anhtld commited on
Commit
f490075
·
verified ·
1 Parent(s): 8809a43

Auto-sync: 2026-06-29 03:12:48 (part 2)

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_grid035045055_safe_margin0p20_mean_by_type_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,346 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_grid035045055_safe_margin0p20_mean_by_type_noopbonus0p03.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.35130434782608694,
9
+ "std_success": 0.009202613255876844,
10
+ "mean_progress": 0.5679968697053118,
11
+ "mean_action_mse_to_best": 0.40172422174391326,
12
+ "gain_vs_h4": 0.05460434782608692,
13
+ "gain_vs_h16_rank_checkpoint": 0.05391304347826087,
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_grid035045055_safe_margin0p20_mean_by_type_noopbonus0p03.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_residual_min_source_progress": 0.0,
32
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
33
+ "retrieval_residual_source_score_bonus_scale": 0.0,
34
+ "retrieval_residual_scale": 0.4,
35
+ "retrieval_residual_scales": [
36
+ 0.35,
37
+ 0.45,
38
+ 0.55
39
+ ],
40
+ "retrieval_residual_anchor": "expert",
41
+ "retrieval_residual_reduce": "mean_by_type",
42
+ "candidate_type_bonuses": {
43
+ "residual_no_op": 0.03
44
+ },
45
+ "selected_residual_scale_counts": {
46
+ "0.35": 537,
47
+ "0.55": 31,
48
+ "0.45": 7
49
+ },
50
+ "policy_rollout_success_rate": 0.3443478260869565,
51
+ "policy_rollout_progress": 0.5549358964865299,
52
+ "oracle_success_rate": 0.8573913043478261,
53
+ "action_mse_to_best": 0.3897169904621399,
54
+ "best_policy_val": {
55
+ "bc_loss": 0.13721593966086706,
56
+ "field_effect_loss": 0.009290305380192068,
57
+ "field_potential_loss": 0.2666468388504452,
58
+ "field_preference_loss": 0.5130573478009965,
59
+ "lattice_edges": 3833.3333333333335,
60
+ "progress_mae": 0.1933159919248687,
61
+ "rank_acc": 0.8265031774838766,
62
+ "rank_loss": 0.5130523675017886,
63
+ "regret_mae": 0.3756548762321472,
64
+ "success_accuracy": 0.8773836526605818,
65
+ "total_loss": 1.5581054819954767
66
+ },
67
+ "per_task": {
68
+ "LiftPegUpright-v1": {
69
+ "action_mse_to_best": 0.31969772460726426,
70
+ "expert_success_rate": 0.8865979381443299,
71
+ "num_groups": 97,
72
+ "oracle_success_rate": 0.9278350515463918,
73
+ "policy_expert_regret": 1.0783334703175063,
74
+ "policy_oracle_regret": 1.1047489480259491,
75
+ "policy_rollout_progress": 0.5714634232914325,
76
+ "policy_rollout_success_rate": 0.2268041237113402,
77
+ "restore_max_error": 2.384185791015625e-07
78
+ },
79
+ "PickCube-v1": {
80
+ "action_mse_to_best": 0.3030652105450057,
81
+ "expert_success_rate": 0.9375,
82
+ "num_groups": 208,
83
+ "oracle_success_rate": 0.9471153846153846,
84
+ "policy_expert_regret": 0.9755131378769875,
85
+ "policy_oracle_regret": 0.9889360053034929,
86
+ "policy_rollout_progress": 0.6113688905651753,
87
+ "policy_rollout_success_rate": 0.3317307692307692,
88
+ "restore_max_error": 2.384185791015625e-07
89
+ },
90
+ "PullCube-v1": {
91
+ "action_mse_to_best": 0.6038731342399275,
92
+ "expert_success_rate": 0.19480519480519481,
93
+ "num_groups": 77,
94
+ "oracle_success_rate": 0.36363636363636365,
95
+ "policy_expert_regret": 0.24397330443894902,
96
+ "policy_oracle_regret": 0.41752934253278684,
97
+ "policy_rollout_progress": 0.31211851256016593,
98
+ "policy_rollout_success_rate": 0.2077922077922078,
99
+ "restore_max_error": 2.3096799850463867e-07
100
+ },
101
+ "PushCube-v1": {
102
+ "action_mse_to_best": 0.3664056982131714,
103
+ "expert_success_rate": 0.8617021276595744,
104
+ "num_groups": 94,
105
+ "oracle_success_rate": 0.9893617021276596,
106
+ "policy_expert_regret": 0.5508279580068081,
107
+ "policy_oracle_regret": 0.6283786044792926,
108
+ "policy_rollout_progress": 0.7014086295632606,
109
+ "policy_rollout_success_rate": 0.6702127659574468,
110
+ "restore_max_error": 2.384185791015625e-07
111
+ },
112
+ "StackCube-v1": {
113
+ "action_mse_to_best": 0.49594605522173824,
114
+ "expert_success_rate": 0.6767676767676768,
115
+ "num_groups": 99,
116
+ "oracle_success_rate": 0.8585858585858586,
117
+ "policy_expert_regret": 0.917089061454089,
118
+ "policy_oracle_regret": 1.0796789195802476,
119
+ "policy_rollout_progress": 0.46995881350353513,
120
+ "policy_rollout_success_rate": 0.2828282828282828,
121
+ "restore_max_error": 2.384185791015625e-07
122
+ }
123
+ }
124
+ },
125
+ {
126
+ "seed": 1,
127
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_grid035045055_safe_margin0p20_mean_by_type_noopbonus0p03.json",
128
+ "num_groups": 575,
129
+ "selection_mode": "retrieval_residual",
130
+ "num_candidates": 18,
131
+ "candidate_sigma": 0.0,
132
+ "selection_margin": 0.2,
133
+ "prepend_policy_candidate": false,
134
+ "field_optim_steps": 0,
135
+ "field_optim_step_size": 0.0,
136
+ "field_optim_trust_radius": 0.0,
137
+ "field_optim_l2_penalty": 0.0,
138
+ "retrieval_neighbors": 4,
139
+ "retrieval_metric": "raw",
140
+ "retrieval_type_min_success": 0.0,
141
+ "retrieval_residual_min_source_progress": 0.0,
142
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
143
+ "retrieval_residual_source_score_bonus_scale": 0.0,
144
+ "retrieval_residual_scale": 0.4,
145
+ "retrieval_residual_scales": [
146
+ 0.35,
147
+ 0.45,
148
+ 0.55
149
+ ],
150
+ "retrieval_residual_anchor": "expert",
151
+ "retrieval_residual_reduce": "mean_by_type",
152
+ "candidate_type_bonuses": {
153
+ "residual_no_op": 0.03
154
+ },
155
+ "selected_residual_scale_counts": {
156
+ "0.35": 548,
157
+ "0.45": 5,
158
+ "0.55": 22
159
+ },
160
+ "policy_rollout_success_rate": 0.34782608695652173,
161
+ "policy_rollout_progress": 0.5698409637744012,
162
+ "oracle_success_rate": 0.8695652173913043,
163
+ "action_mse_to_best": 0.3928398724855936,
164
+ "best_policy_val": {
165
+ "bc_loss": 0.1315989200439718,
166
+ "field_effect_loss": 0.010110370814800262,
167
+ "field_potential_loss": 0.31546536915832096,
168
+ "field_preference_loss": 0.4960637440284093,
169
+ "lattice_edges": 3833.3333333333335,
170
+ "progress_mae": 0.19023226698239645,
171
+ "rank_acc": 0.8255469501018524,
172
+ "rank_loss": 0.4960531195004781,
173
+ "regret_mae": 0.42156532737943864,
174
+ "success_accuracy": 0.8807473679383596,
175
+ "total_loss": 1.561984618504842
176
+ },
177
+ "per_task": {
178
+ "LiftPegUpright-v1": {
179
+ "action_mse_to_best": 0.3261080000104498,
180
+ "expert_success_rate": 0.8584070796460177,
181
+ "num_groups": 113,
182
+ "oracle_success_rate": 0.9380530973451328,
183
+ "policy_expert_regret": 0.9873896409979964,
184
+ "policy_oracle_regret": 1.0665916710828258,
185
+ "policy_rollout_progress": 0.6355200547032651,
186
+ "policy_rollout_success_rate": 0.23008849557522124,
187
+ "restore_max_error": 2.384185791015625e-07
188
+ },
189
+ "PickCube-v1": {
190
+ "action_mse_to_best": 0.3233320013538975,
191
+ "expert_success_rate": 0.9402173913043478,
192
+ "num_groups": 184,
193
+ "oracle_success_rate": 0.9456521739130435,
194
+ "policy_expert_regret": 1.055402592452162,
195
+ "policy_oracle_regret": 1.0554265969270922,
196
+ "policy_rollout_progress": 0.5789296915552214,
197
+ "policy_rollout_success_rate": 0.30434782608695654,
198
+ "restore_max_error": 2.384185791015625e-07
199
+ },
200
+ "PullCube-v1": {
201
+ "action_mse_to_best": 0.5753377006438217,
202
+ "expert_success_rate": 0.25,
203
+ "num_groups": 76,
204
+ "oracle_success_rate": 0.40789473684210525,
205
+ "policy_expert_regret": 0.32425887294505773,
206
+ "policy_oracle_regret": 0.5707975676969478,
207
+ "policy_rollout_progress": 0.26812168328385605,
208
+ "policy_rollout_success_rate": 0.15789473684210525,
209
+ "restore_max_error": 2.384185791015625e-07
210
+ },
211
+ "PushCube-v1": {
212
+ "action_mse_to_best": 0.3614138618514344,
213
+ "expert_success_rate": 0.8198198198198198,
214
+ "num_groups": 111,
215
+ "oracle_success_rate": 1.0,
216
+ "policy_expert_regret": 0.31219839673858507,
217
+ "policy_oracle_regret": 0.3293659638714146,
218
+ "policy_rollout_progress": 0.8418052072997566,
219
+ "policy_rollout_success_rate": 0.8288288288288288,
220
+ "restore_max_error": 2.3795291781425476e-07
221
+ },
222
+ "StackCube-v1": {
223
+ "action_mse_to_best": 0.5021651704887767,
224
+ "expert_success_rate": 0.6923076923076923,
225
+ "num_groups": 91,
226
+ "oracle_success_rate": 0.8571428571428571,
227
+ "policy_expert_regret": 1.1407885481040556,
228
+ "policy_oracle_regret": 1.3004503952605384,
229
+ "policy_rollout_progress": 0.39015493189895545,
230
+ "policy_rollout_success_rate": 0.15384615384615385,
231
+ "restore_max_error": 1.9371509552001953e-07
232
+ }
233
+ }
234
+ },
235
+ {
236
+ "seed": 2,
237
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_grid035045055_safe_margin0p20_mean_by_type_noopbonus0p03.json",
238
+ "num_groups": 575,
239
+ "selection_mode": "retrieval_residual",
240
+ "num_candidates": 18,
241
+ "candidate_sigma": 0.0,
242
+ "selection_margin": 0.2,
243
+ "prepend_policy_candidate": false,
244
+ "field_optim_steps": 0,
245
+ "field_optim_step_size": 0.0,
246
+ "field_optim_trust_radius": 0.0,
247
+ "field_optim_l2_penalty": 0.0,
248
+ "retrieval_neighbors": 4,
249
+ "retrieval_metric": "raw",
250
+ "retrieval_type_min_success": 0.0,
251
+ "retrieval_residual_min_source_progress": 0.0,
252
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
253
+ "retrieval_residual_source_score_bonus_scale": 0.0,
254
+ "retrieval_residual_scale": 0.4,
255
+ "retrieval_residual_scales": [
256
+ 0.35,
257
+ 0.45,
258
+ 0.55
259
+ ],
260
+ "retrieval_residual_anchor": "expert",
261
+ "retrieval_residual_reduce": "mean_by_type",
262
+ "candidate_type_bonuses": {
263
+ "residual_no_op": 0.03
264
+ },
265
+ "selected_residual_scale_counts": {
266
+ "0.35": 542,
267
+ "0.45": 9,
268
+ "0.55": 24
269
+ },
270
+ "policy_rollout_success_rate": 0.3617391304347826,
271
+ "policy_rollout_progress": 0.5792137488550044,
272
+ "oracle_success_rate": 0.8765217391304347,
273
+ "action_mse_to_best": 0.4226158022840062,
274
+ "best_policy_val": {
275
+ "bc_loss": 0.11367896075050037,
276
+ "field_effect_loss": 0.009670218582161598,
277
+ "field_potential_loss": 0.2641640139950646,
278
+ "field_preference_loss": 0.5130490180518892,
279
+ "lattice_edges": 3833.3333333333335,
280
+ "progress_mae": 0.2021729110015763,
281
+ "rank_acc": 0.8333857821093665,
282
+ "rank_loss": 0.5130119257503085,
283
+ "regret_mae": 0.3958987047274907,
284
+ "success_accuracy": 0.8680730561415354,
285
+ "total_loss": 1.4394984311527677
286
+ },
287
+ "per_task": {
288
+ "LiftPegUpright-v1": {
289
+ "action_mse_to_best": 0.35442226944239036,
290
+ "expert_success_rate": 0.8229166666666666,
291
+ "num_groups": 96,
292
+ "oracle_success_rate": 0.9270833333333334,
293
+ "policy_expert_regret": 0.8476849501021206,
294
+ "policy_oracle_regret": 0.9539546399998168,
295
+ "policy_rollout_progress": 0.6463205499264101,
296
+ "policy_rollout_success_rate": 0.3333333333333333,
297
+ "restore_max_error": 1.955777406692505e-07
298
+ },
299
+ "PickCube-v1": {
300
+ "action_mse_to_best": 0.33851213141749004,
301
+ "expert_success_rate": 0.9444444444444444,
302
+ "num_groups": 198,
303
+ "oracle_success_rate": 0.9595959595959596,
304
+ "policy_expert_regret": 0.9878664435254354,
305
+ "policy_oracle_regret": 0.9949498940541437,
306
+ "policy_rollout_progress": 0.6336727752585454,
307
+ "policy_rollout_success_rate": 0.3333333333333333,
308
+ "restore_max_error": 2.384185791015625e-07
309
+ },
310
+ "PullCube-v1": {
311
+ "action_mse_to_best": 0.6319371541341146,
312
+ "expert_success_rate": 0.24444444444444444,
313
+ "num_groups": 90,
314
+ "oracle_success_rate": 0.4666666666666667,
315
+ "policy_expert_regret": 0.3142229907904443,
316
+ "policy_oracle_regret": 0.5429822058382342,
317
+ "policy_rollout_progress": 0.31857849127027293,
318
+ "policy_rollout_success_rate": 0.2111111111111111,
319
+ "restore_max_error": 2.384185791015625e-07
320
+ },
321
+ "PushCube-v1": {
322
+ "action_mse_to_best": 0.39486069888761727,
323
+ "expert_success_rate": 0.8514851485148515,
324
+ "num_groups": 101,
325
+ "oracle_success_rate": 1.0,
326
+ "policy_expert_regret": 0.3989962453004157,
327
+ "policy_oracle_regret": 0.4538984967045265,
328
+ "policy_rollout_progress": 0.7837252656717112,
329
+ "policy_rollout_success_rate": 0.7623762376237624,
330
+ "restore_max_error": 2.384185791015625e-07
331
+ },
332
+ "StackCube-v1": {
333
+ "action_mse_to_best": 0.5022096885161267,
334
+ "expert_success_rate": 0.7666666666666667,
335
+ "num_groups": 90,
336
+ "oracle_success_rate": 0.9111111111111111,
337
+ "policy_expert_regret": 1.162422686484125,
338
+ "policy_oracle_regret": 1.311674097014798,
339
+ "policy_rollout_progress": 0.4189511916703648,
340
+ "policy_rollout_success_rate": 0.15555555555555556,
341
+ "restore_max_error": 2.2351741790771484e-07
342
+ }
343
+ }
344
+ }
345
+ ]
346
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_grid035045055_safe_margin0p20_mean_by_type_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_grid035045055_safe_margin0p20_mean_by_type_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: 35.13% +/- 0.92%
11
+ Gain vs h=16 rank checkpoint: +5.39%
12
+ Mean progress: 56.80%
13
+ Mean action MSE to best: 0.402
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | min source progress | source progress bonus | source score bonus | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---:|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 18 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.000 | 0.40 | 0.35,0.45,0.55 | 0.200 | 0.00 | 0 | 0.00 | 34.43% | 55.49% | 85.74% | 0.390 |
18
+ | 1 | retrieval_residual | 18 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.000 | 0.40 | 0.35,0.45,0.55 | 0.200 | 0.00 | 0 | 0.00 | 34.78% | 56.98% | 86.96% | 0.393 |
19
+ | 2 | retrieval_residual | 18 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.000 | 0.40 | 0.35,0.45,0.55 | 0.200 | 0.00 | 0 | 0.00 | 36.17% | 57.92% | 87.65% | 0.423 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_grid040045050_safe_margin0p20_mean_by_type_srcscorebonus0p02_summary.json ADDED
@@ -0,0 +1,340 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_grid040045050_safe_margin0p20_mean_by_type_srcscorebonus0p02.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.35304347826086957,
9
+ "std_success": 0.0105787174439969,
10
+ "mean_progress": 0.5689121059149256,
11
+ "mean_action_mse_to_best": 0.398193252855183,
12
+ "gain_vs_h4": 0.05634347826086955,
13
+ "gain_vs_h16_rank_checkpoint": 0.0556521739130435,
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_grid040045050_safe_margin0p20_mean_by_type_srcscorebonus0p02.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_residual_min_source_progress": 0.0,
32
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
33
+ "retrieval_residual_source_score_bonus_scale": 0.02,
34
+ "retrieval_residual_scale": 0.4,
35
+ "retrieval_residual_scales": [
36
+ 0.4,
37
+ 0.45,
38
+ 0.5
39
+ ],
40
+ "retrieval_residual_anchor": "expert",
41
+ "retrieval_residual_reduce": "mean_by_type",
42
+ "candidate_type_bonuses": {},
43
+ "selected_residual_scale_counts": {
44
+ "0.4": 544,
45
+ "0.5": 30,
46
+ "0.45": 1
47
+ },
48
+ "policy_rollout_success_rate": 0.34608695652173915,
49
+ "policy_rollout_progress": 0.5549538400178047,
50
+ "oracle_success_rate": 0.8573913043478261,
51
+ "action_mse_to_best": 0.3833637685717448,
52
+ "best_policy_val": {
53
+ "bc_loss": 0.13721593966086706,
54
+ "field_effect_loss": 0.009290305380192068,
55
+ "field_potential_loss": 0.2666468388504452,
56
+ "field_preference_loss": 0.5130573478009965,
57
+ "lattice_edges": 3833.3333333333335,
58
+ "progress_mae": 0.1933159919248687,
59
+ "rank_acc": 0.8265031774838766,
60
+ "rank_loss": 0.5130523675017886,
61
+ "regret_mae": 0.3756548762321472,
62
+ "success_accuracy": 0.8773836526605818,
63
+ "total_loss": 1.5581054819954767
64
+ },
65
+ "per_task": {
66
+ "LiftPegUpright-v1": {
67
+ "action_mse_to_best": 0.31969772460726426,
68
+ "expert_success_rate": 0.8865979381443299,
69
+ "num_groups": 97,
70
+ "oracle_success_rate": 0.9278350515463918,
71
+ "policy_expert_regret": 1.0783334703175063,
72
+ "policy_oracle_regret": 1.1047489480259491,
73
+ "policy_rollout_progress": 0.5714634232914325,
74
+ "policy_rollout_success_rate": 0.2268041237113402,
75
+ "restore_max_error": 2.384185791015625e-07
76
+ },
77
+ "PickCube-v1": {
78
+ "action_mse_to_best": 0.3028640510299458,
79
+ "expert_success_rate": 0.9375,
80
+ "num_groups": 208,
81
+ "oracle_success_rate": 0.9471153846153846,
82
+ "policy_expert_regret": 0.9681451314916978,
83
+ "policy_oracle_regret": 0.9815679989182032,
84
+ "policy_rollout_progress": 0.6139292046427727,
85
+ "policy_rollout_success_rate": 0.33653846153846156,
86
+ "restore_max_error": 2.384185791015625e-07
87
+ },
88
+ "PullCube-v1": {
89
+ "action_mse_to_best": 0.6005741688725236,
90
+ "expert_success_rate": 0.19480519480519481,
91
+ "num_groups": 77,
92
+ "oracle_success_rate": 0.36363636363636365,
93
+ "policy_expert_regret": 0.24742292233609728,
94
+ "policy_oracle_regret": 0.4209789604299351,
95
+ "policy_rollout_progress": 0.308833411997547,
96
+ "policy_rollout_success_rate": 0.2077922077922078,
97
+ "restore_max_error": 2.3096799850463867e-07
98
+ },
99
+ "PushCube-v1": {
100
+ "action_mse_to_best": 0.36579142193844977,
101
+ "expert_success_rate": 0.8617021276595744,
102
+ "num_groups": 94,
103
+ "oracle_success_rate": 0.9893617021276596,
104
+ "policy_expert_regret": 0.5508279580068081,
105
+ "policy_oracle_regret": 0.6283786044792926,
106
+ "policy_rollout_progress": 0.7014086295632606,
107
+ "policy_rollout_success_rate": 0.6702127659574468,
108
+ "restore_max_error": 2.384185791015625e-07
109
+ },
110
+ "StackCube-v1": {
111
+ "action_mse_to_best": 0.4626177814365788,
112
+ "expert_success_rate": 0.6767676767676768,
113
+ "num_groups": 99,
114
+ "oracle_success_rate": 0.8585858585858586,
115
+ "policy_expert_regret": 0.9198090114978829,
116
+ "policy_oracle_regret": 1.0823988696240416,
117
+ "policy_rollout_progress": 0.46723886345974125,
118
+ "policy_rollout_success_rate": 0.2828282828282828,
119
+ "restore_max_error": 2.384185791015625e-07
120
+ }
121
+ }
122
+ },
123
+ {
124
+ "seed": 1,
125
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_grid040045050_safe_margin0p20_mean_by_type_srcscorebonus0p02.json",
126
+ "num_groups": 575,
127
+ "selection_mode": "retrieval_residual",
128
+ "num_candidates": 18,
129
+ "candidate_sigma": 0.0,
130
+ "selection_margin": 0.2,
131
+ "prepend_policy_candidate": false,
132
+ "field_optim_steps": 0,
133
+ "field_optim_step_size": 0.0,
134
+ "field_optim_trust_radius": 0.0,
135
+ "field_optim_l2_penalty": 0.0,
136
+ "retrieval_neighbors": 4,
137
+ "retrieval_metric": "raw",
138
+ "retrieval_type_min_success": 0.0,
139
+ "retrieval_residual_min_source_progress": 0.0,
140
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
141
+ "retrieval_residual_source_score_bonus_scale": 0.02,
142
+ "retrieval_residual_scale": 0.4,
143
+ "retrieval_residual_scales": [
144
+ 0.4,
145
+ 0.45,
146
+ 0.5
147
+ ],
148
+ "retrieval_residual_anchor": "expert",
149
+ "retrieval_residual_reduce": "mean_by_type",
150
+ "candidate_type_bonuses": {},
151
+ "selected_residual_scale_counts": {
152
+ "0.4": 552,
153
+ "0.5": 20,
154
+ "0.45": 3
155
+ },
156
+ "policy_rollout_success_rate": 0.34782608695652173,
157
+ "policy_rollout_progress": 0.5707065432065207,
158
+ "oracle_success_rate": 0.8695652173913043,
159
+ "action_mse_to_best": 0.39096053219521826,
160
+ "best_policy_val": {
161
+ "bc_loss": 0.1315989200439718,
162
+ "field_effect_loss": 0.010110370814800262,
163
+ "field_potential_loss": 0.31546536915832096,
164
+ "field_preference_loss": 0.4960637440284093,
165
+ "lattice_edges": 3833.3333333333335,
166
+ "progress_mae": 0.19023226698239645,
167
+ "rank_acc": 0.8255469501018524,
168
+ "rank_loss": 0.4960531195004781,
169
+ "regret_mae": 0.42156532737943864,
170
+ "success_accuracy": 0.8807473679383596,
171
+ "total_loss": 1.561984618504842
172
+ },
173
+ "per_task": {
174
+ "LiftPegUpright-v1": {
175
+ "action_mse_to_best": 0.326095403600649,
176
+ "expert_success_rate": 0.8584070796460177,
177
+ "num_groups": 113,
178
+ "oracle_success_rate": 0.9380530973451328,
179
+ "policy_expert_regret": 0.99882721795445,
180
+ "policy_oracle_regret": 1.0780292480392795,
181
+ "policy_rollout_progress": 0.6329320352689355,
182
+ "policy_rollout_success_rate": 0.22123893805309736,
183
+ "restore_max_error": 2.384185791015625e-07
184
+ },
185
+ "PickCube-v1": {
186
+ "action_mse_to_best": 0.319761208643246,
187
+ "expert_success_rate": 0.9402173913043478,
188
+ "num_groups": 184,
189
+ "oracle_success_rate": 0.9456521739130435,
190
+ "policy_expert_regret": 1.0551657436476292,
191
+ "policy_oracle_regret": 1.0551897481225594,
192
+ "policy_rollout_progress": 0.5791665403597542,
193
+ "policy_rollout_success_rate": 0.30434782608695654,
194
+ "restore_max_error": 2.384185791015625e-07
195
+ },
196
+ "PullCube-v1": {
197
+ "action_mse_to_best": 0.573825633937591,
198
+ "expert_success_rate": 0.25,
199
+ "num_groups": 76,
200
+ "oracle_success_rate": 0.40789473684210525,
201
+ "policy_expert_regret": 0.3001050222174902,
202
+ "policy_oracle_regret": 0.522557810194006,
203
+ "policy_rollout_progress": 0.29004565131311355,
204
+ "policy_rollout_success_rate": 0.18421052631578946,
205
+ "restore_max_error": 2.384185791015625e-07
206
+ },
207
+ "PushCube-v1": {
208
+ "action_mse_to_best": 0.35999264198917524,
209
+ "expert_success_rate": 0.8198198198198198,
210
+ "num_groups": 111,
211
+ "oracle_success_rate": 1.0,
212
+ "policy_expert_regret": 0.3294863182682175,
213
+ "policy_oracle_regret": 0.346653885401047,
214
+ "policy_rollout_progress": 0.8335262947791332,
215
+ "policy_rollout_success_rate": 0.8198198198198198,
216
+ "restore_max_error": 2.3795291781425476e-07
217
+ },
218
+ "StackCube-v1": {
219
+ "action_mse_to_best": 0.5005223249996101,
220
+ "expert_success_rate": 0.6923076923076923,
221
+ "num_groups": 91,
222
+ "oracle_success_rate": 0.8571428571428571,
223
+ "policy_expert_regret": 1.140796109542742,
224
+ "policy_oracle_regret": 1.3004579566992247,
225
+ "policy_rollout_progress": 0.3901473704602692,
226
+ "policy_rollout_success_rate": 0.15384615384615385,
227
+ "restore_max_error": 1.9371509552001953e-07
228
+ }
229
+ }
230
+ },
231
+ {
232
+ "seed": 2,
233
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_grid040045050_safe_margin0p20_mean_by_type_srcscorebonus0p02.json",
234
+ "num_groups": 575,
235
+ "selection_mode": "retrieval_residual",
236
+ "num_candidates": 18,
237
+ "candidate_sigma": 0.0,
238
+ "selection_margin": 0.2,
239
+ "prepend_policy_candidate": false,
240
+ "field_optim_steps": 0,
241
+ "field_optim_step_size": 0.0,
242
+ "field_optim_trust_radius": 0.0,
243
+ "field_optim_l2_penalty": 0.0,
244
+ "retrieval_neighbors": 4,
245
+ "retrieval_metric": "raw",
246
+ "retrieval_type_min_success": 0.0,
247
+ "retrieval_residual_min_source_progress": 0.0,
248
+ "retrieval_residual_source_progress_bonus_scale": 0.0,
249
+ "retrieval_residual_source_score_bonus_scale": 0.02,
250
+ "retrieval_residual_scale": 0.4,
251
+ "retrieval_residual_scales": [
252
+ 0.4,
253
+ 0.45,
254
+ 0.5
255
+ ],
256
+ "retrieval_residual_anchor": "expert",
257
+ "retrieval_residual_reduce": "mean_by_type",
258
+ "candidate_type_bonuses": {},
259
+ "selected_residual_scale_counts": {
260
+ "0.4": 548,
261
+ "0.45": 8,
262
+ "0.5": 19
263
+ },
264
+ "policy_rollout_success_rate": 0.3652173913043478,
265
+ "policy_rollout_progress": 0.5810759345204513,
266
+ "oracle_success_rate": 0.8765217391304347,
267
+ "action_mse_to_best": 0.4202554577985859,
268
+ "best_policy_val": {
269
+ "bc_loss": 0.11367896075050037,
270
+ "field_effect_loss": 0.009670218582161598,
271
+ "field_potential_loss": 0.2641640139950646,
272
+ "field_preference_loss": 0.5130490180518892,
273
+ "lattice_edges": 3833.3333333333335,
274
+ "progress_mae": 0.2021729110015763,
275
+ "rank_acc": 0.8333857821093665,
276
+ "rank_loss": 0.5130119257503085,
277
+ "regret_mae": 0.3958987047274907,
278
+ "success_accuracy": 0.8680730561415354,
279
+ "total_loss": 1.4394984311527677
280
+ },
281
+ "per_task": {
282
+ "LiftPegUpright-v1": {
283
+ "action_mse_to_best": 0.35520764285562717,
284
+ "expert_success_rate": 0.8229166666666666,
285
+ "num_groups": 96,
286
+ "oracle_success_rate": 0.9270833333333334,
287
+ "policy_expert_regret": 0.8476849501021206,
288
+ "policy_oracle_regret": 0.9539546399998168,
289
+ "policy_rollout_progress": 0.6463205499264101,
290
+ "policy_rollout_success_rate": 0.3333333333333333,
291
+ "restore_max_error": 1.955777406692505e-07
292
+ },
293
+ "PickCube-v1": {
294
+ "action_mse_to_best": 0.3359119386948419,
295
+ "expert_success_rate": 0.9444444444444444,
296
+ "num_groups": 198,
297
+ "oracle_success_rate": 0.9595959595959596,
298
+ "policy_expert_regret": 0.9808061452902327,
299
+ "policy_oracle_regret": 0.9878895958189411,
300
+ "policy_rollout_progress": 0.635682568443243,
301
+ "policy_rollout_success_rate": 0.3383838383838384,
302
+ "restore_max_error": 2.384185791015625e-07
303
+ },
304
+ "PullCube-v1": {
305
+ "action_mse_to_best": 0.6280809841222234,
306
+ "expert_success_rate": 0.24444444444444444,
307
+ "num_groups": 90,
308
+ "oracle_success_rate": 0.4666666666666667,
309
+ "policy_expert_regret": 0.2938225742705375,
310
+ "policy_oracle_regret": 0.5420316076182038,
311
+ "policy_rollout_progress": 0.3195290894903034,
312
+ "policy_rollout_success_rate": 0.2111111111111111,
313
+ "restore_max_error": 2.384185791015625e-07
314
+ },
315
+ "PushCube-v1": {
316
+ "action_mse_to_best": 0.3930995374327839,
317
+ "expert_success_rate": 0.8514851485148515,
318
+ "num_groups": 101,
319
+ "oracle_success_rate": 1.0,
320
+ "policy_expert_regret": 0.3976949640724919,
321
+ "policy_oracle_regret": 0.43470467125425244,
322
+ "policy_rollout_progress": 0.7930181010229753,
323
+ "policy_rollout_success_rate": 0.7722772277227723,
324
+ "restore_max_error": 2.384185791015625e-07
325
+ },
326
+ "StackCube-v1": {
327
+ "action_mse_to_best": 0.4978449867417415,
328
+ "expert_success_rate": 0.7666666666666667,
329
+ "num_groups": 90,
330
+ "oracle_success_rate": 0.9111111111111111,
331
+ "policy_expert_regret": 1.162587201760875,
332
+ "policy_oracle_regret": 1.311838612291548,
333
+ "policy_rollout_progress": 0.4150477194123798,
334
+ "policy_rollout_success_rate": 0.15555555555555556,
335
+ "restore_max_error": 2.2351741790771484e-07
336
+ }
337
+ }
338
+ }
339
+ ]
340
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_grid040045050_safe_margin0p20_mean_by_type_srcscorebonus0p02_summary.md ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # h=16 Best-Policy Checkpoint Rollout
2
+
3
+ Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
4
+ Objective: `near_miss_policy_bc5`
5
+ Result file: `policy_rollout_retrieval_residual_k4_grid040045050_safe_margin0p20_mean_by_type_srcscorebonus0p02.json`
6
+ Completed seeds: 3
7
+ Baseline h=4 policy success: 29.67%
8
+ Baseline h=16 rank-checkpoint success: 29.74%
9
+
10
+ Mean success: 35.30% +/- 1.06%
11
+ Gain vs h=16 rank checkpoint: +5.57%
12
+ Mean progress: 56.89%
13
+ Mean action MSE to best: 0.398
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | min source progress | source progress bonus | source score bonus | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---:|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 18 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.020 | 0.40 | 0.40,0.45,0.50 | 0.200 | 0.00 | 0 | 0.00 | 34.61% | 55.50% | 85.74% | 0.383 |
18
+ | 1 | retrieval_residual | 18 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.020 | 0.40 | 0.40,0.45,0.50 | 0.200 | 0.00 | 0 | 0.00 | 34.78% | 57.07% | 86.96% | 0.391 |
19
+ | 2 | retrieval_residual | 18 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.00 | 0.000 | 0.020 | 0.40 | 0.40,0.45,0.50 | 0.200 | 0.00 | 0 | 0.00 | 36.52% | 58.11% | 87.65% | 0.420 |