anhtld commited on
Commit
22f2a4b
·
verified ·
1 Parent(s): 727dda5

Auto-sync: 2026-06-29 12:25:28 (part 2)

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,364 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_compbonus_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": 0.3536231884057971,
9
+ "std_success": 0.011579701075616344,
10
+ "mean_progress": 0.5698253496913558,
11
+ "mean_action_mse_to_best": 0.4128576582829481,
12
+ "gain_vs_h4": 0.056923188405797076,
13
+ "gain_vs_h16_rank_checkpoint": 0.05623188405797103,
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_compbonus_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_scale": 1.0,
39
+ "retrieval_residual_scales": [
40
+ 0.35,
41
+ 0.4,
42
+ 0.45
43
+ ],
44
+ "retrieval_residual_anchor": "expert",
45
+ "retrieval_residual_direction": "candidate_minus_anchor",
46
+ "retrieval_residual_reduce": "compose_mean_by_type",
47
+ "candidate_type_bonuses": {
48
+ "residual_no_op": 0.03
49
+ },
50
+ "candidate_type_bonus_components": true,
51
+ "selected_residual_scale_counts": {
52
+ "0.35": 529,
53
+ "0.45": 42,
54
+ "0.4": 4
55
+ },
56
+ "policy_rollout_success_rate": 0.34782608695652173,
57
+ "policy_rollout_progress": 0.5560779854971105,
58
+ "oracle_success_rate": 0.8573913043478261,
59
+ "action_mse_to_best": 0.4081159762341691,
60
+ "best_policy_val": {
61
+ "bc_loss": 0.13721593966086706,
62
+ "field_effect_loss": 0.009290305380192068,
63
+ "field_potential_loss": 0.2666468388504452,
64
+ "field_preference_loss": 0.5130573478009965,
65
+ "lattice_edges": 3833.3333333333335,
66
+ "progress_mae": 0.1933159919248687,
67
+ "rank_acc": 0.8265031774838766,
68
+ "rank_loss": 0.5130523675017886,
69
+ "regret_mae": 0.3756548762321472,
70
+ "success_accuracy": 0.8773836526605818,
71
+ "total_loss": 1.5581054819954767
72
+ },
73
+ "per_task": {
74
+ "LiftPegUpright-v1": {
75
+ "action_mse_to_best": 0.34483776046635256,
76
+ "expert_success_rate": 0.8865979381443299,
77
+ "num_groups": 97,
78
+ "oracle_success_rate": 0.9278350515463918,
79
+ "policy_expert_regret": 1.0918974355631268,
80
+ "policy_oracle_regret": 1.1183129132715697,
81
+ "policy_rollout_progress": 0.5682087363963274,
82
+ "policy_rollout_success_rate": 0.21649484536082475,
83
+ "restore_max_error": 2.384185791015625e-07
84
+ },
85
+ "PickCube-v1": {
86
+ "action_mse_to_best": 0.30273427473954284,
87
+ "expert_success_rate": 0.9375,
88
+ "num_groups": 208,
89
+ "oracle_success_rate": 0.9471153846153846,
90
+ "policy_expert_regret": 0.9611748628891431,
91
+ "policy_oracle_regret": 0.9745977303156486,
92
+ "policy_rollout_progress": 0.616091780937635,
93
+ "policy_rollout_success_rate": 0.34134615384615385,
94
+ "restore_max_error": 2.384185791015625e-07
95
+ },
96
+ "PullCube-v1": {
97
+ "action_mse_to_best": 0.6077609157020395,
98
+ "expert_success_rate": 0.19480519480519481,
99
+ "num_groups": 77,
100
+ "oracle_success_rate": 0.36363636363636365,
101
+ "policy_expert_regret": 0.24365403075617822,
102
+ "policy_oracle_regret": 0.41721006885001605,
103
+ "policy_rollout_progress": 0.3124380540767951,
104
+ "policy_rollout_success_rate": 0.2077922077922078,
105
+ "restore_max_error": 2.3096799850463867e-07
106
+ },
107
+ "PushCube-v1": {
108
+ "action_mse_to_best": 0.3675119048578942,
109
+ "expert_success_rate": 0.8617021276595744,
110
+ "num_groups": 94,
111
+ "oracle_success_rate": 0.9893617021276596,
112
+ "policy_expert_regret": 0.5508309629686335,
113
+ "policy_oracle_regret": 0.628381609441118,
114
+ "policy_rollout_progress": 0.7014056246014352,
115
+ "policy_rollout_success_rate": 0.6702127659574468,
116
+ "restore_max_error": 2.384185791015625e-07
117
+ },
118
+ "StackCube-v1": {
119
+ "action_mse_to_best": 0.5747978268473437,
120
+ "expert_success_rate": 0.6767676767676768,
121
+ "num_groups": 99,
122
+ "oracle_success_rate": 0.8585858585858586,
123
+ "policy_expert_regret": 0.9222438893835954,
124
+ "policy_oracle_regret": 1.0699241480442008,
125
+ "policy_rollout_progress": 0.46961257493857184,
126
+ "policy_rollout_success_rate": 0.29292929292929293,
127
+ "restore_max_error": 2.384185791015625e-07
128
+ }
129
+ }
130
+ },
131
+ {
132
+ "seed": 1,
133
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03.json",
134
+ "num_groups": 575,
135
+ "selection_mode": "retrieval_residual",
136
+ "num_candidates": 48,
137
+ "candidate_sigma": 0.0,
138
+ "selection_margin": 0.2,
139
+ "prepend_policy_candidate": false,
140
+ "field_optim_steps": 0,
141
+ "field_optim_step_size": 0.0,
142
+ "field_optim_trust_radius": 0.0,
143
+ "field_optim_l2_penalty": 0.0,
144
+ "retrieval_neighbors": 4,
145
+ "retrieval_metric": "raw",
146
+ "retrieval_type_min_success": 0.0,
147
+ "retrieval_type_success_bonus_scale": 0.0,
148
+ "retrieval_residual_consensus_penalty_scale": 0.0,
149
+ "retrieval_residual_min_source_progress": 0.0,
150
+ "retrieval_residual_min_source_advantage": -1000000000.0,
151
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
152
+ "retrieval_residual_source_score_bonus_scale": 0.0,
153
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
154
+ "retrieval_residual_scale": 1.0,
155
+ "retrieval_residual_scales": [
156
+ 0.35,
157
+ 0.4,
158
+ 0.45
159
+ ],
160
+ "retrieval_residual_anchor": "expert",
161
+ "retrieval_residual_direction": "candidate_minus_anchor",
162
+ "retrieval_residual_reduce": "compose_mean_by_type",
163
+ "candidate_type_bonuses": {
164
+ "residual_no_op": 0.03
165
+ },
166
+ "candidate_type_bonus_components": true,
167
+ "selected_residual_scale_counts": {
168
+ "0.35": 547,
169
+ "0.45": 25,
170
+ "0.4": 3
171
+ },
172
+ "policy_rollout_success_rate": 0.34608695652173915,
173
+ "policy_rollout_progress": 0.5709259671815063,
174
+ "oracle_success_rate": 0.8695652173913043,
175
+ "action_mse_to_best": 0.4013609097946597,
176
+ "best_policy_val": {
177
+ "bc_loss": 0.1315989200439718,
178
+ "field_effect_loss": 0.010110370814800262,
179
+ "field_potential_loss": 0.31546536915832096,
180
+ "field_preference_loss": 0.4960637440284093,
181
+ "lattice_edges": 3833.3333333333335,
182
+ "progress_mae": 0.19023226698239645,
183
+ "rank_acc": 0.8255469501018524,
184
+ "rank_loss": 0.4960531195004781,
185
+ "regret_mae": 0.42156532737943864,
186
+ "success_accuracy": 0.8807473679383596,
187
+ "total_loss": 1.561984618504842
188
+ },
189
+ "per_task": {
190
+ "LiftPegUpright-v1": {
191
+ "action_mse_to_best": 0.33235782895039406,
192
+ "expert_success_rate": 0.8584070796460177,
193
+ "num_groups": 113,
194
+ "oracle_success_rate": 0.9380530973451328,
195
+ "policy_expert_regret": 0.9873896409979964,
196
+ "policy_oracle_regret": 1.0665916710828258,
197
+ "policy_rollout_progress": 0.6355200547032651,
198
+ "policy_rollout_success_rate": 0.23008849557522124,
199
+ "restore_max_error": 2.384185791015625e-07
200
+ },
201
+ "PickCube-v1": {
202
+ "action_mse_to_best": 0.3277512231801191,
203
+ "expert_success_rate": 0.9402173913043478,
204
+ "num_groups": 184,
205
+ "oracle_success_rate": 0.9456521739130435,
206
+ "policy_expert_regret": 1.0705589705308818,
207
+ "policy_oracle_regret": 1.070582975005812,
208
+ "policy_rollout_progress": 0.5746428786938929,
209
+ "policy_rollout_success_rate": 0.29347826086956524,
210
+ "restore_max_error": 2.384185791015625e-07
211
+ },
212
+ "PullCube-v1": {
213
+ "action_mse_to_best": 0.5966075253917983,
214
+ "expert_success_rate": 0.25,
215
+ "num_groups": 76,
216
+ "oracle_success_rate": 0.40789473684210525,
217
+ "policy_expert_regret": 0.300334617103401,
218
+ "policy_oracle_regret": 0.5236276074459678,
219
+ "policy_rollout_progress": 0.2889758540611518,
220
+ "policy_rollout_success_rate": 0.18421052631578946,
221
+ "restore_max_error": 2.384185791015625e-07
222
+ },
223
+ "PushCube-v1": {
224
+ "action_mse_to_best": 0.3571035131677851,
225
+ "expert_success_rate": 0.8198198198198198,
226
+ "num_groups": 111,
227
+ "oracle_success_rate": 1.0,
228
+ "policy_expert_regret": 0.32961789887767656,
229
+ "policy_oracle_regret": 0.3467854660105061,
230
+ "policy_rollout_progress": 0.8333947141696741,
231
+ "policy_rollout_success_rate": 0.8198198198198198,
232
+ "restore_max_error": 2.3795291781425476e-07
233
+ },
234
+ "StackCube-v1": {
235
+ "action_mse_to_best": 0.5268044121317811,
236
+ "expert_success_rate": 0.6923076923076923,
237
+ "num_groups": 91,
238
+ "oracle_success_rate": 0.8571428571428571,
239
+ "policy_expert_regret": 1.1324226328959832,
240
+ "policy_oracle_regret": 1.292084480052466,
241
+ "policy_rollout_progress": 0.3985208471070279,
242
+ "policy_rollout_success_rate": 0.15384615384615385,
243
+ "restore_max_error": 1.9371509552001953e-07
244
+ }
245
+ }
246
+ },
247
+ {
248
+ "seed": 2,
249
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03.json",
250
+ "num_groups": 575,
251
+ "selection_mode": "retrieval_residual",
252
+ "num_candidates": 48,
253
+ "candidate_sigma": 0.0,
254
+ "selection_margin": 0.2,
255
+ "prepend_policy_candidate": false,
256
+ "field_optim_steps": 0,
257
+ "field_optim_step_size": 0.0,
258
+ "field_optim_trust_radius": 0.0,
259
+ "field_optim_l2_penalty": 0.0,
260
+ "retrieval_neighbors": 4,
261
+ "retrieval_metric": "raw",
262
+ "retrieval_type_min_success": 0.0,
263
+ "retrieval_type_success_bonus_scale": 0.0,
264
+ "retrieval_residual_consensus_penalty_scale": 0.0,
265
+ "retrieval_residual_min_source_progress": 0.0,
266
+ "retrieval_residual_min_source_advantage": -1000000000.0,
267
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
268
+ "retrieval_residual_source_score_bonus_scale": 0.0,
269
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
270
+ "retrieval_residual_scale": 1.0,
271
+ "retrieval_residual_scales": [
272
+ 0.35,
273
+ 0.4,
274
+ 0.45
275
+ ],
276
+ "retrieval_residual_anchor": "expert",
277
+ "retrieval_residual_direction": "candidate_minus_anchor",
278
+ "retrieval_residual_reduce": "compose_mean_by_type",
279
+ "candidate_type_bonuses": {
280
+ "residual_no_op": 0.03
281
+ },
282
+ "candidate_type_bonus_components": true,
283
+ "selected_residual_scale_counts": {
284
+ "0.35": 545,
285
+ "0.45": 25,
286
+ "0.4": 5
287
+ },
288
+ "policy_rollout_success_rate": 0.36695652173913046,
289
+ "policy_rollout_progress": 0.5824720963954504,
290
+ "oracle_success_rate": 0.8765217391304347,
291
+ "action_mse_to_best": 0.4290960888200156,
292
+ "best_policy_val": {
293
+ "bc_loss": 0.11367896075050037,
294
+ "field_effect_loss": 0.009670218582161598,
295
+ "field_potential_loss": 0.2641640139950646,
296
+ "field_preference_loss": 0.5130490180518892,
297
+ "lattice_edges": 3833.3333333333335,
298
+ "progress_mae": 0.2021729110015763,
299
+ "rank_acc": 0.8333857821093665,
300
+ "rank_loss": 0.5130119257503085,
301
+ "regret_mae": 0.3958987047274907,
302
+ "success_accuracy": 0.8680730561415354,
303
+ "total_loss": 1.4394984311527677
304
+ },
305
+ "per_task": {
306
+ "LiftPegUpright-v1": {
307
+ "action_mse_to_best": 0.35442226944239036,
308
+ "expert_success_rate": 0.8229166666666666,
309
+ "num_groups": 96,
310
+ "oracle_success_rate": 0.9270833333333334,
311
+ "policy_expert_regret": 0.8476849501021206,
312
+ "policy_oracle_regret": 0.9539546399998168,
313
+ "policy_rollout_progress": 0.6463205499264101,
314
+ "policy_rollout_success_rate": 0.3333333333333333,
315
+ "restore_max_error": 1.955777406692505e-07
316
+ },
317
+ "PickCube-v1": {
318
+ "action_mse_to_best": 0.34171034747527707,
319
+ "expert_success_rate": 0.9444444444444444,
320
+ "num_groups": 198,
321
+ "oracle_success_rate": 0.9595959595959596,
322
+ "policy_expert_regret": 0.9879422967482094,
323
+ "policy_oracle_regret": 0.9950257472769178,
324
+ "policy_rollout_progress": 0.6335969220357712,
325
+ "policy_rollout_success_rate": 0.3333333333333333,
326
+ "restore_max_error": 2.384185791015625e-07
327
+ },
328
+ "PullCube-v1": {
329
+ "action_mse_to_best": 0.6394667098919551,
330
+ "expert_success_rate": 0.24444444444444444,
331
+ "num_groups": 90,
332
+ "oracle_success_rate": 0.4666666666666667,
333
+ "policy_expert_regret": 0.27350674071521,
334
+ "policy_oracle_regret": 0.5016956209377491,
335
+ "policy_rollout_progress": 0.3376428539485359,
336
+ "policy_rollout_success_rate": 0.23333333333333334,
337
+ "restore_max_error": 2.384185791015625e-07
338
+ },
339
+ "PushCube-v1": {
340
+ "action_mse_to_best": 0.397049247997232,
341
+ "expert_success_rate": 0.8514851485148515,
342
+ "num_groups": 101,
343
+ "oracle_success_rate": 1.0,
344
+ "policy_expert_regret": 0.3614020819711213,
345
+ "policy_oracle_regret": 0.41630433337523204,
346
+ "policy_rollout_progress": 0.8015174488029858,
347
+ "policy_rollout_success_rate": 0.7821782178217822,
348
+ "restore_max_error": 2.384185791015625e-07
349
+ },
350
+ "StackCube-v1": {
351
+ "action_mse_to_best": 0.5265898496326473,
352
+ "expert_success_rate": 0.7666666666666667,
353
+ "num_groups": 90,
354
+ "oracle_success_rate": 0.9111111111111111,
355
+ "policy_expert_regret": 1.1748778588242001,
356
+ "policy_oracle_regret": 1.3369984101918009,
357
+ "policy_rollout_progress": 0.4009041431877348,
358
+ "policy_rollout_success_rate": 0.14444444444444443,
359
+ "restore_max_error": 2.2351741790771484e-07
360
+ }
361
+ }
362
+ }
363
+ ]
364
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_compbonus_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_compbonus_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: 35.36% +/- 1.16%
11
+ Gain vs h=16 rank checkpoint: +5.62%
12
+ Mean progress: 56.98%
13
+ Mean action MSE to best: 0.413
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 | 34.78% | 55.61% | 85.74% | 0.408 |
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 | 34.61% | 57.09% | 86.96% | 0.401 |
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% | 58.25% | 87.65% | 0.429 |