anhtld commited on
Commit
620d8b8
·
verified ·
1 Parent(s): 25c4b77

Auto-sync: 2026-06-29 08:05:55 (part 2)

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_nearmiss_k4_grid025035050_margin0p10_summary.json ADDED
@@ -0,0 +1,355 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_repair_nearmiss_k4_grid025035050_margin0p10.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.34144927536231884,
9
+ "std_success": 0.014791131388026771,
10
+ "mean_progress": 0.5600691433255407,
11
+ "mean_action_mse_to_best": 0.3929642635457002,
12
+ "gain_vs_h4": 0.04474927536231882,
13
+ "gain_vs_h16_rank_checkpoint": 0.04405797101449277,
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_repair_nearmiss_k4_grid025035050_margin0p10.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 18,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.1,
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
+ 0.25,
41
+ 0.35,
42
+ 0.5
43
+ ],
44
+ "retrieval_residual_anchor": "expert",
45
+ "retrieval_residual_direction": "anchor_minus_candidate",
46
+ "retrieval_residual_reduce": "mean_by_type",
47
+ "candidate_type_bonuses": {},
48
+ "selected_residual_scale_counts": {
49
+ "0.5": 19,
50
+ "0.25": 552,
51
+ "0.35": 4
52
+ },
53
+ "policy_rollout_success_rate": 0.3356521739130435,
54
+ "policy_rollout_progress": 0.5459818405164001,
55
+ "oracle_success_rate": 0.8573913043478261,
56
+ "action_mse_to_best": 0.3795984360802433,
57
+ "best_policy_val": {
58
+ "bc_loss": 0.13721593966086706,
59
+ "field_effect_loss": 0.009290305380192068,
60
+ "field_potential_loss": 0.2666468388504452,
61
+ "field_preference_loss": 0.5130573478009965,
62
+ "lattice_edges": 3833.3333333333335,
63
+ "progress_mae": 0.1933159919248687,
64
+ "rank_acc": 0.8265031774838766,
65
+ "rank_loss": 0.5130523675017886,
66
+ "regret_mae": 0.3756548762321472,
67
+ "success_accuracy": 0.8773836526605818,
68
+ "total_loss": 1.5581054819954767
69
+ },
70
+ "per_task": {
71
+ "LiftPegUpright-v1": {
72
+ "action_mse_to_best": 0.31828345447663486,
73
+ "expert_success_rate": 0.8865979381443299,
74
+ "num_groups": 97,
75
+ "oracle_success_rate": 0.9278350515463918,
76
+ "policy_expert_regret": 1.078207813280145,
77
+ "policy_oracle_regret": 1.1046232909885878,
78
+ "policy_rollout_progress": 0.5715890803287939,
79
+ "policy_rollout_success_rate": 0.2268041237113402,
80
+ "restore_max_error": 2.384185791015625e-07
81
+ },
82
+ "PickCube-v1": {
83
+ "action_mse_to_best": 0.3025179534362486,
84
+ "expert_success_rate": 0.9375,
85
+ "num_groups": 208,
86
+ "oracle_success_rate": 0.9471153846153846,
87
+ "policy_expert_regret": 0.9750762461469724,
88
+ "policy_oracle_regret": 0.9884991135734779,
89
+ "policy_rollout_progress": 0.6118057822951903,
90
+ "policy_rollout_success_rate": 0.3317307692307692,
91
+ "restore_max_error": 2.384185791015625e-07
92
+ },
93
+ "PullCube-v1": {
94
+ "action_mse_to_best": 0.607571807968152,
95
+ "expert_success_rate": 0.19480519480519481,
96
+ "num_groups": 77,
97
+ "oracle_success_rate": 0.36363636363636365,
98
+ "policy_expert_regret": 0.26381213830590444,
99
+ "policy_oracle_regret": 0.48201371671922216,
100
+ "policy_rollout_progress": 0.2867596946692989,
101
+ "policy_rollout_success_rate": 0.16883116883116883,
102
+ "restore_max_error": 2.3096799850463867e-07
103
+ },
104
+ "PushCube-v1": {
105
+ "action_mse_to_best": 0.3642956905384013,
106
+ "expert_success_rate": 0.8617021276595744,
107
+ "num_groups": 94,
108
+ "oracle_success_rate": 0.9893617021276596,
109
+ "policy_expert_regret": 0.569931797683239,
110
+ "policy_oracle_regret": 0.6663877770621726,
111
+ "policy_rollout_progress": 0.6846760527250615,
112
+ "policy_rollout_success_rate": 0.648936170212766,
113
+ "restore_max_error": 2.384185791015625e-07
114
+ },
115
+ "StackCube-v1": {
116
+ "action_mse_to_best": 0.43883886083847645,
117
+ "expert_success_rate": 0.6767676767676768,
118
+ "num_groups": 99,
119
+ "oracle_success_rate": 0.8585858585858586,
120
+ "policy_expert_regret": 0.9345249523117085,
121
+ "policy_oracle_regret": 1.0971148104378672,
122
+ "policy_rollout_progress": 0.45252292264591565,
123
+ "policy_rollout_success_rate": 0.2828282828282828,
124
+ "restore_max_error": 2.384185791015625e-07
125
+ }
126
+ }
127
+ },
128
+ {
129
+ "seed": 1,
130
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_repair_nearmiss_k4_grid025035050_margin0p10.json",
131
+ "num_groups": 575,
132
+ "selection_mode": "retrieval_residual",
133
+ "num_candidates": 18,
134
+ "candidate_sigma": 0.0,
135
+ "selection_margin": 0.1,
136
+ "prepend_policy_candidate": false,
137
+ "field_optim_steps": 0,
138
+ "field_optim_step_size": 0.0,
139
+ "field_optim_trust_radius": 0.0,
140
+ "field_optim_l2_penalty": 0.0,
141
+ "retrieval_neighbors": 4,
142
+ "retrieval_metric": "raw",
143
+ "retrieval_type_min_success": 0.0,
144
+ "retrieval_type_success_bonus_scale": 0.0,
145
+ "retrieval_residual_consensus_penalty_scale": 0.0,
146
+ "retrieval_residual_min_source_progress": 0.0,
147
+ "retrieval_residual_min_source_advantage": -1000000000.0,
148
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
149
+ "retrieval_residual_source_score_bonus_scale": 0.0,
150
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
151
+ "retrieval_residual_scale": 1.0,
152
+ "retrieval_residual_scales": [
153
+ 0.25,
154
+ 0.35,
155
+ 0.5
156
+ ],
157
+ "retrieval_residual_anchor": "expert",
158
+ "retrieval_residual_direction": "anchor_minus_candidate",
159
+ "retrieval_residual_reduce": "mean_by_type",
160
+ "candidate_type_bonuses": {},
161
+ "selected_residual_scale_counts": {
162
+ "0.25": 547,
163
+ "0.5": 27,
164
+ "0.35": 1
165
+ },
166
+ "policy_rollout_success_rate": 0.33043478260869563,
167
+ "policy_rollout_progress": 0.5586519980981298,
168
+ "oracle_success_rate": 0.8695652173913043,
169
+ "action_mse_to_best": 0.3867724084287234,
170
+ "best_policy_val": {
171
+ "bc_loss": 0.1315989200439718,
172
+ "field_effect_loss": 0.010110370814800262,
173
+ "field_potential_loss": 0.31546536915832096,
174
+ "field_preference_loss": 0.4960637440284093,
175
+ "lattice_edges": 3833.3333333333335,
176
+ "progress_mae": 0.19023226698239645,
177
+ "rank_acc": 0.8255469501018524,
178
+ "rank_loss": 0.4960531195004781,
179
+ "regret_mae": 0.42156532737943864,
180
+ "success_accuracy": 0.8807473679383596,
181
+ "total_loss": 1.561984618504842
182
+ },
183
+ "per_task": {
184
+ "LiftPegUpright-v1": {
185
+ "action_mse_to_best": 0.3242295700644629,
186
+ "expert_success_rate": 0.8584070796460177,
187
+ "num_groups": 113,
188
+ "oracle_success_rate": 0.9380530973451328,
189
+ "policy_expert_regret": 0.9766897917848772,
190
+ "policy_oracle_regret": 1.0558918218697066,
191
+ "policy_rollout_progress": 0.6373703463942604,
192
+ "policy_rollout_success_rate": 0.23893805309734514,
193
+ "restore_max_error": 2.384185791015625e-07
194
+ },
195
+ "PickCube-v1": {
196
+ "action_mse_to_best": 0.3196238886119555,
197
+ "expert_success_rate": 0.9402173913043478,
198
+ "num_groups": 184,
199
+ "oracle_success_rate": 0.9456521739130435,
200
+ "policy_expert_regret": 1.086398953602285,
201
+ "policy_oracle_regret": 1.086422958077215,
202
+ "policy_rollout_progress": 0.5696724608398812,
203
+ "policy_rollout_success_rate": 0.2826086956521739,
204
+ "restore_max_error": 2.384185791015625e-07
205
+ },
206
+ "PullCube-v1": {
207
+ "action_mse_to_best": 0.5711985409847999,
208
+ "expert_success_rate": 0.25,
209
+ "num_groups": 76,
210
+ "oracle_success_rate": 0.40789473684210525,
211
+ "policy_expert_regret": 0.3723486057041507,
212
+ "policy_oracle_regret": 0.6421201418300992,
213
+ "policy_rollout_progress": 0.2362727933612309,
214
+ "policy_rollout_success_rate": 0.11842105263157894,
215
+ "restore_max_error": 2.384185791015625e-07
216
+ },
217
+ "PushCube-v1": {
218
+ "action_mse_to_best": 0.35825215300192703,
219
+ "expert_success_rate": 0.8198198198198198,
220
+ "num_groups": 111,
221
+ "oracle_success_rate": 1.0,
222
+ "policy_expert_regret": 0.3303554279578699,
223
+ "policy_oracle_regret": 0.3799207206513431,
224
+ "policy_rollout_progress": 0.8182774775468551,
225
+ "policy_rollout_success_rate": 0.8018018018018018,
226
+ "restore_max_error": 2.3795291781425476e-07
227
+ },
228
+ "StackCube-v1": {
229
+ "action_mse_to_best": 0.4809705475447597,
230
+ "expert_success_rate": 0.6923076923076923,
231
+ "num_groups": 91,
232
+ "oracle_success_rate": 0.8571428571428571,
233
+ "policy_expert_regret": 1.1381711190218453,
234
+ "policy_oracle_regret": 1.3104208802783883,
235
+ "policy_rollout_progress": 0.3911734578701166,
236
+ "policy_rollout_success_rate": 0.14285714285714285,
237
+ "restore_max_error": 1.9371509552001953e-07
238
+ }
239
+ }
240
+ },
241
+ {
242
+ "seed": 2,
243
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_repair_nearmiss_k4_grid025035050_margin0p10.json",
244
+ "num_groups": 575,
245
+ "selection_mode": "retrieval_residual",
246
+ "num_candidates": 18,
247
+ "candidate_sigma": 0.0,
248
+ "selection_margin": 0.1,
249
+ "prepend_policy_candidate": false,
250
+ "field_optim_steps": 0,
251
+ "field_optim_step_size": 0.0,
252
+ "field_optim_trust_radius": 0.0,
253
+ "field_optim_l2_penalty": 0.0,
254
+ "retrieval_neighbors": 4,
255
+ "retrieval_metric": "raw",
256
+ "retrieval_type_min_success": 0.0,
257
+ "retrieval_type_success_bonus_scale": 0.0,
258
+ "retrieval_residual_consensus_penalty_scale": 0.0,
259
+ "retrieval_residual_min_source_progress": 0.0,
260
+ "retrieval_residual_min_source_advantage": -1000000000.0,
261
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
262
+ "retrieval_residual_source_score_bonus_scale": 0.0,
263
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
264
+ "retrieval_residual_scale": 1.0,
265
+ "retrieval_residual_scales": [
266
+ 0.25,
267
+ 0.35,
268
+ 0.5
269
+ ],
270
+ "retrieval_residual_anchor": "expert",
271
+ "retrieval_residual_direction": "anchor_minus_candidate",
272
+ "retrieval_residual_reduce": "mean_by_type",
273
+ "candidate_type_bonuses": {},
274
+ "selected_residual_scale_counts": {
275
+ "0.25": 537,
276
+ "0.5": 32,
277
+ "0.35": 6
278
+ },
279
+ "policy_rollout_success_rate": 0.3582608695652174,
280
+ "policy_rollout_progress": 0.5755735913620922,
281
+ "oracle_success_rate": 0.8765217391304347,
282
+ "action_mse_to_best": 0.41252194612813387,
283
+ "best_policy_val": {
284
+ "bc_loss": 0.11367896075050037,
285
+ "field_effect_loss": 0.009670218582161598,
286
+ "field_potential_loss": 0.2641640139950646,
287
+ "field_preference_loss": 0.5130490180518892,
288
+ "lattice_edges": 3833.3333333333335,
289
+ "progress_mae": 0.2021729110015763,
290
+ "rank_acc": 0.8333857821093665,
291
+ "rank_loss": 0.5130119257503085,
292
+ "regret_mae": 0.3958987047274907,
293
+ "success_accuracy": 0.8680730561415354,
294
+ "total_loss": 1.4394984311527677
295
+ },
296
+ "per_task": {
297
+ "LiftPegUpright-v1": {
298
+ "action_mse_to_best": 0.35007496579297975,
299
+ "expert_success_rate": 0.8229166666666666,
300
+ "num_groups": 96,
301
+ "oracle_success_rate": 0.9270833333333334,
302
+ "policy_expert_regret": 0.8457174766808748,
303
+ "policy_oracle_regret": 0.951987166578571,
304
+ "policy_rollout_progress": 0.6482880233476559,
305
+ "policy_rollout_success_rate": 0.3333333333333333,
306
+ "restore_max_error": 1.955777406692505e-07
307
+ },
308
+ "PickCube-v1": {
309
+ "action_mse_to_best": 0.33490413741556685,
310
+ "expert_success_rate": 0.9444444444444444,
311
+ "num_groups": 198,
312
+ "oracle_success_rate": 0.9595959595959596,
313
+ "policy_expert_regret": 0.9930475291353885,
314
+ "policy_oracle_regret": 1.000130979664097,
315
+ "policy_rollout_progress": 0.6335421946990971,
316
+ "policy_rollout_success_rate": 0.3282828282828283,
317
+ "restore_max_error": 2.384185791015625e-07
318
+ },
319
+ "PullCube-v1": {
320
+ "action_mse_to_best": 0.6288028371002939,
321
+ "expert_success_rate": 0.24444444444444444,
322
+ "num_groups": 90,
323
+ "oracle_success_rate": 0.4666666666666667,
324
+ "policy_expert_regret": 0.3151884991537574,
325
+ "policy_oracle_regret": 0.5832855659709417,
326
+ "policy_rollout_progress": 0.30096366798022955,
327
+ "policy_rollout_success_rate": 0.18888888888888888,
328
+ "restore_max_error": 2.384185791015625e-07
329
+ },
330
+ "PushCube-v1": {
331
+ "action_mse_to_best": 0.3748256295624346,
332
+ "expert_success_rate": 0.8514851485148515,
333
+ "num_groups": 101,
334
+ "oracle_success_rate": 1.0,
335
+ "policy_expert_regret": 0.3990666143374868,
336
+ "policy_oracle_regret": 0.45396886574159756,
337
+ "policy_rollout_progress": 0.78365489663464,
338
+ "policy_rollout_success_rate": 0.7623762376237624,
339
+ "restore_max_error": 2.384185791015625e-07
340
+ },
341
+ "StackCube-v1": {
342
+ "action_mse_to_best": 0.47591399082707037,
343
+ "expert_success_rate": 0.7666666666666667,
344
+ "num_groups": 90,
345
+ "oracle_success_rate": 0.9111111111111111,
346
+ "policy_expert_regret": 1.167434913251135,
347
+ "policy_oracle_regret": 1.303771203259627,
348
+ "policy_rollout_progress": 0.4115770618120829,
349
+ "policy_rollout_success_rate": 0.16666666666666666,
350
+ "restore_max_error": 2.2351741790771484e-07
351
+ }
352
+ }
353
+ }
354
+ ]
355
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_nearmiss_k4_grid025035050_margin0p10_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_repair_nearmiss_k4_grid025035050_margin0p10.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: 34.14% +/- 1.48%
11
+ Gain vs h=16 rank checkpoint: +4.41%
12
+ Mean progress: 56.01%
13
+ Mean action MSE to best: 0.393
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 | 18 | no | 4 | raw | expert | anchor_minus_candidate | mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.25,0.35,0.50 | 0.100 | 0.00 | 0 | 0.00 | 33.57% | 54.60% | 85.74% | 0.380 |
18
+ | 1 | retrieval_residual | 18 | no | 4 | raw | expert | anchor_minus_candidate | mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.25,0.35,0.50 | 0.100 | 0.00 | 0 | 0.00 | 33.04% | 55.87% | 86.96% | 0.387 |
19
+ | 2 | retrieval_residual | 18 | no | 4 | raw | expert | anchor_minus_candidate | mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.25,0.35,0.50 | 0.100 | 0.00 | 0 | 0.00 | 35.83% | 57.56% | 87.65% | 0.413 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_safe_k4_grid025035050_margin0p20_summary.json ADDED
@@ -0,0 +1,355 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_repair_safe_k4_grid025035050_margin0p20.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3443478260869565,
9
+ "std_success": 0.012541047914657353,
10
+ "mean_progress": 0.5602484929151746,
11
+ "mean_action_mse_to_best": 0.39449909915337744,
12
+ "gain_vs_h4": 0.0476478260869565,
13
+ "gain_vs_h16_rank_checkpoint": 0.04695652173913045,
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_repair_safe_k4_grid025035050_margin0p20.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 18,
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
+ 0.25,
41
+ 0.35,
42
+ 0.5
43
+ ],
44
+ "retrieval_residual_anchor": "expert",
45
+ "retrieval_residual_direction": "anchor_minus_candidate",
46
+ "retrieval_residual_reduce": "mean_by_type",
47
+ "candidate_type_bonuses": {},
48
+ "selected_residual_scale_counts": {
49
+ "0.25": 558,
50
+ "0.5": 13,
51
+ "0.35": 4
52
+ },
53
+ "policy_rollout_success_rate": 0.3408695652173913,
54
+ "policy_rollout_progress": 0.5477940025257514,
55
+ "oracle_success_rate": 0.8573913043478261,
56
+ "action_mse_to_best": 0.38139278126475606,
57
+ "best_policy_val": {
58
+ "bc_loss": 0.13721593966086706,
59
+ "field_effect_loss": 0.009290305380192068,
60
+ "field_potential_loss": 0.2666468388504452,
61
+ "field_preference_loss": 0.5130573478009965,
62
+ "lattice_edges": 3833.3333333333335,
63
+ "progress_mae": 0.1933159919248687,
64
+ "rank_acc": 0.8265031774838766,
65
+ "rank_loss": 0.5130523675017886,
66
+ "regret_mae": 0.3756548762321472,
67
+ "success_accuracy": 0.8773836526605818,
68
+ "total_loss": 1.5581054819954767
69
+ },
70
+ "per_task": {
71
+ "LiftPegUpright-v1": {
72
+ "action_mse_to_best": 0.31969772460726426,
73
+ "expert_success_rate": 0.8865979381443299,
74
+ "num_groups": 97,
75
+ "oracle_success_rate": 0.9278350515463918,
76
+ "policy_expert_regret": 1.0783334703175063,
77
+ "policy_oracle_regret": 1.1047489480259491,
78
+ "policy_rollout_progress": 0.5714634232914325,
79
+ "policy_rollout_success_rate": 0.2268041237113402,
80
+ "restore_max_error": 2.384185791015625e-07
81
+ },
82
+ "PickCube-v1": {
83
+ "action_mse_to_best": 0.3028158242097841,
84
+ "expert_success_rate": 0.9375,
85
+ "num_groups": 208,
86
+ "oracle_success_rate": 0.9471153846153846,
87
+ "policy_expert_regret": 0.9779542152895234,
88
+ "policy_oracle_regret": 0.9913770827160289,
89
+ "policy_rollout_progress": 0.6089278131526393,
90
+ "policy_rollout_success_rate": 0.3317307692307692,
91
+ "restore_max_error": 2.384185791015625e-07
92
+ },
93
+ "PullCube-v1": {
94
+ "action_mse_to_best": 0.6139496880305277,
95
+ "expert_success_rate": 0.19480519480519481,
96
+ "num_groups": 77,
97
+ "oracle_success_rate": 0.36363636363636365,
98
+ "policy_expert_regret": 0.2191846632586139,
99
+ "policy_oracle_regret": 0.4367253573689997,
100
+ "policy_rollout_progress": 0.30607402804549544,
101
+ "policy_rollout_success_rate": 0.19480519480519481,
102
+ "restore_max_error": 2.3096799850463867e-07
103
+ },
104
+ "PushCube-v1": {
105
+ "action_mse_to_best": 0.36453170575042987,
106
+ "expert_success_rate": 0.8617021276595744,
107
+ "num_groups": 94,
108
+ "oracle_success_rate": 0.9893617021276596,
109
+ "policy_expert_regret": 0.5508236038558026,
110
+ "policy_oracle_regret": 0.6472795832347362,
111
+ "policy_rollout_progress": 0.6931459486801573,
112
+ "policy_rollout_success_rate": 0.6595744680851063,
113
+ "restore_max_error": 2.384185791015625e-07
114
+ },
115
+ "StackCube-v1": {
116
+ "action_mse_to_best": 0.4420643655131712,
117
+ "expert_success_rate": 0.6767676767676768,
118
+ "num_groups": 99,
119
+ "oracle_success_rate": 0.8585858585858586,
120
+ "policy_expert_regret": 0.9408943908985214,
121
+ "policy_oracle_regret": 1.1034842490246801,
122
+ "policy_rollout_progress": 0.4461534840591026,
123
+ "policy_rollout_success_rate": 0.2828282828282828,
124
+ "restore_max_error": 2.384185791015625e-07
125
+ }
126
+ }
127
+ },
128
+ {
129
+ "seed": 1,
130
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_repair_safe_k4_grid025035050_margin0p20.json",
131
+ "num_groups": 575,
132
+ "selection_mode": "retrieval_residual",
133
+ "num_candidates": 18,
134
+ "candidate_sigma": 0.0,
135
+ "selection_margin": 0.2,
136
+ "prepend_policy_candidate": false,
137
+ "field_optim_steps": 0,
138
+ "field_optim_step_size": 0.0,
139
+ "field_optim_trust_radius": 0.0,
140
+ "field_optim_l2_penalty": 0.0,
141
+ "retrieval_neighbors": 4,
142
+ "retrieval_metric": "raw",
143
+ "retrieval_type_min_success": 0.0,
144
+ "retrieval_type_success_bonus_scale": 0.0,
145
+ "retrieval_residual_consensus_penalty_scale": 0.0,
146
+ "retrieval_residual_min_source_progress": 0.0,
147
+ "retrieval_residual_min_source_advantage": -1000000000.0,
148
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
149
+ "retrieval_residual_source_score_bonus_scale": 0.0,
150
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
151
+ "retrieval_residual_scale": 1.0,
152
+ "retrieval_residual_scales": [
153
+ 0.25,
154
+ 0.35,
155
+ 0.5
156
+ ],
157
+ "retrieval_residual_anchor": "expert",
158
+ "retrieval_residual_direction": "anchor_minus_candidate",
159
+ "retrieval_residual_reduce": "mean_by_type",
160
+ "candidate_type_bonuses": {},
161
+ "selected_residual_scale_counts": {
162
+ "0.25": 556,
163
+ "0.5": 18,
164
+ "0.35": 1
165
+ },
166
+ "policy_rollout_success_rate": 0.3339130434782609,
167
+ "policy_rollout_progress": 0.5587685327847367,
168
+ "oracle_success_rate": 0.8695652173913043,
169
+ "action_mse_to_best": 0.38822281592728003,
170
+ "best_policy_val": {
171
+ "bc_loss": 0.1315989200439718,
172
+ "field_effect_loss": 0.010110370814800262,
173
+ "field_potential_loss": 0.31546536915832096,
174
+ "field_preference_loss": 0.4960637440284093,
175
+ "lattice_edges": 3833.3333333333335,
176
+ "progress_mae": 0.19023226698239645,
177
+ "rank_acc": 0.8255469501018524,
178
+ "rank_loss": 0.4960531195004781,
179
+ "regret_mae": 0.42156532737943864,
180
+ "success_accuracy": 0.8807473679383596,
181
+ "total_loss": 1.561984618504842
182
+ },
183
+ "per_task": {
184
+ "LiftPegUpright-v1": {
185
+ "action_mse_to_best": 0.3265873762506959,
186
+ "expert_success_rate": 0.8584070796460177,
187
+ "num_groups": 113,
188
+ "oracle_success_rate": 0.9380530973451328,
189
+ "policy_expert_regret": 0.9642004154424751,
190
+ "policy_oracle_regret": 1.0432407254666354,
191
+ "policy_rollout_progress": 0.6436866321394928,
192
+ "policy_rollout_success_rate": 0.25663716814159293,
193
+ "restore_max_error": 2.384185791015625e-07
194
+ },
195
+ "PickCube-v1": {
196
+ "action_mse_to_best": 0.31978633928963024,
197
+ "expert_success_rate": 0.9402173913043478,
198
+ "num_groups": 184,
199
+ "oracle_success_rate": 0.9456521739130435,
200
+ "policy_expert_regret": 1.0868371329371533,
201
+ "policy_oracle_regret": 1.0943740379045028,
202
+ "policy_rollout_progress": 0.567156163621289,
203
+ "policy_rollout_success_rate": 0.27717391304347827,
204
+ "restore_max_error": 2.384185791015625e-07
205
+ },
206
+ "PullCube-v1": {
207
+ "action_mse_to_best": 0.574047543794701,
208
+ "expert_success_rate": 0.25,
209
+ "num_groups": 76,
210
+ "oracle_success_rate": 0.40789473684210525,
211
+ "policy_expert_regret": 0.34984783171431016,
212
+ "policy_oracle_regret": 0.6448580108975109,
213
+ "policy_rollout_progress": 0.23353492429381922,
214
+ "policy_rollout_success_rate": 0.11842105263157894,
215
+ "restore_max_error": 2.384185791015625e-07
216
+ },
217
+ "PushCube-v1": {
218
+ "action_mse_to_best": 0.3581734021101986,
219
+ "expert_success_rate": 0.8198198198198198,
220
+ "num_groups": 111,
221
+ "oracle_success_rate": 1.0,
222
+ "policy_expert_regret": 0.3303554279578699,
223
+ "policy_oracle_regret": 0.3800030283562772,
224
+ "policy_rollout_progress": 0.818195169841921,
225
+ "policy_rollout_success_rate": 0.8018018018018018,
226
+ "restore_max_error": 2.3795291781425476e-07
227
+ },
228
+ "StackCube-v1": {
229
+ "action_mse_to_best": 0.48459558516413304,
230
+ "expert_success_rate": 0.6923076923076923,
231
+ "num_groups": 91,
232
+ "oracle_success_rate": 0.8571428571428571,
233
+ "policy_expert_regret": 1.139402109515536,
234
+ "policy_oracle_regret": 1.2990639566720188,
235
+ "policy_rollout_progress": 0.3915413704874751,
236
+ "policy_rollout_success_rate": 0.15384615384615385,
237
+ "restore_max_error": 1.9371509552001953e-07
238
+ }
239
+ }
240
+ },
241
+ {
242
+ "seed": 2,
243
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_repair_safe_k4_grid025035050_margin0p20.json",
244
+ "num_groups": 575,
245
+ "selection_mode": "retrieval_residual",
246
+ "num_candidates": 18,
247
+ "candidate_sigma": 0.0,
248
+ "selection_margin": 0.2,
249
+ "prepend_policy_candidate": false,
250
+ "field_optim_steps": 0,
251
+ "field_optim_step_size": 0.0,
252
+ "field_optim_trust_radius": 0.0,
253
+ "field_optim_l2_penalty": 0.0,
254
+ "retrieval_neighbors": 4,
255
+ "retrieval_metric": "raw",
256
+ "retrieval_type_min_success": 0.0,
257
+ "retrieval_type_success_bonus_scale": 0.0,
258
+ "retrieval_residual_consensus_penalty_scale": 0.0,
259
+ "retrieval_residual_min_source_progress": 0.0,
260
+ "retrieval_residual_min_source_advantage": -1000000000.0,
261
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
262
+ "retrieval_residual_source_score_bonus_scale": 0.0,
263
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
264
+ "retrieval_residual_scale": 1.0,
265
+ "retrieval_residual_scales": [
266
+ 0.25,
267
+ 0.35,
268
+ 0.5
269
+ ],
270
+ "retrieval_residual_anchor": "expert",
271
+ "retrieval_residual_direction": "anchor_minus_candidate",
272
+ "retrieval_residual_reduce": "mean_by_type",
273
+ "candidate_type_bonuses": {},
274
+ "selected_residual_scale_counts": {
275
+ "0.25": 565,
276
+ "0.5": 9,
277
+ "0.35": 1
278
+ },
279
+ "policy_rollout_success_rate": 0.3582608695652174,
280
+ "policy_rollout_progress": 0.574182943435036,
281
+ "oracle_success_rate": 0.8765217391304347,
282
+ "action_mse_to_best": 0.41388170026809623,
283
+ "best_policy_val": {
284
+ "bc_loss": 0.11367896075050037,
285
+ "field_effect_loss": 0.009670218582161598,
286
+ "field_potential_loss": 0.2641640139950646,
287
+ "field_preference_loss": 0.5130490180518892,
288
+ "lattice_edges": 3833.3333333333335,
289
+ "progress_mae": 0.2021729110015763,
290
+ "rank_acc": 0.8333857821093665,
291
+ "rank_loss": 0.5130119257503085,
292
+ "regret_mae": 0.3958987047274907,
293
+ "success_accuracy": 0.8680730561415354,
294
+ "total_loss": 1.4394984311527677
295
+ },
296
+ "per_task": {
297
+ "LiftPegUpright-v1": {
298
+ "action_mse_to_best": 0.35281265252948896,
299
+ "expert_success_rate": 0.8229166666666666,
300
+ "num_groups": 96,
301
+ "oracle_success_rate": 0.9270833333333334,
302
+ "policy_expert_regret": 0.8479746750866374,
303
+ "policy_oracle_regret": 0.9542443649843335,
304
+ "policy_rollout_progress": 0.6460308249418935,
305
+ "policy_rollout_success_rate": 0.3333333333333333,
306
+ "restore_max_error": 1.955777406692505e-07
307
+ },
308
+ "PickCube-v1": {
309
+ "action_mse_to_best": 0.3358964854381,
310
+ "expert_success_rate": 0.9444444444444444,
311
+ "num_groups": 198,
312
+ "oracle_success_rate": 0.9595959595959596,
313
+ "policy_expert_regret": 0.9954175142703032,
314
+ "policy_oracle_regret": 1.0024487015349095,
315
+ "policy_rollout_progress": 0.6328623286928191,
316
+ "policy_rollout_success_rate": 0.3282828282828283,
317
+ "restore_max_error": 2.384185791015625e-07
318
+ },
319
+ "PullCube-v1": {
320
+ "action_mse_to_best": 0.6294648294647535,
321
+ "expert_success_rate": 0.24444444444444444,
322
+ "num_groups": 90,
323
+ "oracle_success_rate": 0.4666666666666667,
324
+ "policy_expert_regret": 0.3152900996583412,
325
+ "policy_oracle_regret": 0.5833871664755255,
326
+ "policy_rollout_progress": 0.30039575285520387,
327
+ "policy_rollout_success_rate": 0.18888888888888888,
328
+ "restore_max_error": 2.384185791015625e-07
329
+ },
330
+ "PushCube-v1": {
331
+ "action_mse_to_best": 0.3759468028155884,
332
+ "expert_success_rate": 0.8514851485148515,
333
+ "num_groups": 101,
334
+ "oracle_success_rate": 1.0,
335
+ "policy_expert_regret": 0.39930118295815914,
336
+ "policy_oracle_regret": 0.47194982017620957,
337
+ "policy_rollout_progress": 0.775574932299038,
338
+ "policy_rollout_success_rate": 0.7524752475247525,
339
+ "restore_max_error": 2.384185791015625e-07
340
+ },
341
+ "StackCube-v1": {
342
+ "action_mse_to_best": 0.477577746204204,
343
+ "expert_success_rate": 0.7666666666666667,
344
+ "num_groups": 90,
345
+ "oracle_success_rate": 0.9111111111111111,
346
+ "policy_expert_regret": 1.1516696830590567,
347
+ "policy_oracle_regret": 1.2880059730675486,
348
+ "policy_rollout_progress": 0.4162311808930503,
349
+ "policy_rollout_success_rate": 0.17777777777777778,
350
+ "restore_max_error": 2.2351741790771484e-07
351
+ }
352
+ }
353
+ }
354
+ ]
355
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_safe_k4_grid025035050_margin0p20_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_repair_safe_k4_grid025035050_margin0p20.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: 34.43% +/- 1.25%
11
+ Gain vs h=16 rank checkpoint: +4.70%
12
+ Mean progress: 56.02%
13
+ Mean action MSE to best: 0.394
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 | 18 | no | 4 | raw | expert | anchor_minus_candidate | mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.25,0.35,0.50 | 0.200 | 0.00 | 0 | 0.00 | 34.09% | 54.78% | 85.74% | 0.381 |
18
+ | 1 | retrieval_residual | 18 | no | 4 | raw | expert | anchor_minus_candidate | mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.25,0.35,0.50 | 0.200 | 0.00 | 0 | 0.00 | 33.39% | 55.88% | 86.96% | 0.388 |
19
+ | 2 | retrieval_residual | 18 | no | 4 | raw | expert | anchor_minus_candidate | mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.25,0.35,0.50 | 0.200 | 0.00 | 0 | 0.00 | 35.83% | 57.42% | 87.65% | 0.414 |