anhtld commited on
Commit
7780e0a
·
verified ·
1 Parent(s): 7d9b953

Auto-sync: 2026-06-29 02:18:16 (part 2)

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,328 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
3
+ "objective": "near_miss_policy_bc5",
4
+ "out_name": "policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_noopbonus0p03.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.35188405797101446,
9
+ "std_success": 0.011838016148789656,
10
+ "mean_progress": 0.5656723719785794,
11
+ "mean_action_mse_to_best": 0.39419553005419994,
12
+ "gain_vs_h4": 0.055184057971014444,
13
+ "gain_vs_h16_rank_checkpoint": 0.054492753623188395,
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_k4s040_safe_margin0p20_mean_by_type_nooponly_noopbonus0p03.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 6,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.2,
23
+ "prepend_policy_candidate": false,
24
+ "field_optim_steps": 0,
25
+ "field_optim_step_size": 0.0,
26
+ "field_optim_trust_radius": 0.0,
27
+ "field_optim_l2_penalty": 0.0,
28
+ "retrieval_neighbors": 4,
29
+ "retrieval_metric": "raw",
30
+ "retrieval_type_min_success": 0.0,
31
+ "retrieval_residual_min_source_progress": 0.0,
32
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
33
+ "retrieval_residual_source_score_bonus_scale": 0.0,
34
+ "retrieval_residual_scale": 0.4,
35
+ "retrieval_residual_scales": [],
36
+ "retrieval_residual_anchor": "expert",
37
+ "retrieval_residual_reduce": "mean_by_type",
38
+ "candidate_type_bonuses": {
39
+ "residual_no_op": 0.03
40
+ },
41
+ "selected_residual_scale_counts": {
42
+ "0.4": 575
43
+ },
44
+ "policy_rollout_success_rate": 0.34782608695652173,
45
+ "policy_rollout_progress": 0.5525477739123871,
46
+ "oracle_success_rate": 0.8573913043478261,
47
+ "action_mse_to_best": 0.38096396585845427,
48
+ "best_policy_val": {
49
+ "bc_loss": 0.13721593966086706,
50
+ "field_effect_loss": 0.009290305380192068,
51
+ "field_potential_loss": 0.2666468388504452,
52
+ "field_preference_loss": 0.5130573478009965,
53
+ "lattice_edges": 3833.3333333333335,
54
+ "progress_mae": 0.1933159919248687,
55
+ "rank_acc": 0.8265031774838766,
56
+ "rank_loss": 0.5130523675017886,
57
+ "regret_mae": 0.3756548762321472,
58
+ "success_accuracy": 0.8773836526605818,
59
+ "total_loss": 1.5581054819954767
60
+ },
61
+ "per_task": {
62
+ "LiftPegUpright-v1": {
63
+ "action_mse_to_best": 0.31969772460726426,
64
+ "expert_success_rate": 0.8865979381443299,
65
+ "num_groups": 97,
66
+ "oracle_success_rate": 0.9278350515463918,
67
+ "policy_expert_regret": 1.0783334703175063,
68
+ "policy_oracle_regret": 1.1047489480259491,
69
+ "policy_rollout_progress": 0.5714634232914325,
70
+ "policy_rollout_success_rate": 0.2268041237113402,
71
+ "restore_max_error": 2.384185791015625e-07
72
+ },
73
+ "PickCube-v1": {
74
+ "action_mse_to_best": 0.30265616485849023,
75
+ "expert_success_rate": 0.9375,
76
+ "num_groups": 208,
77
+ "oracle_success_rate": 0.9471153846153846,
78
+ "policy_expert_regret": 0.9611708908700026,
79
+ "policy_oracle_regret": 0.9745937582965081,
80
+ "policy_rollout_progress": 0.6160957529567755,
81
+ "policy_rollout_success_rate": 0.34134615384615385,
82
+ "restore_max_error": 2.384185791015625e-07
83
+ },
84
+ "PullCube-v1": {
85
+ "action_mse_to_best": 0.6013353352809881,
86
+ "expert_success_rate": 0.19480519480519481,
87
+ "num_groups": 77,
88
+ "oracle_success_rate": 0.36363636363636365,
89
+ "policy_expert_regret": 0.24465571066316266,
90
+ "policy_oracle_regret": 0.41821174875700046,
91
+ "policy_rollout_progress": 0.31160062367048164,
92
+ "policy_rollout_success_rate": 0.2077922077922078,
93
+ "restore_max_error": 2.3096799850463867e-07
94
+ },
95
+ "PushCube-v1": {
96
+ "action_mse_to_best": 0.36294203449754003,
97
+ "expert_success_rate": 0.8617021276595744,
98
+ "num_groups": 94,
99
+ "oracle_success_rate": 0.9893617021276596,
100
+ "policy_expert_regret": 0.5508309629686335,
101
+ "policy_oracle_regret": 0.628381609441118,
102
+ "policy_rollout_progress": 0.7014056246014352,
103
+ "policy_rollout_success_rate": 0.6702127659574468,
104
+ "restore_max_error": 2.384185791015625e-07
105
+ },
106
+ "StackCube-v1": {
107
+ "action_mse_to_best": 0.45122976496702794,
108
+ "expert_success_rate": 0.6767676767676768,
109
+ "num_groups": 99,
110
+ "oracle_success_rate": 0.8585858585858586,
111
+ "policy_expert_regret": 0.940485000309318,
112
+ "policy_oracle_regret": 1.1030748584354766,
113
+ "policy_rollout_progress": 0.44656287464830613,
114
+ "policy_rollout_success_rate": 0.2828282828282828,
115
+ "restore_max_error": 2.384185791015625e-07
116
+ }
117
+ }
118
+ },
119
+ {
120
+ "seed": 1,
121
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_noopbonus0p03.json",
122
+ "num_groups": 575,
123
+ "selection_mode": "retrieval_residual",
124
+ "num_candidates": 6,
125
+ "candidate_sigma": 0.0,
126
+ "selection_margin": 0.2,
127
+ "prepend_policy_candidate": false,
128
+ "field_optim_steps": 0,
129
+ "field_optim_step_size": 0.0,
130
+ "field_optim_trust_radius": 0.0,
131
+ "field_optim_l2_penalty": 0.0,
132
+ "retrieval_neighbors": 4,
133
+ "retrieval_metric": "raw",
134
+ "retrieval_type_min_success": 0.0,
135
+ "retrieval_residual_min_source_progress": 0.0,
136
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
137
+ "retrieval_residual_source_score_bonus_scale": 0.0,
138
+ "retrieval_residual_scale": 0.4,
139
+ "retrieval_residual_scales": [],
140
+ "retrieval_residual_anchor": "expert",
141
+ "retrieval_residual_reduce": "mean_by_type",
142
+ "candidate_type_bonuses": {
143
+ "residual_no_op": 0.03
144
+ },
145
+ "selected_residual_scale_counts": {
146
+ "0.4": 575
147
+ },
148
+ "policy_rollout_success_rate": 0.3426086956521739,
149
+ "policy_rollout_progress": 0.5665045300385226,
150
+ "oracle_success_rate": 0.8695652173913043,
151
+ "action_mse_to_best": 0.38766786168612866,
152
+ "best_policy_val": {
153
+ "bc_loss": 0.1315989200439718,
154
+ "field_effect_loss": 0.010110370814800262,
155
+ "field_potential_loss": 0.31546536915832096,
156
+ "field_preference_loss": 0.4960637440284093,
157
+ "lattice_edges": 3833.3333333333335,
158
+ "progress_mae": 0.19023226698239645,
159
+ "rank_acc": 0.8255469501018524,
160
+ "rank_loss": 0.4960531195004781,
161
+ "regret_mae": 0.42156532737943864,
162
+ "success_accuracy": 0.8807473679383596,
163
+ "total_loss": 1.561984618504842
164
+ },
165
+ "per_task": {
166
+ "LiftPegUpright-v1": {
167
+ "action_mse_to_best": 0.3255710795406351,
168
+ "expert_success_rate": 0.8584070796460177,
169
+ "num_groups": 113,
170
+ "oracle_success_rate": 0.9380530973451328,
171
+ "policy_expert_regret": 0.9873926745051831,
172
+ "policy_oracle_regret": 1.0665947045900126,
173
+ "policy_rollout_progress": 0.6355170211960784,
174
+ "policy_rollout_success_rate": 0.23008849557522124,
175
+ "restore_max_error": 2.384185791015625e-07
176
+ },
177
+ "PickCube-v1": {
178
+ "action_mse_to_best": 0.3200199794550629,
179
+ "expert_success_rate": 0.9402173913043478,
180
+ "num_groups": 184,
181
+ "oracle_success_rate": 0.9456521739130435,
182
+ "policy_expert_regret": 1.0707762707429735,
183
+ "policy_oracle_regret": 1.0708002752179038,
184
+ "policy_rollout_progress": 0.5744255784818012,
185
+ "policy_rollout_success_rate": 0.29347826086956524,
186
+ "restore_max_error": 2.384185791015625e-07
187
+ },
188
+ "PullCube-v1": {
189
+ "action_mse_to_best": 0.5715771112591028,
190
+ "expert_success_rate": 0.25,
191
+ "num_groups": 76,
192
+ "oracle_success_rate": 0.40789473684210525,
193
+ "policy_expert_regret": 0.30013844498286124,
194
+ "policy_oracle_regret": 0.5234715515061429,
195
+ "policy_rollout_progress": 0.28913191000097677,
196
+ "policy_rollout_success_rate": 0.18421052631578946,
197
+ "restore_max_error": 2.384185791015625e-07
198
+ },
199
+ "PushCube-v1": {
200
+ "action_mse_to_best": 0.35635087504848706,
201
+ "expert_success_rate": 0.8198198198198198,
202
+ "num_groups": 111,
203
+ "oracle_success_rate": 1.0,
204
+ "policy_expert_regret": 0.33058198494417174,
205
+ "policy_oracle_regret": 0.380407250276557,
206
+ "policy_rollout_progress": 0.8177909479216412,
207
+ "policy_rollout_success_rate": 0.8018018018018018,
208
+ "restore_max_error": 2.3795291781425476e-07
209
+ },
210
+ "StackCube-v1": {
211
+ "action_mse_to_best": 0.48616488654534895,
212
+ "expert_success_rate": 0.6923076923076923,
213
+ "num_groups": 91,
214
+ "oracle_success_rate": 0.8571428571428571,
215
+ "policy_expert_regret": 1.1410143090473426,
216
+ "policy_oracle_regret": 1.3006761562038254,
217
+ "policy_rollout_progress": 0.3899291709556684,
218
+ "policy_rollout_success_rate": 0.15384615384615385,
219
+ "restore_max_error": 1.9371509552001953e-07
220
+ }
221
+ }
222
+ },
223
+ {
224
+ "seed": 2,
225
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_noopbonus0p03.json",
226
+ "num_groups": 575,
227
+ "selection_mode": "retrieval_residual",
228
+ "num_candidates": 6,
229
+ "candidate_sigma": 0.0,
230
+ "selection_margin": 0.2,
231
+ "prepend_policy_candidate": false,
232
+ "field_optim_steps": 0,
233
+ "field_optim_step_size": 0.0,
234
+ "field_optim_trust_radius": 0.0,
235
+ "field_optim_l2_penalty": 0.0,
236
+ "retrieval_neighbors": 4,
237
+ "retrieval_metric": "raw",
238
+ "retrieval_type_min_success": 0.0,
239
+ "retrieval_residual_min_source_progress": 0.0,
240
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
241
+ "retrieval_residual_source_score_bonus_scale": 0.0,
242
+ "retrieval_residual_scale": 0.4,
243
+ "retrieval_residual_scales": [],
244
+ "retrieval_residual_anchor": "expert",
245
+ "retrieval_residual_reduce": "mean_by_type",
246
+ "candidate_type_bonuses": {
247
+ "residual_no_op": 0.03
248
+ },
249
+ "selected_residual_scale_counts": {
250
+ "0.4": 575
251
+ },
252
+ "policy_rollout_success_rate": 0.3652173913043478,
253
+ "policy_rollout_progress": 0.5779648119848286,
254
+ "oracle_success_rate": 0.8765217391304347,
255
+ "action_mse_to_best": 0.41395476261801695,
256
+ "best_policy_val": {
257
+ "bc_loss": 0.11367896075050037,
258
+ "field_effect_loss": 0.009670218582161598,
259
+ "field_potential_loss": 0.2641640139950646,
260
+ "field_preference_loss": 0.5130490180518892,
261
+ "lattice_edges": 3833.3333333333335,
262
+ "progress_mae": 0.2021729110015763,
263
+ "rank_acc": 0.8333857821093665,
264
+ "rank_loss": 0.5130119257503085,
265
+ "regret_mae": 0.3958987047274907,
266
+ "success_accuracy": 0.8680730561415354,
267
+ "total_loss": 1.4394984311527677
268
+ },
269
+ "per_task": {
270
+ "LiftPegUpright-v1": {
271
+ "action_mse_to_best": 0.35361564661434386,
272
+ "expert_success_rate": 0.8229166666666666,
273
+ "num_groups": 96,
274
+ "oracle_success_rate": 0.9270833333333334,
275
+ "policy_expert_regret": 0.8476849501021206,
276
+ "policy_oracle_regret": 0.9539546399998168,
277
+ "policy_rollout_progress": 0.6463205499264101,
278
+ "policy_rollout_success_rate": 0.3333333333333333,
279
+ "restore_max_error": 1.955777406692505e-07
280
+ },
281
+ "PickCube-v1": {
282
+ "action_mse_to_best": 0.33552251840856945,
283
+ "expert_success_rate": 0.9444444444444444,
284
+ "num_groups": 198,
285
+ "oracle_success_rate": 0.9595959595959596,
286
+ "policy_expert_regret": 0.9807514474924767,
287
+ "policy_oracle_regret": 0.9878348980211851,
288
+ "policy_rollout_progress": 0.6357372662409989,
289
+ "policy_rollout_success_rate": 0.3383838383838384,
290
+ "restore_max_error": 2.384185791015625e-07
291
+ },
292
+ "PullCube-v1": {
293
+ "action_mse_to_best": 0.6269181397226121,
294
+ "expert_success_rate": 0.24444444444444444,
295
+ "num_groups": 90,
296
+ "oracle_success_rate": 0.4666666666666667,
297
+ "policy_expert_regret": 0.31323022758242586,
298
+ "policy_oracle_regret": 0.5418931306305239,
299
+ "policy_rollout_progress": 0.3196675664779832,
300
+ "policy_rollout_success_rate": 0.2111111111111111,
301
+ "restore_max_error": 2.384185791015625e-07
302
+ },
303
+ "PushCube-v1": {
304
+ "action_mse_to_best": 0.3759468028155884,
305
+ "expert_success_rate": 0.8514851485148515,
306
+ "num_groups": 101,
307
+ "oracle_success_rate": 1.0,
308
+ "policy_expert_regret": 0.39930118295815914,
309
+ "policy_oracle_regret": 0.47194982017620957,
310
+ "policy_rollout_progress": 0.775574932299038,
311
+ "policy_rollout_success_rate": 0.7524752475247525,
312
+ "restore_max_error": 2.384185791015625e-07
313
+ },
314
+ "StackCube-v1": {
315
+ "action_mse_to_best": 0.4805574236230718,
316
+ "expert_success_rate": 0.7666666666666667,
317
+ "num_groups": 90,
318
+ "oracle_success_rate": 0.9111111111111111,
319
+ "policy_expert_regret": 1.1534134613143072,
320
+ "policy_oracle_regret": 1.2897497513227993,
321
+ "policy_rollout_progress": 0.4144874026377996,
322
+ "policy_rollout_success_rate": 0.17777777777777778,
323
+ "restore_max_error": 2.2351741790771484e-07
324
+ }
325
+ }
326
+ }
327
+ ]
328
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_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_k4s040_safe_margin0p20_mean_by_type_nooponly_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: 35.19% +/- 1.18%
11
+ Gain vs h=16 rank checkpoint: +5.45%
12
+ Mean progress: 56.57%
13
+ Mean action MSE to best: 0.394
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | 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 | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.000 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 34.78% | 55.25% | 85.74% | 0.381 |
18
+ | 1 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.000 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 34.26% | 56.65% | 86.96% | 0.388 |
19
+ | 2 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.000 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 36.52% | 57.80% | 87.65% | 0.414 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_srcscorebonus0p02_summary.json ADDED
@@ -0,0 +1,322 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_k4s040_safe_margin0p20_mean_by_type_nooponly_srcscorebonus0p02.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.35188405797101446,
9
+ "std_success": 0.011838016148789656,
10
+ "mean_progress": 0.5655461518514167,
11
+ "mean_action_mse_to_best": 0.3940786650153282,
12
+ "gain_vs_h4": 0.055184057971014444,
13
+ "gain_vs_h16_rank_checkpoint": 0.054492753623188395,
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_k4s040_safe_margin0p20_mean_by_type_nooponly_srcscorebonus0p02.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 6,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.2,
23
+ "prepend_policy_candidate": false,
24
+ "field_optim_steps": 0,
25
+ "field_optim_step_size": 0.0,
26
+ "field_optim_trust_radius": 0.0,
27
+ "field_optim_l2_penalty": 0.0,
28
+ "retrieval_neighbors": 4,
29
+ "retrieval_metric": "raw",
30
+ "retrieval_type_min_success": 0.0,
31
+ "retrieval_residual_min_source_progress": 0.0,
32
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
33
+ "retrieval_residual_source_score_bonus_scale": 0.02,
34
+ "retrieval_residual_scale": 0.4,
35
+ "retrieval_residual_scales": [],
36
+ "retrieval_residual_anchor": "expert",
37
+ "retrieval_residual_reduce": "mean_by_type",
38
+ "candidate_type_bonuses": {},
39
+ "selected_residual_scale_counts": {
40
+ "0.4": 575
41
+ },
42
+ "policy_rollout_success_rate": 0.34782608695652173,
43
+ "policy_rollout_progress": 0.5521566874882126,
44
+ "oracle_success_rate": 0.8573913043478261,
45
+ "action_mse_to_best": 0.38061467294783696,
46
+ "best_policy_val": {
47
+ "bc_loss": 0.13721593966086706,
48
+ "field_effect_loss": 0.009290305380192068,
49
+ "field_potential_loss": 0.2666468388504452,
50
+ "field_preference_loss": 0.5130573478009965,
51
+ "lattice_edges": 3833.3333333333335,
52
+ "progress_mae": 0.1933159919248687,
53
+ "rank_acc": 0.8265031774838766,
54
+ "rank_loss": 0.5130523675017886,
55
+ "regret_mae": 0.3756548762321472,
56
+ "success_accuracy": 0.8773836526605818,
57
+ "total_loss": 1.5581054819954767
58
+ },
59
+ "per_task": {
60
+ "LiftPegUpright-v1": {
61
+ "action_mse_to_best": 0.31969772460726426,
62
+ "expert_success_rate": 0.8865979381443299,
63
+ "num_groups": 97,
64
+ "oracle_success_rate": 0.9278350515463918,
65
+ "policy_expert_regret": 1.0783334703175063,
66
+ "policy_oracle_regret": 1.1047489480259491,
67
+ "policy_rollout_progress": 0.5714634232914325,
68
+ "policy_rollout_success_rate": 0.2268041237113402,
69
+ "restore_max_error": 2.384185791015625e-07
70
+ },
71
+ "PickCube-v1": {
72
+ "action_mse_to_best": 0.30266214839111155,
73
+ "expert_success_rate": 0.9375,
74
+ "num_groups": 208,
75
+ "oracle_success_rate": 0.9471153846153846,
76
+ "policy_expert_regret": 0.9612141423500501,
77
+ "policy_oracle_regret": 0.9746370097765555,
78
+ "policy_rollout_progress": 0.616052501476728,
79
+ "policy_rollout_success_rate": 0.34134615384615385,
80
+ "restore_max_error": 2.384185791015625e-07
81
+ },
82
+ "PullCube-v1": {
83
+ "action_mse_to_best": 0.5987108158213752,
84
+ "expert_success_rate": 0.19480519480519481,
85
+ "num_groups": 77,
86
+ "oracle_success_rate": 0.36363636363636365,
87
+ "policy_expert_regret": 0.2474593260664154,
88
+ "policy_oracle_regret": 0.42101536416025326,
89
+ "policy_rollout_progress": 0.3087970082672289,
90
+ "policy_rollout_success_rate": 0.2077922077922078,
91
+ "restore_max_error": 2.3096799850463867e-07
92
+ },
93
+ "PushCube-v1": {
94
+ "action_mse_to_best": 0.36294203449754003,
95
+ "expert_success_rate": 0.8617021276595744,
96
+ "num_groups": 94,
97
+ "oracle_success_rate": 0.9893617021276596,
98
+ "policy_expert_regret": 0.5508309629686335,
99
+ "policy_oracle_regret": 0.628381609441118,
100
+ "policy_rollout_progress": 0.7014056246014352,
101
+ "policy_rollout_success_rate": 0.6702127659574468,
102
+ "restore_max_error": 2.384185791015625e-07
103
+ },
104
+ "StackCube-v1": {
105
+ "action_mse_to_best": 0.45122976496702794,
106
+ "expert_success_rate": 0.6767676767676768,
107
+ "num_groups": 99,
108
+ "oracle_success_rate": 0.8585858585858586,
109
+ "policy_expert_regret": 0.940485000309318,
110
+ "policy_oracle_regret": 1.1030748584354766,
111
+ "policy_rollout_progress": 0.44656287464830613,
112
+ "policy_rollout_success_rate": 0.2828282828282828,
113
+ "restore_max_error": 2.384185791015625e-07
114
+ }
115
+ }
116
+ },
117
+ {
118
+ "seed": 1,
119
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_srcscorebonus0p02.json",
120
+ "num_groups": 575,
121
+ "selection_mode": "retrieval_residual",
122
+ "num_candidates": 6,
123
+ "candidate_sigma": 0.0,
124
+ "selection_margin": 0.2,
125
+ "prepend_policy_candidate": false,
126
+ "field_optim_steps": 0,
127
+ "field_optim_step_size": 0.0,
128
+ "field_optim_trust_radius": 0.0,
129
+ "field_optim_l2_penalty": 0.0,
130
+ "retrieval_neighbors": 4,
131
+ "retrieval_metric": "raw",
132
+ "retrieval_type_min_success": 0.0,
133
+ "retrieval_residual_min_source_progress": 0.0,
134
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
135
+ "retrieval_residual_source_score_bonus_scale": 0.02,
136
+ "retrieval_residual_scale": 0.4,
137
+ "retrieval_residual_scales": [],
138
+ "retrieval_residual_anchor": "expert",
139
+ "retrieval_residual_reduce": "mean_by_type",
140
+ "candidate_type_bonuses": {},
141
+ "selected_residual_scale_counts": {
142
+ "0.4": 575
143
+ },
144
+ "policy_rollout_success_rate": 0.3426086956521739,
145
+ "policy_rollout_progress": 0.5664698526813932,
146
+ "oracle_success_rate": 0.8695652173913043,
147
+ "action_mse_to_best": 0.38757041037244644,
148
+ "best_policy_val": {
149
+ "bc_loss": 0.1315989200439718,
150
+ "field_effect_loss": 0.010110370814800262,
151
+ "field_potential_loss": 0.31546536915832096,
152
+ "field_preference_loss": 0.4960637440284093,
153
+ "lattice_edges": 3833.3333333333335,
154
+ "progress_mae": 0.19023226698239645,
155
+ "rank_acc": 0.8255469501018524,
156
+ "rank_loss": 0.4960531195004781,
157
+ "regret_mae": 0.42156532737943864,
158
+ "success_accuracy": 0.8807473679383596,
159
+ "total_loss": 1.561984618504842
160
+ },
161
+ "per_task": {
162
+ "LiftPegUpright-v1": {
163
+ "action_mse_to_best": 0.3255710795406351,
164
+ "expert_success_rate": 0.8584070796460177,
165
+ "num_groups": 113,
166
+ "oracle_success_rate": 0.9380530973451328,
167
+ "policy_expert_regret": 0.9873926745051831,
168
+ "policy_oracle_regret": 1.0665947045900126,
169
+ "policy_rollout_progress": 0.6355170211960784,
170
+ "policy_rollout_success_rate": 0.23008849557522124,
171
+ "restore_max_error": 2.384185791015625e-07
172
+ },
173
+ "PickCube-v1": {
174
+ "action_mse_to_best": 0.3197154440998059,
175
+ "expert_success_rate": 0.9402173913043478,
176
+ "num_groups": 184,
177
+ "oracle_success_rate": 0.9456521739130435,
178
+ "policy_expert_regret": 1.070884637484003,
179
+ "policy_oracle_regret": 1.070908641958933,
180
+ "policy_rollout_progress": 0.5743172117407717,
181
+ "policy_rollout_success_rate": 0.29347826086956524,
182
+ "restore_max_error": 2.384185791015625e-07
183
+ },
184
+ "PullCube-v1": {
185
+ "action_mse_to_best": 0.5715771112591028,
186
+ "expert_success_rate": 0.25,
187
+ "num_groups": 76,
188
+ "oracle_success_rate": 0.40789473684210525,
189
+ "policy_expert_regret": 0.30013844498286124,
190
+ "policy_oracle_regret": 0.5234715515061429,
191
+ "policy_rollout_progress": 0.28913191000097677,
192
+ "policy_rollout_success_rate": 0.18421052631578946,
193
+ "restore_max_error": 2.384185791015625e-07
194
+ },
195
+ "PushCube-v1": {
196
+ "action_mse_to_best": 0.35635087504848706,
197
+ "expert_success_rate": 0.8198198198198198,
198
+ "num_groups": 111,
199
+ "oracle_success_rate": 1.0,
200
+ "policy_expert_regret": 0.33058198494417174,
201
+ "policy_oracle_regret": 0.380407250276557,
202
+ "policy_rollout_progress": 0.8177909479216412,
203
+ "policy_rollout_success_rate": 0.8018018018018018,
204
+ "restore_max_error": 2.3795291781425476e-07
205
+ },
206
+ "StackCube-v1": {
207
+ "action_mse_to_best": 0.48616488654534895,
208
+ "expert_success_rate": 0.6923076923076923,
209
+ "num_groups": 91,
210
+ "oracle_success_rate": 0.8571428571428571,
211
+ "policy_expert_regret": 1.1410143090473426,
212
+ "policy_oracle_regret": 1.3006761562038254,
213
+ "policy_rollout_progress": 0.3899291709556684,
214
+ "policy_rollout_success_rate": 0.15384615384615385,
215
+ "restore_max_error": 1.9371509552001953e-07
216
+ }
217
+ }
218
+ },
219
+ {
220
+ "seed": 2,
221
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_srcscorebonus0p02.json",
222
+ "num_groups": 575,
223
+ "selection_mode": "retrieval_residual",
224
+ "num_candidates": 6,
225
+ "candidate_sigma": 0.0,
226
+ "selection_margin": 0.2,
227
+ "prepend_policy_candidate": false,
228
+ "field_optim_steps": 0,
229
+ "field_optim_step_size": 0.0,
230
+ "field_optim_trust_radius": 0.0,
231
+ "field_optim_l2_penalty": 0.0,
232
+ "retrieval_neighbors": 4,
233
+ "retrieval_metric": "raw",
234
+ "retrieval_type_min_success": 0.0,
235
+ "retrieval_residual_min_source_progress": 0.0,
236
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
237
+ "retrieval_residual_source_score_bonus_scale": 0.02,
238
+ "retrieval_residual_scale": 0.4,
239
+ "retrieval_residual_scales": [],
240
+ "retrieval_residual_anchor": "expert",
241
+ "retrieval_residual_reduce": "mean_by_type",
242
+ "candidate_type_bonuses": {},
243
+ "selected_residual_scale_counts": {
244
+ "0.4": 575
245
+ },
246
+ "policy_rollout_success_rate": 0.3652173913043478,
247
+ "policy_rollout_progress": 0.5780119153846444,
248
+ "oracle_success_rate": 0.8765217391304347,
249
+ "action_mse_to_best": 0.4140509117257012,
250
+ "best_policy_val": {
251
+ "bc_loss": 0.11367896075050037,
252
+ "field_effect_loss": 0.009670218582161598,
253
+ "field_potential_loss": 0.2641640139950646,
254
+ "field_preference_loss": 0.5130490180518892,
255
+ "lattice_edges": 3833.3333333333335,
256
+ "progress_mae": 0.2021729110015763,
257
+ "rank_acc": 0.8333857821093665,
258
+ "rank_loss": 0.5130119257503085,
259
+ "regret_mae": 0.3958987047274907,
260
+ "success_accuracy": 0.8680730561415354,
261
+ "total_loss": 1.4394984311527677
262
+ },
263
+ "per_task": {
264
+ "LiftPegUpright-v1": {
265
+ "action_mse_to_best": 0.35361564661434386,
266
+ "expert_success_rate": 0.8229166666666666,
267
+ "num_groups": 96,
268
+ "oracle_success_rate": 0.9270833333333334,
269
+ "policy_expert_regret": 0.8476849501021206,
270
+ "policy_oracle_regret": 0.9539546399998168,
271
+ "policy_rollout_progress": 0.6463205499264101,
272
+ "policy_rollout_success_rate": 0.3333333333333333,
273
+ "restore_max_error": 1.955777406692505e-07
274
+ },
275
+ "PickCube-v1": {
276
+ "action_mse_to_best": 0.33545475296977195,
277
+ "expert_success_rate": 0.9444444444444444,
278
+ "num_groups": 198,
279
+ "oracle_success_rate": 0.9595959595959596,
280
+ "policy_expert_regret": 0.9806582634572429,
281
+ "policy_oracle_regret": 0.9877417139859512,
282
+ "policy_rollout_progress": 0.6358304502762328,
283
+ "policy_rollout_success_rate": 0.3383838383838384,
284
+ "restore_max_error": 2.384185791015625e-07
285
+ },
286
+ "PullCube-v1": {
287
+ "action_mse_to_best": 0.6276815096537273,
288
+ "expert_success_rate": 0.24444444444444444,
289
+ "num_groups": 90,
290
+ "oracle_success_rate": 0.4666666666666667,
291
+ "policy_expert_regret": 0.31323022758242586,
292
+ "policy_oracle_regret": 0.5417971971203264,
293
+ "policy_rollout_progress": 0.3197634999881807,
294
+ "policy_rollout_success_rate": 0.2111111111111111,
295
+ "restore_max_error": 2.384185791015625e-07
296
+ },
297
+ "PushCube-v1": {
298
+ "action_mse_to_best": 0.3759468028155884,
299
+ "expert_success_rate": 0.8514851485148515,
300
+ "num_groups": 101,
301
+ "oracle_success_rate": 1.0,
302
+ "policy_expert_regret": 0.39930118295815914,
303
+ "policy_oracle_regret": 0.47194982017620957,
304
+ "policy_rollout_progress": 0.775574932299038,
305
+ "policy_rollout_success_rate": 0.7524752475247525,
306
+ "restore_max_error": 2.384185791015625e-07
307
+ },
308
+ "StackCube-v1": {
309
+ "action_mse_to_best": 0.4805574236230718,
310
+ "expert_success_rate": 0.7666666666666667,
311
+ "num_groups": 90,
312
+ "oracle_success_rate": 0.9111111111111111,
313
+ "policy_expert_regret": 1.1534134613143072,
314
+ "policy_oracle_regret": 1.2897497513227993,
315
+ "policy_rollout_progress": 0.4144874026377996,
316
+ "policy_rollout_success_rate": 0.17777777777777778,
317
+ "restore_max_error": 2.2351741790771484e-07
318
+ }
319
+ }
320
+ }
321
+ ]
322
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_nooponly_srcscorebonus0p02_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_k4s040_safe_margin0p20_mean_by_type_nooponly_srcscorebonus0p02.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: 35.19% +/- 1.18%
11
+ Gain vs h=16 rank checkpoint: +5.45%
12
+ Mean progress: 56.55%
13
+ Mean action MSE to best: 0.394
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | 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 | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.020 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 34.78% | 55.22% | 85.74% | 0.381 |
18
+ | 1 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.020 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 34.26% | 56.65% | 86.96% | 0.388 |
19
+ | 2 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.020 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 36.52% | 57.80% | 87.65% | 0.414 |
results/paper_analysis.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "best_clean_key": "residual_k4_consensus_noopbonus003",
3
- "generated_utc": "2026-06-29T06:12:41+00:00",
4
  "mechanism_gap": {
5
  "best_clean_vs_direct_same_ckpt": 0.06956521739130428,
6
  "best_clean_vs_h16": 0.05507246376811592,
 
1
  {
2
  "best_clean_key": "residual_k4_consensus_noopbonus003",
3
+ "generated_utc": "2026-06-29T06:15:24+00:00",
4
  "mechanism_gap": {
5
  "best_clean_vs_direct_same_ckpt": 0.06956521739130428,
6
  "best_clean_vs_h16": 0.05507246376811592,