anhtld commited on
Commit
d688d49
·
verified ·
1 Parent(s): beb423d

Auto-sync: 2026-06-28 21:57:10

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p05_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,337 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_fieldsoftmax_grid_safe_margin0p05_noopbonus0p03.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3507246376811594,
9
+ "std_success": 0.011044961671453697,
10
+ "mean_progress": 0.5672646419938816,
11
+ "mean_action_mse_to_best": 0.40925245372929436,
12
+ "gain_vs_h4": 0.05402463768115939,
13
+ "gain_vs_h16_rank_checkpoint": 0.053333333333333344,
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_fieldsoftmax_grid_safe_margin0p05_noopbonus0p03.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 4,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.05,
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_scale": 1.0,
32
+ "retrieval_residual_scales": [
33
+ 0.35,
34
+ 0.4,
35
+ 0.45
36
+ ],
37
+ "retrieval_residual_anchor": "expert",
38
+ "retrieval_residual_reduce": "field_softmax",
39
+ "candidate_type_bonuses": {
40
+ "residual_no_op": 0.03
41
+ },
42
+ "selected_residual_scale_counts": {
43
+ "0.35": 19,
44
+ "0.4": 2,
45
+ "0.45": 46
46
+ },
47
+ "policy_rollout_success_rate": 0.3443478260869565,
48
+ "policy_rollout_progress": 0.5518923319673733,
49
+ "oracle_success_rate": 0.8573913043478261,
50
+ "action_mse_to_best": 0.3920214500521188,
51
+ "best_policy_val": {
52
+ "bc_loss": 0.13721593966086706,
53
+ "field_effect_loss": 0.009290305380192068,
54
+ "field_potential_loss": 0.2666468388504452,
55
+ "field_preference_loss": 0.5130573478009965,
56
+ "lattice_edges": 3833.3333333333335,
57
+ "progress_mae": 0.1933159919248687,
58
+ "rank_acc": 0.8265031774838766,
59
+ "rank_loss": 0.5130523675017886,
60
+ "regret_mae": 0.3756548762321472,
61
+ "success_accuracy": 0.8773836526605818,
62
+ "total_loss": 1.5581054819954767
63
+ },
64
+ "per_task": {
65
+ "LiftPegUpright-v1": {
66
+ "action_mse_to_best": 0.3247723626740968,
67
+ "expert_success_rate": 0.8865979381443299,
68
+ "num_groups": 97,
69
+ "oracle_success_rate": 0.9278350515463918,
70
+ "policy_expert_regret": 1.0917842305198158,
71
+ "policy_oracle_regret": 1.1181997082282586,
72
+ "policy_rollout_progress": 0.5683219414396384,
73
+ "policy_rollout_success_rate": 0.21649484536082475,
74
+ "restore_max_error": 2.384185791015625e-07
75
+ },
76
+ "PickCube-v1": {
77
+ "action_mse_to_best": 0.31156466976524544,
78
+ "expert_success_rate": 0.9375,
79
+ "num_groups": 208,
80
+ "oracle_success_rate": 0.9471153846153846,
81
+ "policy_expert_regret": 0.9720102318276006,
82
+ "policy_oracle_regret": 0.985433099254106,
83
+ "policy_rollout_progress": 0.6100641043068697,
84
+ "policy_rollout_success_rate": 0.33653846153846156,
85
+ "restore_max_error": 2.384185791015625e-07
86
+ },
87
+ "PullCube-v1": {
88
+ "action_mse_to_best": 0.6039306242744644,
89
+ "expert_success_rate": 0.19480519480519481,
90
+ "num_groups": 77,
91
+ "oracle_success_rate": 0.36363636363636365,
92
+ "policy_expert_regret": 0.22348312615695132,
93
+ "policy_oracle_regret": 0.3957931218463202,
94
+ "policy_rollout_progress": 0.3226016153121827,
95
+ "policy_rollout_success_rate": 0.22077922077922077,
96
+ "restore_max_error": 2.3096799850463867e-07
97
+ },
98
+ "PushCube-v1": {
99
+ "action_mse_to_best": 0.36951062051539724,
100
+ "expert_success_rate": 0.8617021276595744,
101
+ "num_groups": 94,
102
+ "oracle_success_rate": 0.9893617021276596,
103
+ "policy_expert_regret": 0.5508309629686335,
104
+ "policy_oracle_regret": 0.628381609441118,
105
+ "policy_rollout_progress": 0.7014056246014352,
106
+ "policy_rollout_success_rate": 0.6702127659574468,
107
+ "restore_max_error": 2.384185791015625e-07
108
+ },
109
+ "StackCube-v1": {
110
+ "action_mse_to_best": 0.4835081504225129,
111
+ "expert_success_rate": 0.6767676767676768,
112
+ "num_groups": 99,
113
+ "oracle_success_rate": 0.8585858585858586,
114
+ "policy_expert_regret": 0.9507471369673507,
115
+ "policy_oracle_regret": 1.109788482839411,
116
+ "policy_rollout_progress": 0.44995026034538194,
117
+ "policy_rollout_success_rate": 0.2727272727272727,
118
+ "restore_max_error": 2.384185791015625e-07
119
+ }
120
+ }
121
+ },
122
+ {
123
+ "seed": 1,
124
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p05_noopbonus0p03.json",
125
+ "num_groups": 575,
126
+ "selection_mode": "retrieval_residual",
127
+ "num_candidates": 4,
128
+ "candidate_sigma": 0.0,
129
+ "selection_margin": 0.05,
130
+ "prepend_policy_candidate": false,
131
+ "field_optim_steps": 0,
132
+ "field_optim_step_size": 0.0,
133
+ "field_optim_trust_radius": 0.0,
134
+ "field_optim_l2_penalty": 0.0,
135
+ "retrieval_neighbors": 4,
136
+ "retrieval_metric": "raw",
137
+ "retrieval_type_min_success": 0.0,
138
+ "retrieval_residual_scale": 1.0,
139
+ "retrieval_residual_scales": [
140
+ 0.35,
141
+ 0.4,
142
+ 0.45
143
+ ],
144
+ "retrieval_residual_anchor": "expert",
145
+ "retrieval_residual_reduce": "field_softmax",
146
+ "candidate_type_bonuses": {
147
+ "residual_no_op": 0.03
148
+ },
149
+ "selected_residual_scale_counts": {
150
+ "0.45": 38,
151
+ "0.35": 28,
152
+ "0.4": 8
153
+ },
154
+ "policy_rollout_success_rate": 0.3443478260869565,
155
+ "policy_rollout_progress": 0.5683723092014374,
156
+ "oracle_success_rate": 0.8695652173913043,
157
+ "action_mse_to_best": 0.3982289221664162,
158
+ "best_policy_val": {
159
+ "bc_loss": 0.1315989200439718,
160
+ "field_effect_loss": 0.010110370814800262,
161
+ "field_potential_loss": 0.31546536915832096,
162
+ "field_preference_loss": 0.4960637440284093,
163
+ "lattice_edges": 3833.3333333333335,
164
+ "progress_mae": 0.19023226698239645,
165
+ "rank_acc": 0.8255469501018524,
166
+ "rank_loss": 0.4960531195004781,
167
+ "regret_mae": 0.42156532737943864,
168
+ "success_accuracy": 0.8807473679383596,
169
+ "total_loss": 1.561984618504842
170
+ },
171
+ "per_task": {
172
+ "LiftPegUpright-v1": {
173
+ "action_mse_to_best": 0.32919689896310456,
174
+ "expert_success_rate": 0.8584070796460177,
175
+ "num_groups": 113,
176
+ "oracle_success_rate": 0.9380530973451328,
177
+ "policy_expert_regret": 0.988306692479986,
178
+ "policy_oracle_regret": 1.0694399586821024,
179
+ "policy_rollout_progress": 0.6326717671039885,
180
+ "policy_rollout_success_rate": 0.23008849557522124,
181
+ "restore_max_error": 2.384185791015625e-07
182
+ },
183
+ "PickCube-v1": {
184
+ "action_mse_to_best": 0.3315969341758477,
185
+ "expert_success_rate": 0.9402173913043478,
186
+ "num_groups": 184,
187
+ "oracle_success_rate": 0.9456521739130435,
188
+ "policy_expert_regret": 1.0729579917742345,
189
+ "policy_oracle_regret": 1.0729819962491647,
190
+ "policy_rollout_progress": 0.5722438574505403,
191
+ "policy_rollout_success_rate": 0.29347826086956524,
192
+ "restore_max_error": 2.384185791015625e-07
193
+ },
194
+ "PullCube-v1": {
195
+ "action_mse_to_best": 0.5645604024788267,
196
+ "expert_success_rate": 0.25,
197
+ "num_groups": 76,
198
+ "oracle_success_rate": 0.40789473684210525,
199
+ "policy_expert_regret": 0.2969187604704578,
200
+ "policy_oracle_regret": 0.5448774912786719,
201
+ "policy_rollout_progress": 0.28509269872876375,
202
+ "policy_rollout_success_rate": 0.17105263157894737,
203
+ "restore_max_error": 2.384185791015625e-07
204
+ },
205
+ "PushCube-v1": {
206
+ "action_mse_to_best": 0.36544963569791467,
207
+ "expert_success_rate": 0.8198198198198198,
208
+ "num_groups": 111,
209
+ "oracle_success_rate": 1.0,
210
+ "policy_expert_regret": 0.32972106874526086,
211
+ "policy_oracle_regret": 0.3468886358780904,
212
+ "policy_rollout_progress": 0.8332915443020898,
213
+ "policy_rollout_success_rate": 0.8198198198198198,
214
+ "restore_max_error": 2.3795291781425476e-07
215
+ },
216
+ "StackCube-v1": {
217
+ "action_mse_to_best": 0.5197477431169578,
218
+ "expert_success_rate": 0.6923076923076923,
219
+ "num_groups": 91,
220
+ "oracle_success_rate": 0.8571428571428571,
221
+ "policy_expert_regret": 1.136801811692479,
222
+ "policy_oracle_regret": 1.2964636588489615,
223
+ "policy_rollout_progress": 0.3941416683105322,
224
+ "policy_rollout_success_rate": 0.15384615384615385,
225
+ "restore_max_error": 1.9371509552001953e-07
226
+ }
227
+ }
228
+ },
229
+ {
230
+ "seed": 2,
231
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p05_noopbonus0p03.json",
232
+ "num_groups": 575,
233
+ "selection_mode": "retrieval_residual",
234
+ "num_candidates": 4,
235
+ "candidate_sigma": 0.0,
236
+ "selection_margin": 0.05,
237
+ "prepend_policy_candidate": false,
238
+ "field_optim_steps": 0,
239
+ "field_optim_step_size": 0.0,
240
+ "field_optim_trust_radius": 0.0,
241
+ "field_optim_l2_penalty": 0.0,
242
+ "retrieval_neighbors": 4,
243
+ "retrieval_metric": "raw",
244
+ "retrieval_type_min_success": 0.0,
245
+ "retrieval_residual_scale": 1.0,
246
+ "retrieval_residual_scales": [
247
+ 0.35,
248
+ 0.4,
249
+ 0.45
250
+ ],
251
+ "retrieval_residual_anchor": "expert",
252
+ "retrieval_residual_reduce": "field_softmax",
253
+ "candidate_type_bonuses": {
254
+ "residual_no_op": 0.03
255
+ },
256
+ "selected_residual_scale_counts": {
257
+ "0.35": 32,
258
+ "0.4": 11,
259
+ "0.45": 83
260
+ },
261
+ "policy_rollout_success_rate": 0.3634782608695652,
262
+ "policy_rollout_progress": 0.581529284812834,
263
+ "oracle_success_rate": 0.8765217391304347,
264
+ "action_mse_to_best": 0.43750698896934803,
265
+ "best_policy_val": {
266
+ "bc_loss": 0.11367896075050037,
267
+ "field_effect_loss": 0.009670218582161598,
268
+ "field_potential_loss": 0.2641640139950646,
269
+ "field_preference_loss": 0.5130490180518892,
270
+ "lattice_edges": 3833.3333333333335,
271
+ "progress_mae": 0.2021729110015763,
272
+ "rank_acc": 0.8333857821093665,
273
+ "rank_loss": 0.5130119257503085,
274
+ "regret_mae": 0.3958987047274907,
275
+ "success_accuracy": 0.8680730561415354,
276
+ "total_loss": 1.4394984311527677
277
+ },
278
+ "per_task": {
279
+ "LiftPegUpright-v1": {
280
+ "action_mse_to_best": 0.3614951341927129,
281
+ "expert_success_rate": 0.8229166666666666,
282
+ "num_groups": 96,
283
+ "oracle_success_rate": 0.9270833333333334,
284
+ "policy_expert_regret": 0.8476849501021206,
285
+ "policy_oracle_regret": 0.9539546399998168,
286
+ "policy_rollout_progress": 0.6463205499264101,
287
+ "policy_rollout_success_rate": 0.3333333333333333,
288
+ "restore_max_error": 1.955777406692505e-07
289
+ },
290
+ "PickCube-v1": {
291
+ "action_mse_to_best": 0.3666959795348271,
292
+ "expert_success_rate": 0.9444444444444444,
293
+ "num_groups": 198,
294
+ "oracle_success_rate": 0.9595959595959596,
295
+ "policy_expert_regret": 1.01903529606308,
296
+ "policy_oracle_regret": 1.0332649915921266,
297
+ "policy_rollout_progress": 0.6206102029730877,
298
+ "policy_rollout_success_rate": 0.30808080808080807,
299
+ "restore_max_error": 2.384185791015625e-07
300
+ },
301
+ "PullCube-v1": {
302
+ "action_mse_to_best": 0.6417228262457583,
303
+ "expert_success_rate": 0.24444444444444444,
304
+ "num_groups": 90,
305
+ "oracle_success_rate": 0.4666666666666667,
306
+ "policy_expert_regret": 0.27700459234830405,
307
+ "policy_oracle_regret": 0.5236944162183338,
308
+ "policy_rollout_progress": 0.33524703457951543,
309
+ "policy_rollout_success_rate": 0.23333333333333334,
310
+ "restore_max_error": 2.384185791015625e-07
311
+ },
312
+ "PushCube-v1": {
313
+ "action_mse_to_best": 0.40282966764551575,
314
+ "expert_success_rate": 0.8514851485148515,
315
+ "num_groups": 101,
316
+ "oracle_success_rate": 1.0,
317
+ "policy_expert_regret": 0.36133055020086835,
318
+ "policy_oracle_regret": 0.39834025738262896,
319
+ "policy_rollout_progress": 0.809580534696579,
320
+ "policy_rollout_success_rate": 0.7920792079207921,
321
+ "restore_max_error": 2.384185791015625e-07
322
+ },
323
+ "StackCube-v1": {
324
+ "action_mse_to_best": 0.5090703448073731,
325
+ "expert_success_rate": 0.7666666666666667,
326
+ "num_groups": 90,
327
+ "oracle_success_rate": 0.9111111111111111,
328
+ "policy_expert_regret": 1.162281624144978,
329
+ "policy_oracle_regret": 1.298549613191022,
330
+ "policy_rollout_progress": 0.41679865188068815,
331
+ "policy_rollout_success_rate": 0.16666666666666666,
332
+ "restore_max_error": 2.2351741790771484e-07
333
+ }
334
+ }
335
+ }
336
+ ]
337
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p05_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_fieldsoftmax_grid_safe_margin0p05_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: 35.07% +/- 1.10%
11
+ Gain vs h=16 rank checkpoint: +5.33%
12
+ Mean progress: 56.73%
13
+ Mean action MSE to best: 0.409
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.050 | 0.00 | 0 | 0.00 | 34.43% | 55.19% | 85.74% | 0.392 |
18
+ | 1 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.050 | 0.00 | 0 | 0.00 | 34.43% | 56.84% | 86.96% | 0.398 |
19
+ | 2 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.050 | 0.00 | 0 | 0.00 | 36.35% | 58.15% | 87.65% | 0.438 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03_summary.json ADDED
@@ -0,0 +1,337 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3518840579710145,
9
+ "std_success": 0.01233843284277841,
10
+ "mean_progress": 0.5673825791396276,
11
+ "mean_action_mse_to_best": 0.4014113438483058,
12
+ "gain_vs_h4": 0.0551840579710145,
13
+ "gain_vs_h16_rank_checkpoint": 0.05449275362318845,
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_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 4,
21
+ "candidate_sigma": 0.0,
22
+ "selection_margin": 0.1,
23
+ "prepend_policy_candidate": false,
24
+ "field_optim_steps": 0,
25
+ "field_optim_step_size": 0.0,
26
+ "field_optim_trust_radius": 0.0,
27
+ "field_optim_l2_penalty": 0.0,
28
+ "retrieval_neighbors": 4,
29
+ "retrieval_metric": "raw",
30
+ "retrieval_type_min_success": 0.0,
31
+ "retrieval_residual_scale": 1.0,
32
+ "retrieval_residual_scales": [
33
+ 0.35,
34
+ 0.4,
35
+ 0.45
36
+ ],
37
+ "retrieval_residual_anchor": "expert",
38
+ "retrieval_residual_reduce": "field_softmax",
39
+ "candidate_type_bonuses": {
40
+ "residual_no_op": 0.03
41
+ },
42
+ "selected_residual_scale_counts": {
43
+ "0.35": 11,
44
+ "0.4": 1,
45
+ "0.45": 26
46
+ },
47
+ "policy_rollout_success_rate": 0.34956521739130436,
48
+ "policy_rollout_progress": 0.5548038772388321,
49
+ "oracle_success_rate": 0.8573913043478261,
50
+ "action_mse_to_best": 0.38557657074831103,
51
+ "best_policy_val": {
52
+ "bc_loss": 0.13721593966086706,
53
+ "field_effect_loss": 0.009290305380192068,
54
+ "field_potential_loss": 0.2666468388504452,
55
+ "field_preference_loss": 0.5130573478009965,
56
+ "lattice_edges": 3833.3333333333335,
57
+ "progress_mae": 0.1933159919248687,
58
+ "rank_acc": 0.8265031774838766,
59
+ "rank_loss": 0.5130523675017886,
60
+ "regret_mae": 0.3756548762321472,
61
+ "success_accuracy": 0.8773836526605818,
62
+ "total_loss": 1.5581054819954767
63
+ },
64
+ "per_task": {
65
+ "LiftPegUpright-v1": {
66
+ "action_mse_to_best": 0.3214408095983654,
67
+ "expert_success_rate": 0.8865979381443299,
68
+ "num_groups": 97,
69
+ "oracle_success_rate": 0.9278350515463918,
70
+ "policy_expert_regret": 1.0917842305198158,
71
+ "policy_oracle_regret": 1.1181997082282586,
72
+ "policy_rollout_progress": 0.5683219414396384,
73
+ "policy_rollout_success_rate": 0.21649484536082475,
74
+ "restore_max_error": 2.384185791015625e-07
75
+ },
76
+ "PickCube-v1": {
77
+ "action_mse_to_best": 0.30813500876180255,
78
+ "expert_success_rate": 0.9375,
79
+ "num_groups": 208,
80
+ "oracle_success_rate": 0.9471153846153846,
81
+ "policy_expert_regret": 0.9544102810323238,
82
+ "policy_oracle_regret": 0.9678331484588293,
83
+ "policy_rollout_progress": 0.6180486704867619,
84
+ "policy_rollout_success_rate": 0.34615384615384615,
85
+ "restore_max_error": 2.384185791015625e-07
86
+ },
87
+ "PullCube-v1": {
88
+ "action_mse_to_best": 0.5976150262278396,
89
+ "expert_success_rate": 0.19480519480519481,
90
+ "num_groups": 77,
91
+ "oracle_success_rate": 0.36363636363636365,
92
+ "policy_expert_regret": 0.22294879098144638,
93
+ "policy_oracle_regret": 0.3965048290752842,
94
+ "policy_rollout_progress": 0.32105197882666797,
95
+ "policy_rollout_success_rate": 0.22077922077922077,
96
+ "restore_max_error": 2.3096799850463867e-07
97
+ },
98
+ "PushCube-v1": {
99
+ "action_mse_to_best": 0.36613771652287624,
100
+ "expert_success_rate": 0.8617021276595744,
101
+ "num_groups": 94,
102
+ "oracle_success_rate": 0.9893617021276596,
103
+ "policy_expert_regret": 0.5508309629686335,
104
+ "policy_oracle_regret": 0.628381609441118,
105
+ "policy_rollout_progress": 0.7014056246014352,
106
+ "policy_rollout_success_rate": 0.6702127659574468,
107
+ "restore_max_error": 2.384185791015625e-07
108
+ },
109
+ "StackCube-v1": {
110
+ "action_mse_to_best": 0.46466045913220655,
111
+ "expert_success_rate": 0.6767676767676768,
112
+ "num_groups": 99,
113
+ "oracle_success_rate": 0.8585858585858586,
114
+ "policy_expert_regret": 0.9393060180273923,
115
+ "policy_oracle_regret": 1.0983473638994525,
116
+ "policy_rollout_progress": 0.45129036918433024,
117
+ "policy_rollout_success_rate": 0.2828282828282828,
118
+ "restore_max_error": 2.384185791015625e-07
119
+ }
120
+ }
121
+ },
122
+ {
123
+ "seed": 1,
124
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03.json",
125
+ "num_groups": 575,
126
+ "selection_mode": "retrieval_residual",
127
+ "num_candidates": 4,
128
+ "candidate_sigma": 0.0,
129
+ "selection_margin": 0.1,
130
+ "prepend_policy_candidate": false,
131
+ "field_optim_steps": 0,
132
+ "field_optim_step_size": 0.0,
133
+ "field_optim_trust_radius": 0.0,
134
+ "field_optim_l2_penalty": 0.0,
135
+ "retrieval_neighbors": 4,
136
+ "retrieval_metric": "raw",
137
+ "retrieval_type_min_success": 0.0,
138
+ "retrieval_residual_scale": 1.0,
139
+ "retrieval_residual_scales": [
140
+ 0.35,
141
+ 0.4,
142
+ 0.45
143
+ ],
144
+ "retrieval_residual_anchor": "expert",
145
+ "retrieval_residual_reduce": "field_softmax",
146
+ "candidate_type_bonuses": {
147
+ "residual_no_op": 0.03
148
+ },
149
+ "selected_residual_scale_counts": {
150
+ "0.45": 21,
151
+ "0.35": 9,
152
+ "0.4": 1
153
+ },
154
+ "policy_rollout_success_rate": 0.3408695652173913,
155
+ "policy_rollout_progress": 0.5665490727742081,
156
+ "oracle_success_rate": 0.8695652173913043,
157
+ "action_mse_to_best": 0.3929036483346768,
158
+ "best_policy_val": {
159
+ "bc_loss": 0.1315989200439718,
160
+ "field_effect_loss": 0.010110370814800262,
161
+ "field_potential_loss": 0.31546536915832096,
162
+ "field_preference_loss": 0.4960637440284093,
163
+ "lattice_edges": 3833.3333333333335,
164
+ "progress_mae": 0.19023226698239645,
165
+ "rank_acc": 0.8255469501018524,
166
+ "rank_loss": 0.4960531195004781,
167
+ "regret_mae": 0.42156532737943864,
168
+ "success_accuracy": 0.8807473679383596,
169
+ "total_loss": 1.561984618504842
170
+ },
171
+ "per_task": {
172
+ "LiftPegUpright-v1": {
173
+ "action_mse_to_best": 0.3286477088911739,
174
+ "expert_success_rate": 0.8584070796460177,
175
+ "num_groups": 113,
176
+ "oracle_success_rate": 0.9380530973451328,
177
+ "policy_expert_regret": 0.9873896409979964,
178
+ "policy_oracle_regret": 1.0665916710828258,
179
+ "policy_rollout_progress": 0.6355200547032651,
180
+ "policy_rollout_success_rate": 0.23008849557522124,
181
+ "restore_max_error": 2.384185791015625e-07
182
+ },
183
+ "PickCube-v1": {
184
+ "action_mse_to_best": 0.3235761510086772,
185
+ "expert_success_rate": 0.9402173913043478,
186
+ "num_groups": 184,
187
+ "oracle_success_rate": 0.9456521739130435,
188
+ "policy_expert_regret": 1.0786200485941346,
189
+ "policy_oracle_regret": 1.0786440530690649,
190
+ "policy_rollout_progress": 0.5720165832393357,
191
+ "policy_rollout_success_rate": 0.28804347826086957,
192
+ "restore_max_error": 2.384185791015625e-07
193
+ },
194
+ "PullCube-v1": {
195
+ "action_mse_to_best": 0.5712389376406607,
196
+ "expert_success_rate": 0.25,
197
+ "num_groups": 76,
198
+ "oracle_success_rate": 0.40789473684210525,
199
+ "policy_expert_regret": 0.3006102462838355,
200
+ "policy_oracle_regret": 0.5470569631771037,
201
+ "policy_rollout_progress": 0.2786732331702584,
202
+ "policy_rollout_success_rate": 0.17105263157894737,
203
+ "restore_max_error": 2.384185791015625e-07
204
+ },
205
+ "PushCube-v1": {
206
+ "action_mse_to_best": 0.354367970816187,
207
+ "expert_success_rate": 0.8198198198198198,
208
+ "num_groups": 111,
209
+ "oracle_success_rate": 1.0,
210
+ "policy_expert_regret": 0.3303900519469837,
211
+ "policy_oracle_regret": 0.3638382223812309,
212
+ "policy_rollout_progress": 0.8253509668079583,
213
+ "policy_rollout_success_rate": 0.8108108108108109,
214
+ "restore_max_error": 2.3795291781425476e-07
215
+ },
216
+ "StackCube-v1": {
217
+ "action_mse_to_best": 0.5109383613280543,
218
+ "expert_success_rate": 0.6923076923076923,
219
+ "num_groups": 91,
220
+ "oracle_success_rate": 0.8571428571428571,
221
+ "policy_expert_regret": 1.136352531530045,
222
+ "policy_oracle_regret": 1.2960143786865277,
223
+ "policy_rollout_progress": 0.3945909484729662,
224
+ "policy_rollout_success_rate": 0.15384615384615385,
225
+ "restore_max_error": 1.9371509552001953e-07
226
+ }
227
+ }
228
+ },
229
+ {
230
+ "seed": 2,
231
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03.json",
232
+ "num_groups": 575,
233
+ "selection_mode": "retrieval_residual",
234
+ "num_candidates": 4,
235
+ "candidate_sigma": 0.0,
236
+ "selection_margin": 0.1,
237
+ "prepend_policy_candidate": false,
238
+ "field_optim_steps": 0,
239
+ "field_optim_step_size": 0.0,
240
+ "field_optim_trust_radius": 0.0,
241
+ "field_optim_l2_penalty": 0.0,
242
+ "retrieval_neighbors": 4,
243
+ "retrieval_metric": "raw",
244
+ "retrieval_type_min_success": 0.0,
245
+ "retrieval_residual_scale": 1.0,
246
+ "retrieval_residual_scales": [
247
+ 0.35,
248
+ 0.4,
249
+ 0.45
250
+ ],
251
+ "retrieval_residual_anchor": "expert",
252
+ "retrieval_residual_reduce": "field_softmax",
253
+ "candidate_type_bonuses": {
254
+ "residual_no_op": 0.03
255
+ },
256
+ "selected_residual_scale_counts": {
257
+ "0.45": 41,
258
+ "0.35": 12,
259
+ "0.4": 6
260
+ },
261
+ "policy_rollout_success_rate": 0.3652173913043478,
262
+ "policy_rollout_progress": 0.5807947874058427,
263
+ "oracle_success_rate": 0.8765217391304347,
264
+ "action_mse_to_best": 0.4257538124619295,
265
+ "best_policy_val": {
266
+ "bc_loss": 0.11367896075050037,
267
+ "field_effect_loss": 0.009670218582161598,
268
+ "field_potential_loss": 0.2641640139950646,
269
+ "field_preference_loss": 0.5130490180518892,
270
+ "lattice_edges": 3833.3333333333335,
271
+ "progress_mae": 0.2021729110015763,
272
+ "rank_acc": 0.8333857821093665,
273
+ "rank_loss": 0.5130119257503085,
274
+ "regret_mae": 0.3958987047274907,
275
+ "success_accuracy": 0.8680730561415354,
276
+ "total_loss": 1.4394984311527677
277
+ },
278
+ "per_task": {
279
+ "LiftPegUpright-v1": {
280
+ "action_mse_to_best": 0.35361564661434386,
281
+ "expert_success_rate": 0.8229166666666666,
282
+ "num_groups": 96,
283
+ "oracle_success_rate": 0.9270833333333334,
284
+ "policy_expert_regret": 0.8476849501021206,
285
+ "policy_oracle_regret": 0.9539546399998168,
286
+ "policy_rollout_progress": 0.6463205499264101,
287
+ "policy_rollout_success_rate": 0.3333333333333333,
288
+ "restore_max_error": 1.955777406692505e-07
289
+ },
290
+ "PickCube-v1": {
291
+ "action_mse_to_best": 0.34668580015575645,
292
+ "expert_success_rate": 0.9444444444444444,
293
+ "num_groups": 198,
294
+ "oracle_success_rate": 0.9595959595959596,
295
+ "policy_expert_regret": 1.0025666135203357,
296
+ "policy_oracle_regret": 1.0167963090493823,
297
+ "policy_rollout_progress": 0.6269778754148219,
298
+ "policy_rollout_success_rate": 0.3181818181818182,
299
+ "restore_max_error": 2.384185791015625e-07
300
+ },
301
+ "PullCube-v1": {
302
+ "action_mse_to_best": 0.6348918165597651,
303
+ "expert_success_rate": 0.24444444444444444,
304
+ "num_groups": 90,
305
+ "oracle_success_rate": 0.4666666666666667,
306
+ "policy_expert_regret": 0.2990271655083739,
307
+ "policy_oracle_regret": 0.546118191553756,
308
+ "policy_rollout_progress": 0.3235071391626156,
309
+ "policy_rollout_success_rate": 0.2222222222222222,
310
+ "restore_max_error": 2.384185791015625e-07
311
+ },
312
+ "PushCube-v1": {
313
+ "action_mse_to_best": 0.39641303030571134,
314
+ "expert_success_rate": 0.8514851485148515,
315
+ "num_groups": 101,
316
+ "oracle_success_rate": 1.0,
317
+ "policy_expert_regret": 0.36133055020086835,
318
+ "policy_oracle_regret": 0.39834025738262896,
319
+ "policy_rollout_progress": 0.809580534696579,
320
+ "policy_rollout_success_rate": 0.7920792079207921,
321
+ "restore_max_error": 2.384185791015625e-07
322
+ },
323
+ "StackCube-v1": {
324
+ "action_mse_to_best": 0.5004396900948551,
325
+ "expert_success_rate": 0.7666666666666667,
326
+ "num_groups": 90,
327
+ "oracle_success_rate": 0.9111111111111111,
328
+ "policy_expert_regret": 1.169243230422338,
329
+ "policy_oracle_regret": 1.3055112194683818,
330
+ "policy_rollout_progress": 0.4098370456033283,
331
+ "policy_rollout_success_rate": 0.16666666666666666,
332
+ "restore_max_error": 2.2351741790771484e-07
333
+ }
334
+ }
335
+ }
336
+ ]
337
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_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_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03.json`
6
+ Completed seeds: 3
7
+ Baseline h=4 policy success: 29.67%
8
+ Baseline h=16 rank-checkpoint success: 29.74%
9
+
10
+ Mean success: 35.19% +/- 1.23%
11
+ Gain vs h=16 rank checkpoint: +5.45%
12
+ Mean progress: 56.74%
13
+ Mean action MSE to best: 0.401
14
+
15
+ | seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---|---:|---|---|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.100 | 0.00 | 0 | 0.00 | 34.96% | 55.48% | 85.74% | 0.386 |
18
+ | 1 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.100 | 0.00 | 0 | 0.00 | 34.09% | 56.65% | 86.96% | 0.393 |
19
+ | 2 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.100 | 0.00 | 0 | 0.00 | 36.52% | 58.08% | 87.65% | 0.426 |