anhtld commited on
Commit
306e16d
·
verified ·
1 Parent(s): 08afaaa

Auto-sync: 2026-06-29 09:58:28

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_compose_grid035040045_safe_margin0p20_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,361 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_compose_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": 0.34144927536231884,
9
+ "std_success": 0.01578047256674342,
10
+ "mean_progress": 0.5600020351956472,
11
+ "mean_action_mse_to_best": 0.48164921497764146,
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_residual_k4_compose_grid035040045_safe_margin0p20_noopbonus0p03.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 48,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.2,
23
+ "prepend_policy_candidate": false,
24
+ "field_optim_steps": 0,
25
+ "field_optim_step_size": 0.0,
26
+ "field_optim_trust_radius": 0.0,
27
+ "field_optim_l2_penalty": 0.0,
28
+ "retrieval_neighbors": 4,
29
+ "retrieval_metric": "raw",
30
+ "retrieval_type_min_success": 0.0,
31
+ "retrieval_type_success_bonus_scale": 0.0,
32
+ "retrieval_residual_consensus_penalty_scale": 0.0,
33
+ "retrieval_residual_min_source_progress": 0.0,
34
+ "retrieval_residual_min_source_advantage": -1000000000.0,
35
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
36
+ "retrieval_residual_source_score_bonus_scale": 0.0,
37
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
38
+ "retrieval_residual_scale": 1.0,
39
+ "retrieval_residual_scales": [
40
+ 0.35,
41
+ 0.4,
42
+ 0.45
43
+ ],
44
+ "retrieval_residual_anchor": "expert",
45
+ "retrieval_residual_direction": "candidate_minus_anchor",
46
+ "retrieval_residual_reduce": "compose_mean_by_type",
47
+ "candidate_type_bonuses": {
48
+ "residual_no_op": 0.03
49
+ },
50
+ "selected_residual_scale_counts": {
51
+ "0.45": 83,
52
+ "0.4": 16,
53
+ "0.35": 476
54
+ },
55
+ "policy_rollout_success_rate": 0.3234782608695652,
56
+ "policy_rollout_progress": 0.5323185247628262,
57
+ "oracle_success_rate": 0.8573913043478261,
58
+ "action_mse_to_best": 0.563375490748364,
59
+ "best_policy_val": {
60
+ "bc_loss": 0.13721593966086706,
61
+ "field_effect_loss": 0.009290305380192068,
62
+ "field_potential_loss": 0.2666468388504452,
63
+ "field_preference_loss": 0.5130573478009965,
64
+ "lattice_edges": 3833.3333333333335,
65
+ "progress_mae": 0.1933159919248687,
66
+ "rank_acc": 0.8265031774838766,
67
+ "rank_loss": 0.5130523675017886,
68
+ "regret_mae": 0.3756548762321472,
69
+ "success_accuracy": 0.8773836526605818,
70
+ "total_loss": 1.5581054819954767
71
+ },
72
+ "per_task": {
73
+ "LiftPegUpright-v1": {
74
+ "action_mse_to_best": 1.1220989288358958,
75
+ "expert_success_rate": 0.8865979381443299,
76
+ "num_groups": 97,
77
+ "oracle_success_rate": 0.9278350515463918,
78
+ "policy_expert_regret": 1.2778108448711867,
79
+ "policy_oracle_regret": 1.3174253004727905,
80
+ "policy_rollout_progress": 0.45157057599923045,
81
+ "policy_rollout_success_rate": 0.13402061855670103,
82
+ "restore_max_error": 2.384185791015625e-07
83
+ },
84
+ "PickCube-v1": {
85
+ "action_mse_to_best": 0.30399148212745786,
86
+ "expert_success_rate": 0.9375,
87
+ "num_groups": 208,
88
+ "oracle_success_rate": 0.9471153846153846,
89
+ "policy_expert_regret": 0.9612944226425427,
90
+ "policy_oracle_regret": 0.9747172900690482,
91
+ "policy_rollout_progress": 0.6159722211842353,
92
+ "policy_rollout_success_rate": 0.34134615384615385,
93
+ "restore_max_error": 2.384185791015625e-07
94
+ },
95
+ "PullCube-v1": {
96
+ "action_mse_to_best": 0.6033564091889889,
97
+ "expert_success_rate": 0.19480519480519481,
98
+ "num_groups": 77,
99
+ "oracle_success_rate": 0.36363636363636365,
100
+ "policy_expert_regret": 0.24368244630296704,
101
+ "policy_oracle_regret": 0.41723848439680483,
102
+ "policy_rollout_progress": 0.31240937069614794,
103
+ "policy_rollout_success_rate": 0.2077922077922078,
104
+ "restore_max_error": 2.3096799850463867e-07
105
+ },
106
+ "PushCube-v1": {
107
+ "action_mse_to_best": 0.3681779884277506,
108
+ "expert_success_rate": 0.8617021276595744,
109
+ "num_groups": 94,
110
+ "oracle_success_rate": 0.9893617021276596,
111
+ "policy_expert_regret": 0.5513087531670611,
112
+ "policy_oracle_regret": 0.6481983095724532,
113
+ "policy_rollout_progress": 0.6922272223424404,
114
+ "policy_rollout_success_rate": 0.6595744680851063,
115
+ "restore_max_error": 2.384185791015625e-07
116
+ },
117
+ "StackCube-v1": {
118
+ "action_mse_to_best": 0.7151505897066208,
119
+ "expert_success_rate": 0.6767676767676768,
120
+ "num_groups": 99,
121
+ "oracle_success_rate": 0.8585858585858586,
122
+ "policy_expert_regret": 0.9875731102444909,
123
+ "policy_oracle_regret": 1.1351558803909956,
124
+ "policy_rollout_progress": 0.4548858930968275,
125
+ "policy_rollout_success_rate": 0.24242424242424243,
126
+ "restore_max_error": 2.384185791015625e-07
127
+ }
128
+ }
129
+ },
130
+ {
131
+ "seed": 1,
132
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_compose_grid035040045_safe_margin0p20_noopbonus0p03.json",
133
+ "num_groups": 575,
134
+ "selection_mode": "retrieval_residual",
135
+ "num_candidates": 48,
136
+ "candidate_sigma": 0.0,
137
+ "selection_margin": 0.2,
138
+ "prepend_policy_candidate": false,
139
+ "field_optim_steps": 0,
140
+ "field_optim_step_size": 0.0,
141
+ "field_optim_trust_radius": 0.0,
142
+ "field_optim_l2_penalty": 0.0,
143
+ "retrieval_neighbors": 4,
144
+ "retrieval_metric": "raw",
145
+ "retrieval_type_min_success": 0.0,
146
+ "retrieval_type_success_bonus_scale": 0.0,
147
+ "retrieval_residual_consensus_penalty_scale": 0.0,
148
+ "retrieval_residual_min_source_progress": 0.0,
149
+ "retrieval_residual_min_source_advantage": -1000000000.0,
150
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
151
+ "retrieval_residual_source_score_bonus_scale": 0.0,
152
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
153
+ "retrieval_residual_scale": 1.0,
154
+ "retrieval_residual_scales": [
155
+ 0.35,
156
+ 0.4,
157
+ 0.45
158
+ ],
159
+ "retrieval_residual_anchor": "expert",
160
+ "retrieval_residual_direction": "candidate_minus_anchor",
161
+ "retrieval_residual_reduce": "compose_mean_by_type",
162
+ "candidate_type_bonuses": {
163
+ "residual_no_op": 0.03
164
+ },
165
+ "selected_residual_scale_counts": {
166
+ "0.35": 538,
167
+ "0.45": 33,
168
+ "0.4": 4
169
+ },
170
+ "policy_rollout_success_rate": 0.34782608695652173,
171
+ "policy_rollout_progress": 0.5697598444961983,
172
+ "oracle_success_rate": 0.8695652173913043,
173
+ "action_mse_to_best": 0.4286549947750957,
174
+ "best_policy_val": {
175
+ "bc_loss": 0.1315989200439718,
176
+ "field_effect_loss": 0.010110370814800262,
177
+ "field_potential_loss": 0.31546536915832096,
178
+ "field_preference_loss": 0.4960637440284093,
179
+ "lattice_edges": 3833.3333333333335,
180
+ "progress_mae": 0.19023226698239645,
181
+ "rank_acc": 0.8255469501018524,
182
+ "rank_loss": 0.4960531195004781,
183
+ "regret_mae": 0.42156532737943864,
184
+ "success_accuracy": 0.8807473679383596,
185
+ "total_loss": 1.561984618504842
186
+ },
187
+ "per_task": {
188
+ "LiftPegUpright-v1": {
189
+ "action_mse_to_best": 0.4459682410187294,
190
+ "expert_success_rate": 0.8584070796460177,
191
+ "num_groups": 113,
192
+ "oracle_success_rate": 0.9380530973451328,
193
+ "policy_expert_regret": 1.0089420887221277,
194
+ "policy_oracle_regret": 1.0881441188069572,
195
+ "policy_rollout_progress": 0.6228171645012577,
196
+ "policy_rollout_success_rate": 0.22123893805309736,
197
+ "restore_max_error": 2.384185791015625e-07
198
+ },
199
+ "PickCube-v1": {
200
+ "action_mse_to_best": 0.32267473227061005,
201
+ "expert_success_rate": 0.9402173913043478,
202
+ "num_groups": 184,
203
+ "oracle_success_rate": 0.9456521739130435,
204
+ "policy_expert_regret": 1.0550734593290025,
205
+ "policy_oracle_regret": 1.0550974638039328,
206
+ "policy_rollout_progress": 0.5792588246783809,
207
+ "policy_rollout_success_rate": 0.30434782608695654,
208
+ "restore_max_error": 2.384185791015625e-07
209
+ },
210
+ "PullCube-v1": {
211
+ "action_mse_to_best": 0.6011673336554515,
212
+ "expert_success_rate": 0.25,
213
+ "num_groups": 76,
214
+ "oracle_success_rate": 0.40789473684210525,
215
+ "policy_expert_regret": 0.27572859843310554,
216
+ "policy_oracle_regret": 0.5229248910358077,
217
+ "policy_rollout_progress": 0.28967857047131185,
218
+ "policy_rollout_success_rate": 0.18421052631578946,
219
+ "restore_max_error": 2.384185791015625e-07
220
+ },
221
+ "PushCube-v1": {
222
+ "action_mse_to_best": 0.3571399895994513,
223
+ "expert_success_rate": 0.8198198198198198,
224
+ "num_groups": 111,
225
+ "oracle_success_rate": 1.0,
226
+ "policy_expert_regret": 0.32961789887767656,
227
+ "policy_oracle_regret": 0.3467854660105061,
228
+ "policy_rollout_progress": 0.8333947141696741,
229
+ "policy_rollout_success_rate": 0.8198198198198198,
230
+ "restore_max_error": 2.3795291781425476e-07
231
+ },
232
+ "StackCube-v1": {
233
+ "action_mse_to_best": 0.5646022397738236,
234
+ "expert_success_rate": 0.6923076923076923,
235
+ "num_groups": 91,
236
+ "oracle_success_rate": 0.8571428571428571,
237
+ "policy_expert_regret": 1.133937297286568,
238
+ "policy_oracle_regret": 1.2935991444430508,
239
+ "policy_rollout_progress": 0.397006182716443,
240
+ "policy_rollout_success_rate": 0.15384615384615385,
241
+ "restore_max_error": 1.9371509552001953e-07
242
+ }
243
+ }
244
+ },
245
+ {
246
+ "seed": 2,
247
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_compose_grid035040045_safe_margin0p20_noopbonus0p03.json",
248
+ "num_groups": 575,
249
+ "selection_mode": "retrieval_residual",
250
+ "num_candidates": 48,
251
+ "candidate_sigma": 0.0,
252
+ "selection_margin": 0.2,
253
+ "prepend_policy_candidate": false,
254
+ "field_optim_steps": 0,
255
+ "field_optim_step_size": 0.0,
256
+ "field_optim_trust_radius": 0.0,
257
+ "field_optim_l2_penalty": 0.0,
258
+ "retrieval_neighbors": 4,
259
+ "retrieval_metric": "raw",
260
+ "retrieval_type_min_success": 0.0,
261
+ "retrieval_type_success_bonus_scale": 0.0,
262
+ "retrieval_residual_consensus_penalty_scale": 0.0,
263
+ "retrieval_residual_min_source_progress": 0.0,
264
+ "retrieval_residual_min_source_advantage": -1000000000.0,
265
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
266
+ "retrieval_residual_source_score_bonus_scale": 0.0,
267
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
268
+ "retrieval_residual_scale": 1.0,
269
+ "retrieval_residual_scales": [
270
+ 0.35,
271
+ 0.4,
272
+ 0.45
273
+ ],
274
+ "retrieval_residual_anchor": "expert",
275
+ "retrieval_residual_direction": "candidate_minus_anchor",
276
+ "retrieval_residual_reduce": "compose_mean_by_type",
277
+ "candidate_type_bonuses": {
278
+ "residual_no_op": 0.03
279
+ },
280
+ "selected_residual_scale_counts": {
281
+ "0.35": 537,
282
+ "0.45": 31,
283
+ "0.4": 7
284
+ },
285
+ "policy_rollout_success_rate": 0.35304347826086957,
286
+ "policy_rollout_progress": 0.577927736327917,
287
+ "oracle_success_rate": 0.8765217391304347,
288
+ "action_mse_to_best": 0.4529171594094647,
289
+ "best_policy_val": {
290
+ "bc_loss": 0.11367896075050037,
291
+ "field_effect_loss": 0.009670218582161598,
292
+ "field_potential_loss": 0.2641640139950646,
293
+ "field_preference_loss": 0.5130490180518892,
294
+ "lattice_edges": 3833.3333333333335,
295
+ "progress_mae": 0.2021729110015763,
296
+ "rank_acc": 0.8333857821093665,
297
+ "rank_loss": 0.5130119257503085,
298
+ "regret_mae": 0.3958987047274907,
299
+ "success_accuracy": 0.8680730561415354,
300
+ "total_loss": 1.4394984311527677
301
+ },
302
+ "per_task": {
303
+ "LiftPegUpright-v1": {
304
+ "action_mse_to_best": 0.3684568781318376,
305
+ "expert_success_rate": 0.8229166666666666,
306
+ "num_groups": 96,
307
+ "oracle_success_rate": 0.9270833333333334,
308
+ "policy_expert_regret": 0.8480236708807448,
309
+ "policy_oracle_regret": 0.9542933607784411,
310
+ "policy_rollout_progress": 0.6459818291477859,
311
+ "policy_rollout_success_rate": 0.3333333333333333,
312
+ "restore_max_error": 1.955777406692505e-07
313
+ },
314
+ "PickCube-v1": {
315
+ "action_mse_to_best": 0.3601090619810904,
316
+ "expert_success_rate": 0.9444444444444444,
317
+ "num_groups": 198,
318
+ "oracle_success_rate": 0.9595959595959596,
319
+ "policy_expert_regret": 1.0246550937493641,
320
+ "policy_oracle_regret": 1.038836064588542,
321
+ "policy_rollout_progress": 0.6179832959867487,
322
+ "policy_rollout_success_rate": 0.29797979797979796,
323
+ "restore_max_error": 2.384185791015625e-07
324
+ },
325
+ "PullCube-v1": {
326
+ "action_mse_to_best": 0.646510697570112,
327
+ "expert_success_rate": 0.24444444444444444,
328
+ "num_groups": 90,
329
+ "oracle_success_rate": 0.4666666666666667,
330
+ "policy_expert_regret": 0.29380798702713945,
331
+ "policy_oracle_regret": 0.5224730442864408,
332
+ "policy_rollout_progress": 0.32797654171095525,
333
+ "policy_rollout_success_rate": 0.2222222222222222,
334
+ "restore_max_error": 2.384185791015625e-07
335
+ },
336
+ "PushCube-v1": {
337
+ "action_mse_to_best": 0.3970620724381787,
338
+ "expert_success_rate": 0.8514851485148515,
339
+ "num_groups": 101,
340
+ "oracle_success_rate": 1.0,
341
+ "policy_expert_regret": 0.3614020819711213,
342
+ "policy_oracle_regret": 0.41630433337523204,
343
+ "policy_rollout_progress": 0.8015174488029858,
344
+ "policy_rollout_success_rate": 0.7821782178217822,
345
+ "restore_max_error": 2.384185791015625e-07
346
+ },
347
+ "StackCube-v1": {
348
+ "action_mse_to_best": 0.6162742221107086,
349
+ "expert_success_rate": 0.7666666666666667,
350
+ "num_groups": 90,
351
+ "oracle_success_rate": 0.9111111111111111,
352
+ "policy_expert_regret": 1.1633344186676873,
353
+ "policy_oracle_regret": 1.325454970035288,
354
+ "policy_rollout_progress": 0.41624832335445616,
355
+ "policy_rollout_success_rate": 0.14444444444444443,
356
+ "restore_max_error": 2.2351741790771484e-07
357
+ }
358
+ }
359
+ }
360
+ ]
361
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_compose_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_compose_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: 34.14% +/- 1.58%
11
+ Gain vs h=16 rank checkpoint: +4.41%
12
+ Mean progress: 56.00%
13
+ Mean action MSE to best: 0.482
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 | 32.35% | 53.23% | 85.74% | 0.563 |
18
+ | 1 | retrieval_residual | 48 | no | 4 | raw | expert | candidate_minus_anchor | compose_mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.200 | 0.00 | 0 | 0.00 | 34.78% | 56.98% | 86.96% | 0.429 |
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 | 35.30% | 57.79% | 87.65% | 0.453 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_compose_grid035040045_safe_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_residual_k4_compose_grid035040045_safe_margin0p20.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3408695652173913,
9
+ "std_success": 0.015457729421418423,
10
+ "mean_progress": 0.5596326749873518,
11
+ "mean_action_mse_to_best": 0.4816767735026129,
12
+ "gain_vs_h4": 0.04416956521739129,
13
+ "gain_vs_h16_rank_checkpoint": 0.043478260869565244,
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_compose_grid035040045_safe_margin0p20.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 48,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.2,
23
+ "prepend_policy_candidate": false,
24
+ "field_optim_steps": 0,
25
+ "field_optim_step_size": 0.0,
26
+ "field_optim_trust_radius": 0.0,
27
+ "field_optim_l2_penalty": 0.0,
28
+ "retrieval_neighbors": 4,
29
+ "retrieval_metric": "raw",
30
+ "retrieval_type_min_success": 0.0,
31
+ "retrieval_type_success_bonus_scale": 0.0,
32
+ "retrieval_residual_consensus_penalty_scale": 0.0,
33
+ "retrieval_residual_min_source_progress": 0.0,
34
+ "retrieval_residual_min_source_advantage": -1000000000.0,
35
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
36
+ "retrieval_residual_source_score_bonus_scale": 0.0,
37
+ "retrieval_residual_source_advantage_bonus_scale": 0.0,
38
+ "retrieval_residual_scale": 1.0,
39
+ "retrieval_residual_scales": [
40
+ 0.35,
41
+ 0.4,
42
+ 0.45
43
+ ],
44
+ "retrieval_residual_anchor": "expert",
45
+ "retrieval_residual_direction": "candidate_minus_anchor",
46
+ "retrieval_residual_reduce": "compose_mean_by_type",
47
+ "candidate_type_bonuses": {},
48
+ "selected_residual_scale_counts": {
49
+ "0.45": 79,
50
+ "0.4": 15,
51
+ "0.35": 481
52
+ },
53
+ "policy_rollout_success_rate": 0.3234782608695652,
54
+ "policy_rollout_progress": 0.5319681812231631,
55
+ "oracle_success_rate": 0.8573913043478261,
56
+ "action_mse_to_best": 0.5628787715733051,
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": 1.1220989288358958,
73
+ "expert_success_rate": 0.8865979381443299,
74
+ "num_groups": 97,
75
+ "oracle_success_rate": 0.9278350515463918,
76
+ "policy_expert_regret": 1.2778108448711867,
77
+ "policy_oracle_regret": 1.3174253004727905,
78
+ "policy_rollout_progress": 0.45157057599923045,
79
+ "policy_rollout_success_rate": 0.13402061855670103,
80
+ "restore_max_error": 2.384185791015625e-07
81
+ },
82
+ "PickCube-v1": {
83
+ "action_mse_to_best": 0.30374566746804005,
84
+ "expert_success_rate": 0.9375,
85
+ "num_groups": 208,
86
+ "oracle_success_rate": 0.9471153846153846,
87
+ "policy_expert_regret": 0.9612364508211613,
88
+ "policy_oracle_regret": 0.9746593182476667,
89
+ "policy_rollout_progress": 0.6160301930056169,
90
+ "policy_rollout_success_rate": 0.34134615384615385,
91
+ "restore_max_error": 2.384185791015625e-07
92
+ },
93
+ "PullCube-v1": {
94
+ "action_mse_to_best": 0.6003111614422365,
95
+ "expert_success_rate": 0.19480519480519481,
96
+ "num_groups": 77,
97
+ "oracle_success_rate": 0.36363636363636365,
98
+ "policy_expert_regret": 0.24661976460052967,
99
+ "policy_oracle_regret": 0.4201758026943675,
100
+ "policy_rollout_progress": 0.3096365697331146,
101
+ "policy_rollout_success_rate": 0.2077922077922078,
102
+ "restore_max_error": 2.3096799850463867e-07
103
+ },
104
+ "PushCube-v1": {
105
+ "action_mse_to_best": 0.3681779884277506,
106
+ "expert_success_rate": 0.8617021276595744,
107
+ "num_groups": 94,
108
+ "oracle_success_rate": 0.9893617021276596,
109
+ "policy_expert_regret": 0.5513087531670611,
110
+ "policy_oracle_regret": 0.6481983095724532,
111
+ "policy_rollout_progress": 0.6922272223424404,
112
+ "policy_rollout_success_rate": 0.6595744680851063,
113
+ "restore_max_error": 2.384185791015625e-07
114
+ },
115
+ "StackCube-v1": {
116
+ "action_mse_to_best": 0.7151505897066208,
117
+ "expert_success_rate": 0.6767676767676768,
118
+ "num_groups": 99,
119
+ "oracle_success_rate": 0.8585858585858586,
120
+ "policy_expert_regret": 0.9875731102444909,
121
+ "policy_oracle_regret": 1.1351558803909956,
122
+ "policy_rollout_progress": 0.4548858930968275,
123
+ "policy_rollout_success_rate": 0.24242424242424243,
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_residual_k4_compose_grid035040045_safe_margin0p20.json",
131
+ "num_groups": 575,
132
+ "selection_mode": "retrieval_residual",
133
+ "num_candidates": 48,
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
+ 0.35,
154
+ 0.4,
155
+ 0.45
156
+ ],
157
+ "retrieval_residual_anchor": "expert",
158
+ "retrieval_residual_direction": "candidate_minus_anchor",
159
+ "retrieval_residual_reduce": "compose_mean_by_type",
160
+ "candidate_type_bonuses": {},
161
+ "selected_residual_scale_counts": {
162
+ "0.35": 540,
163
+ "0.45": 31,
164
+ "0.4": 4
165
+ },
166
+ "policy_rollout_success_rate": 0.34608695652173915,
167
+ "policy_rollout_progress": 0.5689693509657746,
168
+ "oracle_success_rate": 0.8695652173913043,
169
+ "action_mse_to_best": 0.4294346837553641,
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": 0.4459682410187294,
186
+ "expert_success_rate": 0.8584070796460177,
187
+ "num_groups": 113,
188
+ "oracle_success_rate": 0.9380530973451328,
189
+ "policy_expert_regret": 1.0089420887221277,
190
+ "policy_oracle_regret": 1.0881441188069572,
191
+ "policy_rollout_progress": 0.6228171645012577,
192
+ "policy_rollout_success_rate": 0.22123893805309736,
193
+ "restore_max_error": 2.384185791015625e-07
194
+ },
195
+ "PickCube-v1": {
196
+ "action_mse_to_best": 0.3251112603339488,
197
+ "expert_success_rate": 0.9402173913043478,
198
+ "num_groups": 184,
199
+ "oracle_success_rate": 0.9456521739130435,
200
+ "policy_expert_regret": 1.0629785342202724,
201
+ "policy_oracle_regret": 1.0630025386952027,
202
+ "policy_rollout_progress": 0.5767885323958066,
203
+ "policy_rollout_success_rate": 0.29891304347826086,
204
+ "restore_max_error": 2.384185791015625e-07
205
+ },
206
+ "PullCube-v1": {
207
+ "action_mse_to_best": 0.6011673336554515,
208
+ "expert_success_rate": 0.25,
209
+ "num_groups": 76,
210
+ "oracle_success_rate": 0.40789473684210525,
211
+ "policy_expert_regret": 0.27572859843310554,
212
+ "policy_oracle_regret": 0.5229248910358077,
213
+ "policy_rollout_progress": 0.28967857047131185,
214
+ "policy_rollout_success_rate": 0.18421052631578946,
215
+ "restore_max_error": 2.384185791015625e-07
216
+ },
217
+ "PushCube-v1": {
218
+ "action_mse_to_best": 0.3571399895994513,
219
+ "expert_success_rate": 0.8198198198198198,
220
+ "num_groups": 111,
221
+ "oracle_success_rate": 1.0,
222
+ "policy_expert_regret": 0.32961789887767656,
223
+ "policy_oracle_regret": 0.3467854660105061,
224
+ "policy_rollout_progress": 0.8333947141696741,
225
+ "policy_rollout_success_rate": 0.8198198198198198,
226
+ "restore_max_error": 2.3795291781425476e-07
227
+ },
228
+ "StackCube-v1": {
229
+ "action_mse_to_best": 0.5646022397738236,
230
+ "expert_success_rate": 0.6923076923076923,
231
+ "num_groups": 91,
232
+ "oracle_success_rate": 0.8571428571428571,
233
+ "policy_expert_regret": 1.133937297286568,
234
+ "policy_oracle_regret": 1.2935991444430508,
235
+ "policy_rollout_progress": 0.397006182716443,
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_residual_k4_compose_grid035040045_safe_margin0p20.json",
244
+ "num_groups": 575,
245
+ "selection_mode": "retrieval_residual",
246
+ "num_candidates": 48,
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
+ 0.35,
267
+ 0.4,
268
+ 0.45
269
+ ],
270
+ "retrieval_residual_anchor": "expert",
271
+ "retrieval_residual_direction": "candidate_minus_anchor",
272
+ "retrieval_residual_reduce": "compose_mean_by_type",
273
+ "candidate_type_bonuses": {},
274
+ "selected_residual_scale_counts": {
275
+ "0.35": 541,
276
+ "0.45": 28,
277
+ "0.4": 6
278
+ },
279
+ "policy_rollout_success_rate": 0.35304347826086957,
280
+ "policy_rollout_progress": 0.5779604927731176,
281
+ "oracle_success_rate": 0.8765217391304347,
282
+ "action_mse_to_best": 0.4527168651791694,
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": 0.3684568781318376,
299
+ "expert_success_rate": 0.8229166666666666,
300
+ "num_groups": 96,
301
+ "oracle_success_rate": 0.9270833333333334,
302
+ "policy_expert_regret": 0.8480236708807448,
303
+ "policy_oracle_regret": 0.9542933607784411,
304
+ "policy_rollout_progress": 0.6459818291477859,
305
+ "policy_rollout_success_rate": 0.3333333333333333,
306
+ "restore_max_error": 1.955777406692505e-07
307
+ },
308
+ "PickCube-v1": {
309
+ "action_mse_to_best": 0.35918041311129173,
310
+ "expert_success_rate": 0.9444444444444444,
311
+ "num_groups": 198,
312
+ "oracle_success_rate": 0.9595959595959596,
313
+ "policy_expert_regret": 1.0246035738500079,
314
+ "policy_oracle_regret": 1.0387845446891857,
315
+ "policy_rollout_progress": 0.618034815886105,
316
+ "policy_rollout_success_rate": 0.29797979797979796,
317
+ "restore_max_error": 2.384185791015625e-07
318
+ },
319
+ "PullCube-v1": {
320
+ "action_mse_to_best": 0.6472740675012271,
321
+ "expert_success_rate": 0.24444444444444444,
322
+ "num_groups": 90,
323
+ "oracle_success_rate": 0.4666666666666667,
324
+ "policy_expert_regret": 0.29380798702713945,
325
+ "policy_oracle_regret": 0.5223771107762433,
326
+ "policy_rollout_progress": 0.3280724752211528,
327
+ "policy_rollout_success_rate": 0.2222222222222222,
328
+ "restore_max_error": 2.384185791015625e-07
329
+ },
330
+ "PushCube-v1": {
331
+ "action_mse_to_best": 0.3970620724381787,
332
+ "expert_success_rate": 0.8514851485148515,
333
+ "num_groups": 101,
334
+ "oracle_success_rate": 1.0,
335
+ "policy_expert_regret": 0.3614020819711213,
336
+ "policy_oracle_regret": 0.41630433337523204,
337
+ "policy_rollout_progress": 0.8015174488029858,
338
+ "policy_rollout_success_rate": 0.7821782178217822,
339
+ "restore_max_error": 2.384185791015625e-07
340
+ },
341
+ "StackCube-v1": {
342
+ "action_mse_to_best": 0.6162742221107086,
343
+ "expert_success_rate": 0.7666666666666667,
344
+ "num_groups": 90,
345
+ "oracle_success_rate": 0.9111111111111111,
346
+ "policy_expert_regret": 1.1633344186676873,
347
+ "policy_oracle_regret": 1.325454970035288,
348
+ "policy_rollout_progress": 0.41624832335445616,
349
+ "policy_rollout_success_rate": 0.14444444444444443,
350
+ "restore_max_error": 2.2351741790771484e-07
351
+ }
352
+ }
353
+ }
354
+ ]
355
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_compose_grid035040045_safe_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_residual_k4_compose_grid035040045_safe_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: 34.09% +/- 1.55%
11
+ Gain vs h=16 rank checkpoint: +4.35%
12
+ Mean progress: 55.96%
13
+ Mean action MSE to best: 0.482
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 | 32.35% | 53.20% | 85.74% | 0.563 |
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% | 56.90% | 86.96% | 0.429 |
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 | 35.30% | 57.80% | 87.65% | 0.453 |