anhtld commited on
Commit
a6cbea9
·
verified ·
1 Parent(s): 8a40d3f

Auto-sync: 2026-06-29 12:50:13 (part 2)

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,367 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_l2comp002_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": 0.3553623188405797,
9
+ "std_success": 0.010626262481057023,
10
+ "mean_progress": 0.5703304209737425,
11
+ "mean_action_mse_to_best": 0.40788532738533356,
12
+ "gain_vs_h4": 0.05866231884057971,
13
+ "gain_vs_h16_rank_checkpoint": 0.05797101449275366,
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_l2comp002_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": 0.02,
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
+ "candidate_type_bonuses": {
49
+ "residual_no_op": 0.03
50
+ },
51
+ "candidate_type_bonus_components": false,
52
+ "selected_residual_scale_counts": {
53
+ "0.35": 532,
54
+ "0.45": 39,
55
+ "0.4": 4
56
+ },
57
+ "policy_rollout_success_rate": 0.35304347826086957,
58
+ "policy_rollout_progress": 0.5580391604931134,
59
+ "oracle_success_rate": 0.8573913043478261,
60
+ "action_mse_to_best": 0.40025916028443886,
61
+ "best_policy_val": {
62
+ "bc_loss": 0.13721593966086706,
63
+ "field_effect_loss": 0.009290305380192068,
64
+ "field_potential_loss": 0.2666468388504452,
65
+ "field_preference_loss": 0.5130573478009965,
66
+ "lattice_edges": 3833.3333333333335,
67
+ "progress_mae": 0.1933159919248687,
68
+ "rank_acc": 0.8265031774838766,
69
+ "rank_loss": 0.5130523675017886,
70
+ "regret_mae": 0.3756548762321472,
71
+ "success_accuracy": 0.8773836526605818,
72
+ "total_loss": 1.5581054819954767
73
+ },
74
+ "per_task": {
75
+ "LiftPegUpright-v1": {
76
+ "action_mse_to_best": 0.3327169909631622,
77
+ "expert_success_rate": 0.8865979381443299,
78
+ "num_groups": 97,
79
+ "oracle_success_rate": 0.9278350515463918,
80
+ "policy_expert_regret": 1.0783523666797226,
81
+ "policy_oracle_regret": 1.1047678443881654,
82
+ "policy_rollout_progress": 0.5714445269292163,
83
+ "policy_rollout_success_rate": 0.2268041237113402,
84
+ "restore_max_error": 2.384185791015625e-07
85
+ },
86
+ "PickCube-v1": {
87
+ "action_mse_to_best": 0.30273104519941485,
88
+ "expert_success_rate": 0.9375,
89
+ "num_groups": 208,
90
+ "oracle_success_rate": 0.9471153846153846,
91
+ "policy_expert_regret": 0.9543058642974267,
92
+ "policy_oracle_regret": 0.9677287317239321,
93
+ "policy_rollout_progress": 0.6181530872216592,
94
+ "policy_rollout_success_rate": 0.34615384615384615,
95
+ "restore_max_error": 2.384185791015625e-07
96
+ },
97
+ "PullCube-v1": {
98
+ "action_mse_to_best": 0.6033577046224049,
99
+ "expert_success_rate": 0.19480519480519481,
100
+ "num_groups": 77,
101
+ "oracle_success_rate": 0.36363636363636365,
102
+ "policy_expert_regret": 0.24368244630296704,
103
+ "policy_oracle_regret": 0.41723848439680483,
104
+ "policy_rollout_progress": 0.31240937069614794,
105
+ "policy_rollout_success_rate": 0.2077922077922078,
106
+ "restore_max_error": 2.3096799850463867e-07
107
+ },
108
+ "PushCube-v1": {
109
+ "action_mse_to_best": 0.3675119048578942,
110
+ "expert_success_rate": 0.8617021276595744,
111
+ "num_groups": 94,
112
+ "oracle_success_rate": 0.9893617021276596,
113
+ "policy_expert_regret": 0.5508309629686335,
114
+ "policy_oracle_regret": 0.628381609441118,
115
+ "policy_rollout_progress": 0.7014056246014352,
116
+ "policy_rollout_success_rate": 0.6702127659574468,
117
+ "restore_max_error": 2.384185791015625e-07
118
+ },
119
+ "StackCube-v1": {
120
+ "action_mse_to_best": 0.5444722154149504,
121
+ "expert_success_rate": 0.6767676767676768,
122
+ "num_groups": 99,
123
+ "oracle_success_rate": 0.8585858585858586,
124
+ "policy_expert_regret": 0.9082311534821385,
125
+ "policy_oracle_regret": 1.055911412142744,
126
+ "policy_rollout_progress": 0.4735243007390186,
127
+ "policy_rollout_success_rate": 0.30303030303030304,
128
+ "restore_max_error": 2.384185791015625e-07
129
+ }
130
+ }
131
+ },
132
+ {
133
+ "seed": 1,
134
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03.json",
135
+ "num_groups": 575,
136
+ "selection_mode": "retrieval_residual",
137
+ "num_candidates": 48,
138
+ "candidate_sigma": 0.0,
139
+ "selection_margin": 0.2,
140
+ "prepend_policy_candidate": false,
141
+ "field_optim_steps": 0,
142
+ "field_optim_step_size": 0.0,
143
+ "field_optim_trust_radius": 0.0,
144
+ "field_optim_l2_penalty": 0.0,
145
+ "retrieval_neighbors": 4,
146
+ "retrieval_metric": "raw",
147
+ "retrieval_type_min_success": 0.0,
148
+ "retrieval_type_success_bonus_scale": 0.0,
149
+ "retrieval_residual_consensus_penalty_scale": 0.0,
150
+ "retrieval_residual_min_source_progress": 0.0,
151
+ "retrieval_residual_min_source_advantage": -1000000000.0,
152
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
153
+ "retrieval_residual_source_score_bonus_scale": 0.0,
154
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
155
+ "retrieval_residual_composite_l2_penalty_scale": 0.02,
156
+ "retrieval_residual_scale": 1.0,
157
+ "retrieval_residual_scales": [
158
+ 0.35,
159
+ 0.4,
160
+ 0.45
161
+ ],
162
+ "retrieval_residual_anchor": "expert",
163
+ "retrieval_residual_direction": "candidate_minus_anchor",
164
+ "retrieval_residual_reduce": "compose_mean_by_type",
165
+ "candidate_type_bonuses": {
166
+ "residual_no_op": 0.03
167
+ },
168
+ "candidate_type_bonus_components": false,
169
+ "selected_residual_scale_counts": {
170
+ "0.35": 549,
171
+ "0.45": 23,
172
+ "0.4": 3
173
+ },
174
+ "policy_rollout_success_rate": 0.34608695652173915,
175
+ "policy_rollout_progress": 0.5702176707529504,
176
+ "oracle_success_rate": 0.8695652173913043,
177
+ "action_mse_to_best": 0.3954629489504125,
178
+ "best_policy_val": {
179
+ "bc_loss": 0.1315989200439718,
180
+ "field_effect_loss": 0.010110370814800262,
181
+ "field_potential_loss": 0.31546536915832096,
182
+ "field_preference_loss": 0.4960637440284093,
183
+ "lattice_edges": 3833.3333333333335,
184
+ "progress_mae": 0.19023226698239645,
185
+ "rank_acc": 0.8255469501018524,
186
+ "rank_loss": 0.4960531195004781,
187
+ "regret_mae": 0.42156532737943864,
188
+ "success_accuracy": 0.8807473679383596,
189
+ "total_loss": 1.561984618504842
190
+ },
191
+ "per_task": {
192
+ "LiftPegUpright-v1": {
193
+ "action_mse_to_best": 0.3261080000104498,
194
+ "expert_success_rate": 0.8584070796460177,
195
+ "num_groups": 113,
196
+ "oracle_success_rate": 0.9380530973451328,
197
+ "policy_expert_regret": 0.9873896409979964,
198
+ "policy_oracle_regret": 1.0665916710828258,
199
+ "policy_rollout_progress": 0.6355200547032651,
200
+ "policy_rollout_success_rate": 0.23008849557522124,
201
+ "restore_max_error": 2.384185791015625e-07
202
+ },
203
+ "PickCube-v1": {
204
+ "action_mse_to_best": 0.3202790089316018,
205
+ "expert_success_rate": 0.9402173913043478,
206
+ "num_groups": 184,
207
+ "oracle_success_rate": 0.9456521739130435,
208
+ "policy_expert_regret": 1.062856701546875,
209
+ "policy_oracle_regret": 1.062880706021805,
210
+ "policy_rollout_progress": 0.576910365069204,
211
+ "policy_rollout_success_rate": 0.29891304347826086,
212
+ "restore_max_error": 2.384185791015625e-07
213
+ },
214
+ "PullCube-v1": {
215
+ "action_mse_to_best": 0.5793679281088867,
216
+ "expert_success_rate": 0.25,
217
+ "num_groups": 76,
218
+ "oracle_success_rate": 0.40789473684210525,
219
+ "policy_expert_regret": 0.32425887294505773,
220
+ "policy_oracle_regret": 0.5476340277022437,
221
+ "policy_rollout_progress": 0.278127328541718,
222
+ "policy_rollout_success_rate": 0.17105263157894737,
223
+ "restore_max_error": 2.384185791015625e-07
224
+ },
225
+ "PushCube-v1": {
226
+ "action_mse_to_best": 0.3571035131677851,
227
+ "expert_success_rate": 0.8198198198198198,
228
+ "num_groups": 111,
229
+ "oracle_success_rate": 1.0,
230
+ "policy_expert_regret": 0.32961789887767656,
231
+ "policy_oracle_regret": 0.3467854660105061,
232
+ "policy_rollout_progress": 0.8333947141696741,
233
+ "policy_rollout_success_rate": 0.8198198198198198,
234
+ "restore_max_error": 2.3795291781425476e-07
235
+ },
236
+ "StackCube-v1": {
237
+ "action_mse_to_best": 0.5268044121317811,
238
+ "expert_success_rate": 0.6923076923076923,
239
+ "num_groups": 91,
240
+ "oracle_success_rate": 0.8571428571428571,
241
+ "policy_expert_regret": 1.1324226328959832,
242
+ "policy_oracle_regret": 1.292084480052466,
243
+ "policy_rollout_progress": 0.3985208471070279,
244
+ "policy_rollout_success_rate": 0.15384615384615385,
245
+ "restore_max_error": 1.9371509552001953e-07
246
+ }
247
+ }
248
+ },
249
+ {
250
+ "seed": 2,
251
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03.json",
252
+ "num_groups": 575,
253
+ "selection_mode": "retrieval_residual",
254
+ "num_candidates": 48,
255
+ "candidate_sigma": 0.0,
256
+ "selection_margin": 0.2,
257
+ "prepend_policy_candidate": false,
258
+ "field_optim_steps": 0,
259
+ "field_optim_step_size": 0.0,
260
+ "field_optim_trust_radius": 0.0,
261
+ "field_optim_l2_penalty": 0.0,
262
+ "retrieval_neighbors": 4,
263
+ "retrieval_metric": "raw",
264
+ "retrieval_type_min_success": 0.0,
265
+ "retrieval_type_success_bonus_scale": 0.0,
266
+ "retrieval_residual_consensus_penalty_scale": 0.0,
267
+ "retrieval_residual_min_source_progress": 0.0,
268
+ "retrieval_residual_min_source_advantage": -1000000000.0,
269
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
270
+ "retrieval_residual_source_score_bonus_scale": 0.0,
271
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
272
+ "retrieval_residual_composite_l2_penalty_scale": 0.02,
273
+ "retrieval_residual_scale": 1.0,
274
+ "retrieval_residual_scales": [
275
+ 0.35,
276
+ 0.4,
277
+ 0.45
278
+ ],
279
+ "retrieval_residual_anchor": "expert",
280
+ "retrieval_residual_direction": "candidate_minus_anchor",
281
+ "retrieval_residual_reduce": "compose_mean_by_type",
282
+ "candidate_type_bonuses": {
283
+ "residual_no_op": 0.03
284
+ },
285
+ "candidate_type_bonus_components": false,
286
+ "selected_residual_scale_counts": {
287
+ "0.35": 547,
288
+ "0.45": 25,
289
+ "0.4": 3
290
+ },
291
+ "policy_rollout_success_rate": 0.36695652173913046,
292
+ "policy_rollout_progress": 0.5827344316751639,
293
+ "oracle_success_rate": 0.8765217391304347,
294
+ "action_mse_to_best": 0.42793387292114937,
295
+ "best_policy_val": {
296
+ "bc_loss": 0.11367896075050037,
297
+ "field_effect_loss": 0.009670218582161598,
298
+ "field_potential_loss": 0.2641640139950646,
299
+ "field_preference_loss": 0.5130490180518892,
300
+ "lattice_edges": 3833.3333333333335,
301
+ "progress_mae": 0.2021729110015763,
302
+ "rank_acc": 0.8333857821093665,
303
+ "rank_loss": 0.5130119257503085,
304
+ "regret_mae": 0.3958987047274907,
305
+ "success_accuracy": 0.8680730561415354,
306
+ "total_loss": 1.4394984311527677
307
+ },
308
+ "per_task": {
309
+ "LiftPegUpright-v1": {
310
+ "action_mse_to_best": 0.35442226944239036,
311
+ "expert_success_rate": 0.8229166666666666,
312
+ "num_groups": 96,
313
+ "oracle_success_rate": 0.9270833333333334,
314
+ "policy_expert_regret": 0.8476849501021206,
315
+ "policy_oracle_regret": 0.9539546399998168,
316
+ "policy_rollout_progress": 0.6463205499264101,
317
+ "policy_rollout_success_rate": 0.3333333333333333,
318
+ "restore_max_error": 1.955777406692505e-07
319
+ },
320
+ "PickCube-v1": {
321
+ "action_mse_to_best": 0.33665372297694585,
322
+ "expert_success_rate": 0.9444444444444444,
323
+ "num_groups": 198,
324
+ "oracle_success_rate": 0.9595959595959596,
325
+ "policy_expert_regret": 0.980730776533936,
326
+ "policy_oracle_regret": 0.9878142270626444,
327
+ "policy_rollout_progress": 0.6357579371995397,
328
+ "policy_rollout_success_rate": 0.3383838383838384,
329
+ "restore_max_error": 2.384185791015625e-07
330
+ },
331
+ "PullCube-v1": {
332
+ "action_mse_to_best": 0.6442032671637006,
333
+ "expert_success_rate": 0.24444444444444444,
334
+ "num_groups": 90,
335
+ "oracle_success_rate": 0.4666666666666667,
336
+ "policy_expert_regret": 0.2941978946190122,
337
+ "policy_oracle_regret": 0.5228573401010079,
338
+ "policy_rollout_progress": 0.3275922458963881,
339
+ "policy_rollout_success_rate": 0.2222222222222222,
340
+ "restore_max_error": 2.384185791015625e-07
341
+ },
342
+ "PushCube-v1": {
343
+ "action_mse_to_best": 0.397049247997232,
344
+ "expert_success_rate": 0.8514851485148515,
345
+ "num_groups": 101,
346
+ "oracle_success_rate": 1.0,
347
+ "policy_expert_regret": 0.3614020819711213,
348
+ "policy_oracle_regret": 0.41630433337523204,
349
+ "policy_rollout_progress": 0.8015174488029858,
350
+ "policy_rollout_success_rate": 0.7821782178217822,
351
+ "restore_max_error": 2.384185791015625e-07
352
+ },
353
+ "StackCube-v1": {
354
+ "action_mse_to_best": 0.525552598014474,
355
+ "expert_success_rate": 0.7666666666666667,
356
+ "num_groups": 90,
357
+ "oracle_success_rate": 0.9111111111111111,
358
+ "policy_expert_regret": 1.1679054531786177,
359
+ "policy_oracle_regret": 1.3300260045462184,
360
+ "policy_rollout_progress": 0.40787654883331725,
361
+ "policy_rollout_success_rate": 0.14444444444444443,
362
+ "restore_max_error": 2.2351741790771484e-07
363
+ }
364
+ }
365
+ }
366
+ ]
367
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp002_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_l2comp002_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: 35.54% +/- 1.06%
11
+ Gain vs h=16 rank checkpoint: +5.80%
12
+ Mean progress: 57.03%
13
+ Mean action MSE to best: 0.408
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% | 55.80% | 85.74% | 0.400 |
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% | 57.02% | 86.96% | 0.395 |
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% | 58.27% | 87.65% | 0.428 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,367 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_l2comp005_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": 0.3547826086956522,
9
+ "std_success": 0.00968306845709568,
10
+ "mean_progress": 0.5694483268211469,
11
+ "mean_action_mse_to_best": 0.40516445952778063,
12
+ "gain_vs_h4": 0.05808260869565218,
13
+ "gain_vs_h16_rank_checkpoint": 0.05739130434782613,
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_l2comp005_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": 0.05,
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
+ "candidate_type_bonuses": {
49
+ "residual_no_op": 0.03
50
+ },
51
+ "candidate_type_bonus_components": false,
52
+ "selected_residual_scale_counts": {
53
+ "0.35": 532,
54
+ "0.45": 38,
55
+ "0.4": 5
56
+ },
57
+ "policy_rollout_success_rate": 0.35304347826086957,
58
+ "policy_rollout_progress": 0.5575822278395619,
59
+ "oracle_success_rate": 0.8573913043478261,
60
+ "action_mse_to_best": 0.39534642129002706,
61
+ "best_policy_val": {
62
+ "bc_loss": 0.13721593966086706,
63
+ "field_effect_loss": 0.009290305380192068,
64
+ "field_potential_loss": 0.2666468388504452,
65
+ "field_preference_loss": 0.5130573478009965,
66
+ "lattice_edges": 3833.3333333333335,
67
+ "progress_mae": 0.1933159919248687,
68
+ "rank_acc": 0.8265031774838766,
69
+ "rank_loss": 0.5130523675017886,
70
+ "regret_mae": 0.3756548762321472,
71
+ "success_accuracy": 0.8773836526605818,
72
+ "total_loss": 1.5581054819954767
73
+ },
74
+ "per_task": {
75
+ "LiftPegUpright-v1": {
76
+ "action_mse_to_best": 0.3327169909631622,
77
+ "expert_success_rate": 0.8865979381443299,
78
+ "num_groups": 97,
79
+ "oracle_success_rate": 0.9278350515463918,
80
+ "policy_expert_regret": 1.0783523666797226,
81
+ "policy_oracle_regret": 1.1047678443881654,
82
+ "policy_rollout_progress": 0.5714445269292163,
83
+ "policy_rollout_success_rate": 0.2268041237113402,
84
+ "restore_max_error": 2.384185791015625e-07
85
+ },
86
+ "PickCube-v1": {
87
+ "action_mse_to_best": 0.30273104519941485,
88
+ "expert_success_rate": 0.9375,
89
+ "num_groups": 208,
90
+ "oracle_success_rate": 0.9471153846153846,
91
+ "policy_expert_regret": 0.9543058642974267,
92
+ "policy_oracle_regret": 0.9677287317239321,
93
+ "policy_rollout_progress": 0.6181530872216592,
94
+ "policy_rollout_success_rate": 0.34615384615384615,
95
+ "restore_max_error": 2.384185791015625e-07
96
+ },
97
+ "PullCube-v1": {
98
+ "action_mse_to_best": 0.6033577046224049,
99
+ "expert_success_rate": 0.19480519480519481,
100
+ "num_groups": 77,
101
+ "oracle_success_rate": 0.36363636363636365,
102
+ "policy_expert_regret": 0.24368244630296704,
103
+ "policy_oracle_regret": 0.41723848439680483,
104
+ "policy_rollout_progress": 0.31240937069614794,
105
+ "policy_rollout_success_rate": 0.2077922077922078,
106
+ "restore_max_error": 2.3096799850463867e-07
107
+ },
108
+ "PushCube-v1": {
109
+ "action_mse_to_best": 0.3675119048578942,
110
+ "expert_success_rate": 0.8617021276595744,
111
+ "num_groups": 94,
112
+ "oracle_success_rate": 0.9893617021276596,
113
+ "policy_expert_regret": 0.5508309629686335,
114
+ "policy_oracle_regret": 0.628381609441118,
115
+ "policy_rollout_progress": 0.7014056246014352,
116
+ "policy_rollout_success_rate": 0.6702127659574468,
117
+ "restore_max_error": 2.384185791015625e-07
118
+ },
119
+ "StackCube-v1": {
120
+ "action_mse_to_best": 0.515938630346397,
121
+ "expert_success_rate": 0.6767676767676768,
122
+ "num_groups": 99,
123
+ "oracle_success_rate": 0.8585858585858586,
124
+ "policy_expert_regret": 0.9108850552578165,
125
+ "policy_oracle_regret": 1.058565313918422,
126
+ "policy_rollout_progress": 0.4708703989633406,
127
+ "policy_rollout_success_rate": 0.30303030303030304,
128
+ "restore_max_error": 2.384185791015625e-07
129
+ }
130
+ }
131
+ },
132
+ {
133
+ "seed": 1,
134
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03.json",
135
+ "num_groups": 575,
136
+ "selection_mode": "retrieval_residual",
137
+ "num_candidates": 48,
138
+ "candidate_sigma": 0.0,
139
+ "selection_margin": 0.2,
140
+ "prepend_policy_candidate": false,
141
+ "field_optim_steps": 0,
142
+ "field_optim_step_size": 0.0,
143
+ "field_optim_trust_radius": 0.0,
144
+ "field_optim_l2_penalty": 0.0,
145
+ "retrieval_neighbors": 4,
146
+ "retrieval_metric": "raw",
147
+ "retrieval_type_min_success": 0.0,
148
+ "retrieval_type_success_bonus_scale": 0.0,
149
+ "retrieval_residual_consensus_penalty_scale": 0.0,
150
+ "retrieval_residual_min_source_progress": 0.0,
151
+ "retrieval_residual_min_source_advantage": -1000000000.0,
152
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
153
+ "retrieval_residual_source_score_bonus_scale": 0.0,
154
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
155
+ "retrieval_residual_composite_l2_penalty_scale": 0.05,
156
+ "retrieval_residual_scale": 1.0,
157
+ "retrieval_residual_scales": [
158
+ 0.35,
159
+ 0.4,
160
+ 0.45
161
+ ],
162
+ "retrieval_residual_anchor": "expert",
163
+ "retrieval_residual_direction": "candidate_minus_anchor",
164
+ "retrieval_residual_reduce": "compose_mean_by_type",
165
+ "candidate_type_bonuses": {
166
+ "residual_no_op": 0.03
167
+ },
168
+ "candidate_type_bonus_components": false,
169
+ "selected_residual_scale_counts": {
170
+ "0.35": 550,
171
+ "0.45": 22,
172
+ "0.4": 3
173
+ },
174
+ "policy_rollout_success_rate": 0.34608695652173915,
175
+ "policy_rollout_progress": 0.5696634550133477,
176
+ "oracle_success_rate": 0.8695652173913043,
177
+ "action_mse_to_best": 0.3933826872949367,
178
+ "best_policy_val": {
179
+ "bc_loss": 0.1315989200439718,
180
+ "field_effect_loss": 0.010110370814800262,
181
+ "field_potential_loss": 0.31546536915832096,
182
+ "field_preference_loss": 0.4960637440284093,
183
+ "lattice_edges": 3833.3333333333335,
184
+ "progress_mae": 0.19023226698239645,
185
+ "rank_acc": 0.8255469501018524,
186
+ "rank_loss": 0.4960531195004781,
187
+ "regret_mae": 0.42156532737943864,
188
+ "success_accuracy": 0.8807473679383596,
189
+ "total_loss": 1.561984618504842
190
+ },
191
+ "per_task": {
192
+ "LiftPegUpright-v1": {
193
+ "action_mse_to_best": 0.3261080000104498,
194
+ "expert_success_rate": 0.8584070796460177,
195
+ "num_groups": 113,
196
+ "oracle_success_rate": 0.9380530973451328,
197
+ "policy_expert_regret": 0.9873896409979964,
198
+ "policy_oracle_regret": 1.0665916710828258,
199
+ "policy_rollout_progress": 0.6355200547032651,
200
+ "policy_rollout_success_rate": 0.23008849557522124,
201
+ "restore_max_error": 2.384185791015625e-07
202
+ },
203
+ "PickCube-v1": {
204
+ "action_mse_to_best": 0.3202790089316018,
205
+ "expert_success_rate": 0.9402173913043478,
206
+ "num_groups": 184,
207
+ "oracle_success_rate": 0.9456521739130435,
208
+ "policy_expert_regret": 1.062856701546875,
209
+ "policy_oracle_regret": 1.062880706021805,
210
+ "policy_rollout_progress": 0.576910365069204,
211
+ "policy_rollout_success_rate": 0.29891304347826086,
212
+ "restore_max_error": 2.384185791015625e-07
213
+ },
214
+ "PullCube-v1": {
215
+ "action_mse_to_best": 0.5788939923636223,
216
+ "expert_success_rate": 0.25,
217
+ "num_groups": 76,
218
+ "oracle_success_rate": 0.40789473684210525,
219
+ "policy_expert_regret": 0.32425887294505773,
220
+ "policy_oracle_regret": 0.547437375999595,
221
+ "policy_rollout_progress": 0.2783239802443667,
222
+ "policy_rollout_success_rate": 0.17105263157894737,
223
+ "restore_max_error": 2.384185791015625e-07
224
+ },
225
+ "PushCube-v1": {
226
+ "action_mse_to_best": 0.3571035131677851,
227
+ "expert_success_rate": 0.8198198198198198,
228
+ "num_groups": 111,
229
+ "oracle_success_rate": 1.0,
230
+ "policy_expert_regret": 0.32961789887767656,
231
+ "policy_oracle_regret": 0.3467854660105061,
232
+ "policy_rollout_progress": 0.8333947141696741,
233
+ "policy_rollout_success_rate": 0.8198198198198198,
234
+ "restore_max_error": 2.3795291781425476e-07
235
+ },
236
+ "StackCube-v1": {
237
+ "action_mse_to_best": 0.5140557161399296,
238
+ "expert_success_rate": 0.6923076923076923,
239
+ "num_groups": 91,
240
+ "oracle_success_rate": 0.8571428571428571,
241
+ "policy_expert_regret": 1.136088782123157,
242
+ "policy_oracle_regret": 1.2957506292796397,
243
+ "policy_rollout_progress": 0.39485469787985417,
244
+ "policy_rollout_success_rate": 0.15384615384615385,
245
+ "restore_max_error": 1.9371509552001953e-07
246
+ }
247
+ }
248
+ },
249
+ {
250
+ "seed": 2,
251
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03.json",
252
+ "num_groups": 575,
253
+ "selection_mode": "retrieval_residual",
254
+ "num_candidates": 48,
255
+ "candidate_sigma": 0.0,
256
+ "selection_margin": 0.2,
257
+ "prepend_policy_candidate": false,
258
+ "field_optim_steps": 0,
259
+ "field_optim_step_size": 0.0,
260
+ "field_optim_trust_radius": 0.0,
261
+ "field_optim_l2_penalty": 0.0,
262
+ "retrieval_neighbors": 4,
263
+ "retrieval_metric": "raw",
264
+ "retrieval_type_min_success": 0.0,
265
+ "retrieval_type_success_bonus_scale": 0.0,
266
+ "retrieval_residual_consensus_penalty_scale": 0.0,
267
+ "retrieval_residual_min_source_progress": 0.0,
268
+ "retrieval_residual_min_source_advantage": -1000000000.0,
269
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
270
+ "retrieval_residual_source_score_bonus_scale": 0.0,
271
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
272
+ "retrieval_residual_composite_l2_penalty_scale": 0.05,
273
+ "retrieval_residual_scale": 1.0,
274
+ "retrieval_residual_scales": [
275
+ 0.35,
276
+ 0.4,
277
+ 0.45
278
+ ],
279
+ "retrieval_residual_anchor": "expert",
280
+ "retrieval_residual_direction": "candidate_minus_anchor",
281
+ "retrieval_residual_reduce": "compose_mean_by_type",
282
+ "candidate_type_bonuses": {
283
+ "residual_no_op": 0.03
284
+ },
285
+ "candidate_type_bonus_components": false,
286
+ "selected_residual_scale_counts": {
287
+ "0.35": 546,
288
+ "0.45": 26,
289
+ "0.4": 3
290
+ },
291
+ "policy_rollout_success_rate": 0.3652173913043478,
292
+ "policy_rollout_progress": 0.581099297610531,
293
+ "oracle_success_rate": 0.8765217391304347,
294
+ "action_mse_to_best": 0.4267642699983781,
295
+ "best_policy_val": {
296
+ "bc_loss": 0.11367896075050037,
297
+ "field_effect_loss": 0.009670218582161598,
298
+ "field_potential_loss": 0.2641640139950646,
299
+ "field_preference_loss": 0.5130490180518892,
300
+ "lattice_edges": 3833.3333333333335,
301
+ "progress_mae": 0.2021729110015763,
302
+ "rank_acc": 0.8333857821093665,
303
+ "rank_loss": 0.5130119257503085,
304
+ "regret_mae": 0.3958987047274907,
305
+ "success_accuracy": 0.8680730561415354,
306
+ "total_loss": 1.4394984311527677
307
+ },
308
+ "per_task": {
309
+ "LiftPegUpright-v1": {
310
+ "action_mse_to_best": 0.35442226944239036,
311
+ "expert_success_rate": 0.8229166666666666,
312
+ "num_groups": 96,
313
+ "oracle_success_rate": 0.9270833333333334,
314
+ "policy_expert_regret": 0.8476849501021206,
315
+ "policy_oracle_regret": 0.9539546399998168,
316
+ "policy_rollout_progress": 0.6463205499264101,
317
+ "policy_rollout_success_rate": 0.3333333333333333,
318
+ "restore_max_error": 1.955777406692505e-07
319
+ },
320
+ "PickCube-v1": {
321
+ "action_mse_to_best": 0.33665372297694585,
322
+ "expert_success_rate": 0.9444444444444444,
323
+ "num_groups": 198,
324
+ "oracle_success_rate": 0.9595959595959596,
325
+ "policy_expert_regret": 0.980730776533936,
326
+ "policy_oracle_regret": 0.9878142270626444,
327
+ "policy_rollout_progress": 0.6357579371995397,
328
+ "policy_rollout_success_rate": 0.3383838383838384,
329
+ "restore_max_error": 2.384185791015625e-07
330
+ },
331
+ "PullCube-v1": {
332
+ "action_mse_to_best": 0.6442032671637006,
333
+ "expert_success_rate": 0.24444444444444444,
334
+ "num_groups": 90,
335
+ "oracle_success_rate": 0.4666666666666667,
336
+ "policy_expert_regret": 0.2941978946190122,
337
+ "policy_oracle_regret": 0.5228573401010079,
338
+ "policy_rollout_progress": 0.3275922458963881,
339
+ "policy_rollout_success_rate": 0.2222222222222222,
340
+ "restore_max_error": 2.384185791015625e-07
341
+ },
342
+ "PushCube-v1": {
343
+ "action_mse_to_best": 0.3903906174963064,
344
+ "expert_success_rate": 0.8514851485148515,
345
+ "num_groups": 101,
346
+ "oracle_success_rate": 1.0,
347
+ "policy_expert_regret": 0.38061200362620967,
348
+ "policy_oracle_regret": 0.43551425503032043,
349
+ "policy_rollout_progress": 0.7922085172469073,
350
+ "policy_rollout_success_rate": 0.7722772277227723,
351
+ "restore_max_error": 2.384185791015625e-07
352
+ },
353
+ "StackCube-v1": {
354
+ "action_mse_to_best": 0.525552598014474,
355
+ "expert_success_rate": 0.7666666666666667,
356
+ "num_groups": 90,
357
+ "oracle_success_rate": 0.9111111111111111,
358
+ "policy_expert_regret": 1.1679054531786177,
359
+ "policy_oracle_regret": 1.3300260045462184,
360
+ "policy_rollout_progress": 0.40787654883331725,
361
+ "policy_rollout_success_rate": 0.14444444444444443,
362
+ "restore_max_error": 2.2351741790771484e-07
363
+ }
364
+ }
365
+ }
366
+ ]
367
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp005_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_l2comp005_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: 35.48% +/- 0.97%
11
+ Gain vs h=16 rank checkpoint: +5.74%
12
+ Mean progress: 56.94%
13
+ Mean action MSE to best: 0.405
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% | 55.76% | 85.74% | 0.395 |
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% | 56.97% | 86.96% | 0.393 |
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 | 36.52% | 58.11% | 87.65% | 0.427 |