "action_mse_to_best": 0.3675119048578942,
115
+ "expert_success_rate": 0.8617021276595744,
116
+ "num_groups": 94,
117
+ "oracle_success_rate": 0.9893617021276596,
118
+ "policy_expert_regret": 0.5508309629686335,
119
+ "policy_oracle_regret": 0.628381609441118,
120
+ "policy_rollout_progress": 0.7014056246014352,
121
+ "policy_rollout_success_rate": 0.6702127659574468,
122
+ "restore_max_error": 2.384185791015625e-07
123
+ },
124
+ "StackCube-v1": {
125
+ "action_mse_to_best": 0.5142107287214862,
126
+ "expert_success_rate": 0.6767676767676768,
127
+ "num_groups": 99,
128
+ "oracle_success_rate": 0.8585858585858586,
129
+ "policy_expert_regret": 0.9066177073753241,
130
+ "policy_oracle_regret": 1.0542979660359295,
131
+ "policy_rollout_progress": 0.475137746845833,
132
+ "policy_rollout_success_rate": 0.30303030303030304,
133
+ "restore_max_error": 2.384185791015625e-07
134
+ }
135
+ }
136
+ },
137
+ {
138
+ "seed": 1,
139
+ "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.json",
140
+ "num_groups": 575,
141
+ "selection_mode": "retrieval_residual",
142
+ "num_candidates": 48,
143
+ "candidate_sigma": 0.0,
144
+ "selection_margin": 0.2,
145
+ "prepend_policy_candidate": false,
146
+ "field_optim_steps": 0,
147
+ "field_optim_step_size": 0.0,
148
+ "field_optim_trust_radius": 0.0,
149
+ "field_optim_l2_penalty": 0.0,
150
+ "retrieval_neighbors": 4,
151
+ "retrieval_metric": "raw",
152
+ "retrieval_type_min_success": 0.0,
153
+ "retrieval_type_success_bonus_scale": 0.0,
154
+ "retrieval_residual_consensus_penalty_scale": 0.0,
155
+ "retrieval_residual_min_source_progress": 0.0,
156
+ "retrieval_residual_min_source_advantage": -1000000000.0,
157
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
158
+ "retrieval_residual_source_score_bonus_scale": 0.0,
159
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
160
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
161
+ "retrieval_residual_scale": 1.0,
162
+ "retrieval_residual_scales": [
163
+ 0.35,
164
+ 0.4,
165
+ 0.45
166
+ ],
167
+ "retrieval_residual_anchor": "expert",
168
+ "retrieval_residual_direction": "candidate_minus_anchor",
169
+ "retrieval_residual_reduce": "compose_mean_by_type",
170
+ "lattice_exclude_types": [
171
+ "residual_random_negative",
172
+ "residual_wrong_direction",
173
+ "residual_near_miss+residual_no_op"
174
+ ],
175
+ "candidate_type_bonuses": {
176
+ "residual_no_op": 0.03
177
+ },
178
+ "candidate_type_bonus_components": false,
179
+ "selected_residual_scale_counts": {
180
+ "0.35": 549,
181
+ "0.45": 23,
182
+ "0.4": 3
183
+ },
184
+ "policy_rollout_success_rate": 0.34782608695652173,
185
+ "policy_rollout_progress": 0.5714951352645522,
186
+ "oracle_success_rate": 0.8695652173913043,
187
+ "action_mse_to_best": 0.39824521353549286,
188
+ "best_policy_val": {
189
+ "bc_loss": 0.1315989200439718,
190
+ "field_effect_loss": 0.010110370814800262,
191
+ "field_potential_loss": 0.31546536915832096,
192
+ "field_preference_loss": 0.4960637440284093,
193
+ "lattice_edges": 3833.3333333333335,
194
+ "progress_mae": 0.19023226698239645,
195
+ "rank_acc": 0.8255469501018524,
196
+ "rank_loss": 0.4960531195004781,
197
+ "regret_mae": 0.42156532737943864,
198
+ "success_accuracy": 0.8807473679383596,
199
+ "total_loss": 1.561984618504842
200
+ },
201
+ "per_task": {
202
+ "LiftPegUpright-v1": {
203
+ "action_mse_to_best": 0.3261080000104498,
204
+ "expert_success_rate": 0.8584070796460177,
205
+ "num_groups": 113,
206
+ "oracle_success_rate": 0.9380530973451328,
207
+ "policy_expert_regret": 0.9873896409979964,
208
+ "policy_oracle_regret": 1.0665916710828258,
209
+ "policy_rollout_progress": 0.6355200547032651,
210
+ "policy_rollout_success_rate": 0.23008849557522124,
211
+ "restore_max_error": 2.384185791015625e-07
212
+ },
213
+ "PickCube-v1": {
214
+ "action_mse_to_best": 0.32250052587250655,
215
+ "expert_success_rate": 0.9402173913043478,
216
+ "num_groups": 184,
217
+ "oracle_success_rate": 0.9456521739130435,
218
+ "policy_expert_regret": 1.062856701546875,
219
+ "policy_oracle_regret": 1.062880706021805,
220
+ "policy_rollout_progress": 0.576910365069204,
221
+ "policy_rollout_success_rate": 0.29891304347826086,
222
+ "restore_max_error": 2.384185791015625e-07
223
+ },
224
+ "PullCube-v1": {
225
+ "action_mse_to_best": 0.5971616980080542,
226
+ "expert_success_rate": 0.25,
227
+ "num_groups": 76,
228
+ "oracle_success_rate": 0.40789473684210525,
229
+ "policy_expert_regret": 0.3014062279345174,
230
+ "policy_oracle_regret": 0.5247813826917034,
231
+ "policy_rollout_progress": 0.28782207881541627,
232
+ "policy_rollout_success_rate": 0.18421052631578946,
233
+ "restore_max_error": 2.384185791015625e-07
234
+ },
235
+ "PushCube-v1": {
236
+ "action_mse_to_best": 0.3571035131677851,
237
+ "expert_success_rate": 0.8198198198198198,
238
+ "num_groups": 111,
239
+ "oracle_success_rate": 1.0,
240
+ "policy_expert_regret": 0.32961789887767656,
241
+ "policy_oracle_regret": 0.3467854660105061,
242
+ "policy_rollout_progress": 0.8333947141696741,
243
+ "policy_rollout_success_rate": 0.8198198198198198,
244
+ "restore_max_error": 2.3795291781425476e-07
245
+ },
246
+ "StackCube-v1": {
247
+ "action_mse_to_best": 0.5250320660543966,
248
+ "expert_success_rate": 0.6923076923076923,
249
+ "num_groups": 91,
250
+ "oracle_success_rate": 0.8571428571428571,
251
+ "policy_expert_regret": 1.1324474562655438,
252
+ "policy_oracle_regret": 1.2921093034220266,
253
+ "policy_rollout_progress": 0.39849602373746723,
254
+ "policy_rollout_success_rate": 0.15384615384615385,
255
+ "restore_max_error": 1.9371509552001953e-07
256
+ }
257
+ }
258
+ },
259
+ {
260
+ "seed": 2,
261
+ "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.json",
262
+ "num_groups": 575,
263
+ "selection_mode": "retrieval_residual",
264
+ "num_candidates": 48,
265
+ "candidate_sigma": 0.0,
266
+ "selection_margin": 0.2,
267
+ "prepend_policy_candidate": false,
268
+ "field_optim_steps": 0,
269
+ "field_optim_step_size": 0.0,
270
+ "field_optim_trust_radius": 0.0,
271
+ "field_optim_l2_penalty": 0.0,
272
+ "retrieval_neighbors": 4,
273
+ "retrieval_metric": "raw",
274
+ "retrieval_type_min_success": 0.0,
275
+ "retrieval_type_success_bonus_scale": 0.0,
276
+ "retrieval_residual_consensus_penalty_scale": 0.0,
277
+ "retrieval_residual_min_source_progress": 0.0,
278
+ "retrieval_residual_min_source_advantage": -1000000000.0,
279
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
280
+ "retrieval_residual_source_score_bonus_scale": 0.0,
281
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
282
+ "retrieval_residual_composite_l2_penalty_scale": 0.0,
283
+ "retrieval_residual_scale": 1.0,
284
+ "retrieval_residual_scales": [
285
+ 0.35,
286
+ 0.4,
287
+ 0.45
288
+ ],
289
+ "retrieval_residual_anchor": "expert",
290
+ "retrieval_residual_direction": "candidate_minus_anchor",
291
+ "retrieval_residual_reduce": "compose_mean_by_type",
292
+ "lattice_exclude_types": [
293
+ "residual_random_negative",
294
+ "residual_wrong_direction",
295
+ "residual_near_miss+residual_no_op"
296
+ ],
297
+ "candidate_type_bonuses": {
298
+ "residual_no_op": 0.03
299
+ },
300
+ "candidate_type_bonus_components": false,
301
+ "selected_residual_scale_counts": {
302
+ "0.35": 546,
303
+ "0.45": 25,
304
+ "0.4": 4
305
+ },
306
+ "policy_rollout_success_rate": 0.36695652173913046,
307
+ "policy_rollout_progress": 0.5822394171216793,
308
+ "oracle_success_rate": 0.8765217391304347,
309
+ "action_mse_to_best": 0.4277913651607283,
310
+ "best_policy_val": {
311
+ "bc_loss": 0.11367896075050037,
312
+ "field_effect_loss": 0.009670218582161598,
313
+ "field_potential_loss": 0.2641640139950646,
314
+ "field_preference_loss": 0.5130490180518892,
315
+ "lattice_edges": 3833.3333333333335,
316
+ "progress_mae": 0.2021729110015763,
317
+ "rank_acc": 0.8333857821093665,
318
+ "rank_loss": 0.5130119257503085,
319
+ "regret_mae": 0.3958987047274907,
320
+ "success_accuracy": 0.8680730561415354,
321
+ "total_loss": 1.4394984311527677
322
+ },
323
+ "per_task": {
324
+ "LiftPegUpright-v1": {
325
+ "action_mse_to_best": 0.35442226944239036,
326
+ "expert_success_rate": 0.8229166666666666,
327
+ "num_groups": 96,
328
+ "oracle_success_rate": 0.9270833333333334,
329
+ "policy_expert_regret": 0.8476849501021206,
330
+ "policy_oracle_regret": 0.9539546399998168,
331
+ "policy_rollout_progress": 0.6463205499264101,
332
+ "policy_rollout_success_rate": 0.3333333333333333,
333
+ "restore_max_error": 1.955777406692505e-07
334
+ },
335
+ "PickCube-v1": {
336
+ "action_mse_to_best": 0.3390917487705898,
337
+ "expert_success_rate": 0.9444444444444444,
338
+ "num_groups": 198,
339
+ "oracle_success_rate": 0.9595959595959596,
340
+ "policy_expert_regret": 0.980730776533936,
341
+ "policy_oracle_regret": 0.9878142270626444,
342
+ "policy_rollout_progress": 0.6357579371995397,
343
+ "policy_rollout_success_rate": 0.3383838383838384,
344
+ "restore_max_error": 2.384185791015625e-07
345
+ },
346
+ "PullCube-v1": {
347
+ "action_mse_to_best": 0.6408373928732343,
348
+ "expert_success_rate": 0.24444444444444444,
349
+ "num_groups": 90,
350
+ "oracle_success_rate": 0.4666666666666667,
351
+ "policy_expert_regret": 0.2941978946190122,
352
+ "policy_oracle_regret": 0.522877308153713,
353
+ "policy_rollout_progress": 0.327572277843683,
354
+ "policy_rollout_success_rate": 0.2222222222222222,
355
+ "restore_max_error": 2.384185791015625e-07
356
+ },
357
+ "PushCube-v1": {
358
+ "action_mse_to_best": 0.397049247997232,
359
+ "expert_success_rate": 0.8514851485148515,
360
+ "num_groups": 101,
361
+ "oracle_success_rate": 1.0,
362
+ "policy_expert_regret": 0.3614020819711213,
363
+ "policy_oracle_regret": 0.41630433337523204,
364
+ "policy_rollout_progress": 0.8015174488029858,
365
+ "policy_rollout_success_rate": 0.7821782178217822,
366
+ "restore_max_error": 2.384185791015625e-07
367
+ },
368
+ "StackCube-v1": {
369
+ "action_mse_to_best": 0.5226443493117888,
370
+ "expert_success_rate": 0.7666666666666667,
371
+ "num_groups": 90,
372
+ "oracle_success_rate": 0.9111111111111111,
373
+ "policy_expert_regret": 1.1748778588242001,
374
+ "policy_oracle_regret": 1.3369984101918009,
375
+ "policy_rollout_progress": 0.4047339239054256,
376
+ "policy_rollout_success_rate": 0.14444444444444443,
377
+ "restore_max_error": 2.2351741790771484e-07
378
+ }
379
+ }
380
+ }
381
+ ]
382
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_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_composemasked_dropnmnoop_grid035040045_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: 35.59% +/- 0.99%
11
+ Gain vs h=16 rank checkpoint: +5.86%
12
+ Mean progress: 57.07%
13
+ Mean action MSE to best: 0.406
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 | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---|---:|---:|---:|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type | 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 | 35.30% | 55.83% | 85.74% | 0.393 |
18
+ | 1 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type | 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% | 57.15% | 86.96% | 0.398 |
19
+ | 2 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type | 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 | 36.70% | 58.22% | 87.65% | 0.428 |