anhtld commited on
Commit
494f875
·
verified ·
1 Parent(s): e2ce49c

Auto-sync: 2026-06-28 07:49:25 (part 2)

Browse files
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p35_safe_noexpert_summary.json ADDED
@@ -0,0 +1,289 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_policy_anchor_scale0p35_safe_noexpert.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3373913043478261,
9
+ "std_success": 0.004601306627938417,
10
+ "mean_progress": 0.554790610001031,
11
+ "mean_action_mse_to_best": 0.41091299527203257,
12
+ "gain_vs_h4": 0.040691304347826085,
13
+ "gain_vs_h16_rank_checkpoint": 0.040000000000000036,
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_policy_anchor_scale0p35_safe_noexpert.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 16,
21
+ "candidate_sigma": 0.0,
22
+ "field_optim_steps": 0,
23
+ "field_optim_step_size": 0.0,
24
+ "field_optim_trust_radius": 0.0,
25
+ "field_optim_l2_penalty": 0.0,
26
+ "retrieval_neighbors": 1,
27
+ "retrieval_metric": "raw",
28
+ "retrieval_type_min_success": 0.0,
29
+ "retrieval_residual_scale": 0.35,
30
+ "retrieval_residual_anchor": "policy",
31
+ "policy_rollout_success_rate": 0.33217391304347826,
32
+ "policy_rollout_progress": 0.5461682539368453,
33
+ "oracle_success_rate": 0.8573913043478261,
34
+ "action_mse_to_best": 0.3909499480874966,
35
+ "best_policy_val": {
36
+ "bc_loss": 0.13721593966086706,
37
+ "field_effect_loss": 0.009290305380192068,
38
+ "field_potential_loss": 0.2666468388504452,
39
+ "field_preference_loss": 0.5130573478009965,
40
+ "lattice_edges": 3833.3333333333335,
41
+ "progress_mae": 0.1933159919248687,
42
+ "rank_acc": 0.8265031774838766,
43
+ "rank_loss": 0.5130523675017886,
44
+ "regret_mae": 0.3756548762321472,
45
+ "success_accuracy": 0.8773836526605818,
46
+ "total_loss": 1.5581054819954767
47
+ },
48
+ "per_task": {
49
+ "LiftPegUpright-v1": {
50
+ "action_mse_to_best": 0.32116451618962527,
51
+ "expert_success_rate": 0.8865979381443299,
52
+ "num_groups": 97,
53
+ "oracle_success_rate": 0.9278350515463918,
54
+ "policy_expert_regret": 1.0792337135248578,
55
+ "policy_oracle_regret": 1.1056491912333006,
56
+ "policy_rollout_progress": 0.570563180084081,
57
+ "policy_rollout_success_rate": 0.2268041237113402,
58
+ "restore_max_error": 4.76837158203125e-07
59
+ },
60
+ "PickCube-v1": {
61
+ "action_mse_to_best": 0.3100338529867048,
62
+ "expert_success_rate": 0.9375,
63
+ "num_groups": 208,
64
+ "oracle_success_rate": 0.9471153846153846,
65
+ "policy_expert_regret": 1.022702943139638,
66
+ "policy_oracle_regret": 1.0294804712399268,
67
+ "policy_rollout_progress": 0.5900551938595107,
68
+ "policy_rollout_success_rate": 0.3125,
69
+ "restore_max_error": 4.76837158203125e-07
70
+ },
71
+ "PullCube-v1": {
72
+ "action_mse_to_best": 0.6079731457806253,
73
+ "expert_success_rate": 0.19480519480519481,
74
+ "num_groups": 77,
75
+ "oracle_success_rate": 0.36363636363636365,
76
+ "policy_expert_regret": 0.19322007524405982,
77
+ "policy_oracle_regret": 0.36429626477712934,
78
+ "policy_rollout_progress": 0.3495055760994747,
79
+ "policy_rollout_success_rate": 0.24675324675324675,
80
+ "restore_max_error": 4.76837158203125e-07
81
+ },
82
+ "PushCube-v1": {
83
+ "action_mse_to_best": 0.3693886375886963,
84
+ "expert_success_rate": 0.8617021276595744,
85
+ "num_groups": 94,
86
+ "oracle_success_rate": 0.9893617021276596,
87
+ "policy_expert_regret": 0.5726731605035194,
88
+ "policy_oracle_regret": 0.6695628316795572,
89
+ "policy_rollout_progress": 0.6815009981076768,
90
+ "policy_rollout_success_rate": 0.648936170212766,
91
+ "restore_max_error": 4.76837158203125e-07
92
+ },
93
+ "StackCube-v1": {
94
+ "action_mse_to_best": 0.48100764141653224,
95
+ "expert_success_rate": 0.6767676767676768,
96
+ "num_groups": 99,
97
+ "oracle_success_rate": 0.8585858585858586,
98
+ "policy_expert_regret": 0.9669641819265153,
99
+ "policy_oracle_regret": 1.1261755732273815,
100
+ "policy_rollout_progress": 0.4545213539492,
101
+ "policy_rollout_success_rate": 0.24242424242424243,
102
+ "restore_max_error": 3.948807716369629e-07
103
+ }
104
+ }
105
+ },
106
+ {
107
+ "seed": 1,
108
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_policy_anchor_scale0p35_safe_noexpert.json",
109
+ "num_groups": 575,
110
+ "selection_mode": "retrieval_residual",
111
+ "num_candidates": 16,
112
+ "candidate_sigma": 0.0,
113
+ "field_optim_steps": 0,
114
+ "field_optim_step_size": 0.0,
115
+ "field_optim_trust_radius": 0.0,
116
+ "field_optim_l2_penalty": 0.0,
117
+ "retrieval_neighbors": 1,
118
+ "retrieval_metric": "raw",
119
+ "retrieval_type_min_success": 0.0,
120
+ "retrieval_residual_scale": 0.35,
121
+ "retrieval_residual_anchor": "policy",
122
+ "policy_rollout_success_rate": 0.3391304347826087,
123
+ "policy_rollout_progress": 0.5584853339502993,
124
+ "oracle_success_rate": 0.8695652173913043,
125
+ "action_mse_to_best": 0.4024488061164384,
126
+ "best_policy_val": {
127
+ "bc_loss": 0.1315989200439718,
128
+ "field_effect_loss": 0.010110370814800262,
129
+ "field_potential_loss": 0.31546536915832096,
130
+ "field_preference_loss": 0.4960637440284093,
131
+ "lattice_edges": 3833.3333333333335,
132
+ "progress_mae": 0.19023226698239645,
133
+ "rank_acc": 0.8255469501018524,
134
+ "rank_loss": 0.4960531195004781,
135
+ "regret_mae": 0.42156532737943864,
136
+ "success_accuracy": 0.8807473679383596,
137
+ "total_loss": 1.561984618504842
138
+ },
139
+ "per_task": {
140
+ "LiftPegUpright-v1": {
141
+ "action_mse_to_best": 0.3343446365005411,
142
+ "expert_success_rate": 0.8584070796460177,
143
+ "num_groups": 113,
144
+ "oracle_success_rate": 0.9380530973451328,
145
+ "policy_expert_regret": 1.0064765876900834,
146
+ "policy_oracle_regret": 1.0859059261537232,
147
+ "policy_rollout_progress": 0.6226141650592332,
148
+ "policy_rollout_success_rate": 0.21238938053097345,
149
+ "restore_max_error": 4.76837158203125e-07
150
+ },
151
+ "PickCube-v1": {
152
+ "action_mse_to_best": 0.34227522861455445,
153
+ "expert_success_rate": 0.9402173913043478,
154
+ "num_groups": 184,
155
+ "oracle_success_rate": 0.9456521739130435,
156
+ "policy_expert_regret": 1.0596443941543365,
157
+ "policy_oracle_regret": 1.0596683986292665,
158
+ "policy_rollout_progress": 0.5714469643906973,
159
+ "policy_rollout_success_rate": 0.31521739130434784,
160
+ "restore_max_error": 4.76837158203125e-07
161
+ },
162
+ "PullCube-v1": {
163
+ "action_mse_to_best": 0.5759768897765561,
164
+ "expert_success_rate": 0.25,
165
+ "num_groups": 76,
166
+ "oracle_success_rate": 0.40789473684210525,
167
+ "policy_expert_regret": 0.2709970424736017,
168
+ "policy_oracle_regret": 0.5241831319994832,
169
+ "policy_rollout_progress": 0.28006213075040204,
170
+ "policy_rollout_success_rate": 0.17105263157894737,
171
+ "restore_max_error": 4.76837158203125e-07
172
+ },
173
+ "PushCube-v1": {
174
+ "action_mse_to_best": 0.3723334042472882,
175
+ "expert_success_rate": 0.8198198198198198,
176
+ "num_groups": 111,
177
+ "oracle_success_rate": 1.0,
178
+ "policy_expert_regret": 0.364276541782929,
179
+ "policy_oracle_regret": 0.4308047579215454,
180
+ "policy_rollout_progress": 0.7944204673036799,
181
+ "policy_rollout_success_rate": 0.7747747747747747,
182
+ "restore_max_error": 4.76837158203125e-07
183
+ },
184
+ "StackCube-v1": {
185
+ "action_mse_to_best": 0.5004969893719305,
186
+ "expert_success_rate": 0.6923076923076923,
187
+ "num_groups": 91,
188
+ "oracle_success_rate": 0.8571428571428571,
189
+ "policy_expert_regret": 1.1210726157322035,
190
+ "policy_oracle_regret": 1.2932203698616762,
191
+ "policy_rollout_progress": 0.3973849572978177,
192
+ "policy_rollout_success_rate": 0.15384615384615385,
193
+ "restore_max_error": 3.948807716369629e-07
194
+ }
195
+ }
196
+ },
197
+ {
198
+ "seed": 2,
199
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_policy_anchor_scale0p35_safe_noexpert.json",
200
+ "num_groups": 575,
201
+ "selection_mode": "retrieval_residual",
202
+ "num_candidates": 16,
203
+ "candidate_sigma": 0.0,
204
+ "field_optim_steps": 0,
205
+ "field_optim_step_size": 0.0,
206
+ "field_optim_trust_radius": 0.0,
207
+ "field_optim_l2_penalty": 0.0,
208
+ "retrieval_neighbors": 1,
209
+ "retrieval_metric": "raw",
210
+ "retrieval_type_min_success": 0.0,
211
+ "retrieval_residual_scale": 0.35,
212
+ "retrieval_residual_anchor": "policy",
213
+ "policy_rollout_success_rate": 0.3408695652173913,
214
+ "policy_rollout_progress": 0.5597182421159486,
215
+ "oracle_success_rate": 0.8765217391304347,
216
+ "action_mse_to_best": 0.43934023161216273,
217
+ "best_policy_val": {
218
+ "bc_loss": 0.11367896075050037,
219
+ "field_effect_loss": 0.009670218582161598,
220
+ "field_potential_loss": 0.2641640139950646,
221
+ "field_preference_loss": 0.5130490180518892,
222
+ "lattice_edges": 3833.3333333333335,
223
+ "progress_mae": 0.2021729110015763,
224
+ "rank_acc": 0.8333857821093665,
225
+ "rank_loss": 0.5130119257503085,
226
+ "regret_mae": 0.3958987047274907,
227
+ "success_accuracy": 0.8680730561415354,
228
+ "total_loss": 1.4394984311527677
229
+ },
230
+ "per_task": {
231
+ "LiftPegUpright-v1": {
232
+ "action_mse_to_best": 0.35727682683985523,
233
+ "expert_success_rate": 0.8229166666666666,
234
+ "num_groups": 96,
235
+ "oracle_success_rate": 0.9270833333333334,
236
+ "policy_expert_regret": 0.8651364526400963,
237
+ "policy_oracle_regret": 0.9713857987274727,
238
+ "policy_rollout_progress": 0.6393060578654209,
239
+ "policy_rollout_success_rate": 0.3229166666666667,
240
+ "restore_max_error": 3.5762786865234375e-07
241
+ },
242
+ "PickCube-v1": {
243
+ "action_mse_to_best": 0.3646057000703583,
244
+ "expert_success_rate": 0.9444444444444444,
245
+ "num_groups": 198,
246
+ "oracle_success_rate": 0.9595959595959596,
247
+ "policy_expert_regret": 1.0403793521526485,
248
+ "policy_oracle_regret": 1.047462802681357,
249
+ "policy_rollout_progress": 0.6043060528434286,
250
+ "policy_rollout_success_rate": 0.30303030303030304,
251
+ "restore_max_error": 4.76837158203125e-07
252
+ },
253
+ "PullCube-v1": {
254
+ "action_mse_to_best": 0.6706881789697541,
255
+ "expert_success_rate": 0.24444444444444444,
256
+ "num_groups": 90,
257
+ "oracle_success_rate": 0.4666666666666667,
258
+ "policy_expert_regret": 0.319284730259743,
259
+ "policy_oracle_regret": 0.5459516277950671,
260
+ "policy_rollout_progress": 0.3085310854638616,
261
+ "policy_rollout_success_rate": 0.2,
262
+ "restore_max_error": 4.0978193283081055e-07
263
+ },
264
+ "PushCube-v1": {
265
+ "action_mse_to_best": 0.3935000709820502,
266
+ "expert_success_rate": 0.8514851485148515,
267
+ "num_groups": 101,
268
+ "oracle_success_rate": 1.0,
269
+ "policy_expert_regret": 0.41857334041949545,
270
+ "policy_oracle_regret": 0.49220996445948534,
271
+ "policy_rollout_progress": 0.765215778114772,
272
+ "policy_rollout_success_rate": 0.7425742574257426,
273
+ "restore_max_error": 4.76837158203125e-07
274
+ },
275
+ "StackCube-v1": {
276
+ "action_mse_to_best": 0.5113853989996844,
277
+ "expert_success_rate": 0.7666666666666667,
278
+ "num_groups": 90,
279
+ "oracle_success_rate": 0.9111111111111111,
280
+ "policy_expert_regret": 1.1934410845239958,
281
+ "policy_oracle_regret": 1.3555614117119048,
282
+ "policy_rollout_progress": 0.39730464352501765,
283
+ "policy_rollout_success_rate": 0.13333333333333333,
284
+ "restore_max_error": 4.76837158203125e-07
285
+ }
286
+ }
287
+ }
288
+ ]
289
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p35_safe_noexpert_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_policy_anchor_scale0p35_safe_noexpert.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: 33.74% +/- 0.46%
11
+ Gain vs h=16 rank checkpoint: +4.00%
12
+ Mean progress: 55.48%
13
+ Mean action MSE to best: 0.411
14
+
15
+ | seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 16 | 1 | raw | policy | 0.00 | 0.35 | 0.00 | 0 | 0.00 | 33.22% | 54.62% | 85.74% | 0.391 |
18
+ | 1 | retrieval_residual | 16 | 1 | raw | policy | 0.00 | 0.35 | 0.00 | 0 | 0.00 | 33.91% | 55.85% | 86.96% | 0.402 |
19
+ | 2 | retrieval_residual | 16 | 1 | raw | policy | 0.00 | 0.35 | 0.00 | 0 | 0.00 | 34.09% | 55.97% | 87.65% | 0.439 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p50_safe_noexpert_summary.json ADDED
@@ -0,0 +1,289 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_policy_anchor_scale0p50_safe_noexpert.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3368115942028986,
9
+ "std_success": 0.006584241560348136,
10
+ "mean_progress": 0.5550795713346451,
11
+ "mean_action_mse_to_best": 0.4211789192766815,
12
+ "gain_vs_h4": 0.04011159420289856,
13
+ "gain_vs_h16_rank_checkpoint": 0.03942028985507251,
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_policy_anchor_scale0p50_safe_noexpert.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 16,
21
+ "candidate_sigma": 0.0,
22
+ "field_optim_steps": 0,
23
+ "field_optim_step_size": 0.0,
24
+ "field_optim_trust_radius": 0.0,
25
+ "field_optim_l2_penalty": 0.0,
26
+ "retrieval_neighbors": 1,
27
+ "retrieval_metric": "raw",
28
+ "retrieval_type_min_success": 0.0,
29
+ "retrieval_residual_scale": 0.5,
30
+ "retrieval_residual_anchor": "policy",
31
+ "policy_rollout_success_rate": 0.33217391304347826,
32
+ "policy_rollout_progress": 0.5437370682017797,
33
+ "oracle_success_rate": 0.8573913043478261,
34
+ "action_mse_to_best": 0.40017282182188785,
35
+ "best_policy_val": {
36
+ "bc_loss": 0.13721593966086706,
37
+ "field_effect_loss": 0.009290305380192068,
38
+ "field_potential_loss": 0.2666468388504452,
39
+ "field_preference_loss": 0.5130573478009965,
40
+ "lattice_edges": 3833.3333333333335,
41
+ "progress_mae": 0.1933159919248687,
42
+ "rank_acc": 0.8265031774838766,
43
+ "rank_loss": 0.5130523675017886,
44
+ "regret_mae": 0.3756548762321472,
45
+ "success_accuracy": 0.8773836526605818,
46
+ "total_loss": 1.5581054819954767
47
+ },
48
+ "per_task": {
49
+ "LiftPegUpright-v1": {
50
+ "action_mse_to_best": 0.3308537736602281,
51
+ "expert_success_rate": 0.8865979381443299,
52
+ "num_groups": 97,
53
+ "oracle_success_rate": 0.9278350515463918,
54
+ "policy_expert_regret": 1.0794352798425044,
55
+ "policy_oracle_regret": 1.1058507575509475,
56
+ "policy_rollout_progress": 0.5703616137664342,
57
+ "policy_rollout_success_rate": 0.2268041237113402,
58
+ "restore_max_error": 4.76837158203125e-07
59
+ },
60
+ "PickCube-v1": {
61
+ "action_mse_to_best": 0.3139645314345566,
62
+ "expert_success_rate": 0.9375,
63
+ "num_groups": 208,
64
+ "oracle_success_rate": 0.9471153846153846,
65
+ "policy_expert_regret": 1.0052076294624175,
66
+ "policy_oracle_regret": 1.018630496888923,
67
+ "policy_rollout_progress": 0.5912897835951298,
68
+ "policy_rollout_success_rate": 0.32211538461538464,
69
+ "restore_max_error": 4.76837158203125e-07
70
+ },
71
+ "PullCube-v1": {
72
+ "action_mse_to_best": 0.6237646672245744,
73
+ "expert_success_rate": 0.19480519480519481,
74
+ "num_groups": 77,
75
+ "oracle_success_rate": 0.36363636363636365,
76
+ "policy_expert_regret": 0.19896453391917823,
77
+ "policy_oracle_regret": 0.41623874391163035,
78
+ "policy_rollout_progress": 0.3232533583846911,
79
+ "policy_rollout_success_rate": 0.22077922077922077,
80
+ "restore_max_error": 4.76837158203125e-07
81
+ },
82
+ "PushCube-v1": {
83
+ "action_mse_to_best": 0.3713671865060608,
84
+ "expert_success_rate": 0.8617021276595744,
85
+ "num_groups": 94,
86
+ "oracle_success_rate": 0.9893617021276596,
87
+ "policy_expert_regret": 0.5719942746961371,
88
+ "policy_oracle_regret": 0.6684110583143031,
89
+ "policy_rollout_progress": 0.6826527714729309,
90
+ "policy_rollout_success_rate": 0.648936170212766,
91
+ "restore_max_error": 4.76837158203125e-07
92
+ },
93
+ "StackCube-v1": {
94
+ "action_mse_to_best": 0.502662010669633,
95
+ "expert_success_rate": 0.6767676767676768,
96
+ "num_groups": 99,
97
+ "oracle_success_rate": 0.8585858585858586,
98
+ "policy_expert_regret": 0.9582149278033864,
99
+ "policy_oracle_regret": 1.1211562764765037,
100
+ "policy_rollout_progress": 0.457329228068843,
101
+ "policy_rollout_success_rate": 0.24242424242424243,
102
+ "restore_max_error": 3.948807716369629e-07
103
+ }
104
+ }
105
+ },
106
+ {
107
+ "seed": 1,
108
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_policy_anchor_scale0p50_safe_noexpert.json",
109
+ "num_groups": 575,
110
+ "selection_mode": "retrieval_residual",
111
+ "num_candidates": 16,
112
+ "candidate_sigma": 0.0,
113
+ "field_optim_steps": 0,
114
+ "field_optim_step_size": 0.0,
115
+ "field_optim_trust_radius": 0.0,
116
+ "field_optim_l2_penalty": 0.0,
117
+ "retrieval_neighbors": 1,
118
+ "retrieval_metric": "raw",
119
+ "retrieval_type_min_success": 0.0,
120
+ "retrieval_residual_scale": 0.5,
121
+ "retrieval_residual_anchor": "policy",
122
+ "policy_rollout_success_rate": 0.3339130434782609,
123
+ "policy_rollout_progress": 0.5564318828381922,
124
+ "oracle_success_rate": 0.8695652173913043,
125
+ "action_mse_to_best": 0.4136547504009112,
126
+ "best_policy_val": {
127
+ "bc_loss": 0.1315989200439718,
128
+ "field_effect_loss": 0.010110370814800262,
129
+ "field_potential_loss": 0.31546536915832096,
130
+ "field_preference_loss": 0.4960637440284093,
131
+ "lattice_edges": 3833.3333333333335,
132
+ "progress_mae": 0.19023226698239645,
133
+ "rank_acc": 0.8255469501018524,
134
+ "rank_loss": 0.4960531195004781,
135
+ "regret_mae": 0.42156532737943864,
136
+ "success_accuracy": 0.8807473679383596,
137
+ "total_loss": 1.561984618504842
138
+ },
139
+ "per_task": {
140
+ "LiftPegUpright-v1": {
141
+ "action_mse_to_best": 0.3306181395129688,
142
+ "expert_success_rate": 0.8584070796460177,
143
+ "num_groups": 113,
144
+ "oracle_success_rate": 0.9380530973451328,
145
+ "policy_expert_regret": 0.9980413421856619,
146
+ "policy_oracle_regret": 1.077886569157111,
147
+ "policy_rollout_progress": 0.6217839645337214,
148
+ "policy_rollout_success_rate": 0.22123893805309736,
149
+ "restore_max_error": 4.76837158203125e-07
150
+ },
151
+ "PickCube-v1": {
152
+ "action_mse_to_best": 0.34642869535752613,
153
+ "expert_success_rate": 0.9402173913043478,
154
+ "num_groups": 184,
155
+ "oracle_success_rate": 0.9456521739130435,
156
+ "policy_expert_regret": 1.1042199022561798,
157
+ "policy_oracle_regret": 1.1117568072235293,
158
+ "policy_rollout_progress": 0.549697133074717,
159
+ "policy_rollout_success_rate": 0.27717391304347827,
160
+ "restore_max_error": 4.76837158203125e-07
161
+ },
162
+ "PullCube-v1": {
163
+ "action_mse_to_best": 0.6093367022511206,
164
+ "expert_success_rate": 0.25,
165
+ "num_groups": 76,
166
+ "oracle_success_rate": 0.40789473684210525,
167
+ "policy_expert_regret": 0.2725211790223655,
168
+ "policy_oracle_regret": 0.5186420567333698,
169
+ "policy_rollout_progress": 0.2949514557656489,
170
+ "policy_rollout_success_rate": 0.18421052631578946,
171
+ "restore_max_error": 4.76837158203125e-07
172
+ },
173
+ "PushCube-v1": {
174
+ "action_mse_to_best": 0.38682191167865787,
175
+ "expert_success_rate": 0.8198198198198198,
176
+ "num_groups": 111,
177
+ "oracle_success_rate": 1.0,
178
+ "policy_expert_regret": 0.34637723379843943,
179
+ "policy_oracle_regret": 0.39610034146824397,
180
+ "policy_rollout_progress": 0.8111068657389632,
181
+ "policy_rollout_success_rate": 0.7927927927927928,
182
+ "restore_max_error": 4.76837158203125e-07
183
+ },
184
+ "StackCube-v1": {
185
+ "action_mse_to_best": 0.5219992329929393,
186
+ "expert_success_rate": 0.6923076923076923,
187
+ "num_groups": 91,
188
+ "oracle_success_rate": 0.8571428571428571,
189
+ "policy_expert_regret": 1.1218279090883967,
190
+ "policy_oracle_regret": 1.2939756632178694,
191
+ "policy_rollout_progress": 0.3966296639416244,
192
+ "policy_rollout_success_rate": 0.15384615384615385,
193
+ "restore_max_error": 3.948807716369629e-07
194
+ }
195
+ }
196
+ },
197
+ {
198
+ "seed": 2,
199
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_policy_anchor_scale0p50_safe_noexpert.json",
200
+ "num_groups": 575,
201
+ "selection_mode": "retrieval_residual",
202
+ "num_candidates": 16,
203
+ "candidate_sigma": 0.0,
204
+ "field_optim_steps": 0,
205
+ "field_optim_step_size": 0.0,
206
+ "field_optim_trust_radius": 0.0,
207
+ "field_optim_l2_penalty": 0.0,
208
+ "retrieval_neighbors": 1,
209
+ "retrieval_metric": "raw",
210
+ "retrieval_type_min_success": 0.0,
211
+ "retrieval_residual_scale": 0.5,
212
+ "retrieval_residual_anchor": "policy",
213
+ "policy_rollout_success_rate": 0.3443478260869565,
214
+ "policy_rollout_progress": 0.5650697629639636,
215
+ "oracle_success_rate": 0.8765217391304347,
216
+ "action_mse_to_best": 0.44970918560724543,
217
+ "best_policy_val": {
218
+ "bc_loss": 0.11367896075050037,
219
+ "field_effect_loss": 0.009670218582161598,
220
+ "field_potential_loss": 0.2641640139950646,
221
+ "field_preference_loss": 0.5130490180518892,
222
+ "lattice_edges": 3833.3333333333335,
223
+ "progress_mae": 0.2021729110015763,
224
+ "rank_acc": 0.8333857821093665,
225
+ "rank_loss": 0.5130119257503085,
226
+ "regret_mae": 0.3958987047274907,
227
+ "success_accuracy": 0.8680730561415354,
228
+ "total_loss": 1.4394984311527677
229
+ },
230
+ "per_task": {
231
+ "LiftPegUpright-v1": {
232
+ "action_mse_to_best": 0.36153392906514153,
233
+ "expert_success_rate": 0.8229166666666666,
234
+ "num_groups": 96,
235
+ "oracle_success_rate": 0.9270833333333334,
236
+ "policy_expert_regret": 0.8829973929872116,
237
+ "policy_oracle_regret": 0.9760098507006963,
238
+ "policy_rollout_progress": 0.6346820058921973,
239
+ "policy_rollout_success_rate": 0.3229166666666667,
240
+ "restore_max_error": 3.5762786865234375e-07
241
+ },
242
+ "PickCube-v1": {
243
+ "action_mse_to_best": 0.37210406688502945,
244
+ "expert_success_rate": 0.9444444444444444,
245
+ "num_groups": 198,
246
+ "oracle_success_rate": 0.9595959595959596,
247
+ "policy_expert_regret": 1.0698605704525805,
248
+ "policy_oracle_regret": 1.0769440209812888,
249
+ "policy_rollout_progress": 0.5929881188115387,
250
+ "policy_rollout_success_rate": 0.2777777777777778,
251
+ "restore_max_error": 4.76837158203125e-07
252
+ },
253
+ "PullCube-v1": {
254
+ "action_mse_to_best": 0.6711744833323691,
255
+ "expert_success_rate": 0.24444444444444444,
256
+ "num_groups": 90,
257
+ "oracle_success_rate": 0.4666666666666667,
258
+ "policy_expert_regret": 0.3375589284011059,
259
+ "policy_oracle_regret": 0.5657348255523377,
260
+ "policy_rollout_progress": 0.3060183207814892,
261
+ "policy_rollout_success_rate": 0.2,
262
+ "restore_max_error": 4.0978193283081055e-07
263
+ },
264
+ "PushCube-v1": {
265
+ "action_mse_to_best": 0.4170426452661505,
266
+ "expert_success_rate": 0.8514851485148515,
267
+ "num_groups": 101,
268
+ "oracle_success_rate": 1.0,
269
+ "policy_expert_regret": 0.36247750423332253,
270
+ "policy_oracle_regret": 0.3994881963670844,
271
+ "policy_rollout_progress": 0.8084325957121236,
272
+ "policy_rollout_success_rate": 0.7920792079207921,
273
+ "restore_max_error": 4.76837158203125e-07
274
+ },
275
+ "StackCube-v1": {
276
+ "action_mse_to_best": 0.5296878735431366,
277
+ "expert_success_rate": 0.7666666666666667,
278
+ "num_groups": 90,
279
+ "oracle_success_rate": 0.9111111111111111,
280
+ "policy_expert_regret": 1.1491315063503054,
281
+ "policy_oracle_regret": 1.3111187919974328,
282
+ "policy_rollout_progress": 0.41534058418538833,
283
+ "policy_rollout_success_rate": 0.15555555555555556,
284
+ "restore_max_error": 4.76837158203125e-07
285
+ }
286
+ }
287
+ }
288
+ ]
289
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p50_safe_noexpert_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_policy_anchor_scale0p50_safe_noexpert.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: 33.68% +/- 0.66%
11
+ Gain vs h=16 rank checkpoint: +3.94%
12
+ Mean progress: 55.51%
13
+ Mean action MSE to best: 0.421
14
+
15
+ | seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 16 | 1 | raw | policy | 0.00 | 0.50 | 0.00 | 0 | 0.00 | 33.22% | 54.37% | 85.74% | 0.400 |
18
+ | 1 | retrieval_residual | 16 | 1 | raw | policy | 0.00 | 0.50 | 0.00 | 0 | 0.00 | 33.39% | 55.64% | 86.96% | 0.414 |
19
+ | 2 | retrieval_residual | 16 | 1 | raw | policy | 0.00 | 0.50 | 0.00 | 0 | 0.00 | 34.43% | 56.51% | 87.65% | 0.450 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success010_summary.json ADDED
@@ -0,0 +1,289 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_scale0p35_type_success010.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3333333333333333,
9
+ "std_success": 0.012819330079707808,
10
+ "mean_progress": 0.551121375710042,
11
+ "mean_action_mse_to_best": 0.40250270315540443,
12
+ "gain_vs_h4": 0.036633333333333296,
13
+ "gain_vs_h16_rank_checkpoint": 0.035942028985507246,
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_scale0p35_type_success010.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 5,
21
+ "candidate_sigma": 0.0,
22
+ "field_optim_steps": 0,
23
+ "field_optim_step_size": 0.0,
24
+ "field_optim_trust_radius": 0.0,
25
+ "field_optim_l2_penalty": 0.0,
26
+ "retrieval_neighbors": 1,
27
+ "retrieval_metric": "raw",
28
+ "retrieval_type_min_success": 0.1,
29
+ "retrieval_residual_scale": 0.35,
30
+ "retrieval_residual_anchor": "expert",
31
+ "policy_rollout_success_rate": 0.32869565217391306,
32
+ "policy_rollout_progress": 0.5414418399556662,
33
+ "oracle_success_rate": 0.8573913043478261,
34
+ "action_mse_to_best": 0.3892755141927172,
35
+ "best_policy_val": {
36
+ "bc_loss": 0.13721593966086706,
37
+ "field_effect_loss": 0.009290305380192068,
38
+ "field_potential_loss": 0.2666468388504452,
39
+ "field_preference_loss": 0.5130573478009965,
40
+ "lattice_edges": 3833.3333333333335,
41
+ "progress_mae": 0.1933159919248687,
42
+ "rank_acc": 0.8265031774838766,
43
+ "rank_loss": 0.5130523675017886,
44
+ "regret_mae": 0.3756548762321472,
45
+ "success_accuracy": 0.8773836526605818,
46
+ "total_loss": 1.5581054819954767
47
+ },
48
+ "per_task": {
49
+ "LiftPegUpright-v1": {
50
+ "action_mse_to_best": 0.3268045182147822,
51
+ "expert_success_rate": 0.8865979381443299,
52
+ "num_groups": 97,
53
+ "oracle_success_rate": 0.9278350515463918,
54
+ "policy_expert_regret": 1.0923947420009632,
55
+ "policy_oracle_regret": 1.118810219709406,
56
+ "policy_rollout_progress": 0.567711429958491,
57
+ "policy_rollout_success_rate": 0.21649484536082475,
58
+ "restore_max_error": 4.76837158203125e-07
59
+ },
60
+ "PickCube-v1": {
61
+ "action_mse_to_best": 0.3092051282023581,
62
+ "expert_success_rate": 0.9375,
63
+ "num_groups": 208,
64
+ "oracle_success_rate": 0.9471153846153846,
65
+ "policy_expert_regret": 1.025236448774544,
66
+ "policy_oracle_regret": 1.0386593162010496,
67
+ "policy_rollout_progress": 0.5837433214227741,
68
+ "policy_rollout_success_rate": 0.30288461538461536,
69
+ "restore_max_error": 4.76837158203125e-07
70
+ },
71
+ "PullCube-v1": {
72
+ "action_mse_to_best": 0.6071846692206023,
73
+ "expert_success_rate": 0.19480519480519481,
74
+ "num_groups": 77,
75
+ "oracle_success_rate": 0.36363636363636365,
76
+ "policy_expert_regret": 0.21639780477959325,
77
+ "policy_oracle_regret": 0.40973563862710516,
78
+ "policy_rollout_progress": 0.33957307715295487,
79
+ "policy_rollout_success_rate": 0.23376623376623376,
80
+ "restore_max_error": 4.76837158203125e-07
81
+ },
82
+ "PushCube-v1": {
83
+ "action_mse_to_best": 0.3701509776663907,
84
+ "expert_success_rate": 0.8617021276595744,
85
+ "num_groups": 94,
86
+ "oracle_success_rate": 0.9893617021276596,
87
+ "policy_expert_regret": 0.5719673947450963,
88
+ "policy_oracle_regret": 0.668413888742315,
89
+ "policy_rollout_progress": 0.682649941044919,
90
+ "policy_rollout_success_rate": 0.648936170212766,
91
+ "restore_max_error": 4.76837158203125e-07
92
+ },
93
+ "StackCube-v1": {
94
+ "action_mse_to_best": 0.4673869120935448,
95
+ "expert_success_rate": 0.6767676767676768,
96
+ "num_groups": 99,
97
+ "oracle_success_rate": 0.8585858585858586,
98
+ "policy_expert_regret": 0.9574467485601251,
99
+ "policy_oracle_regret": 1.1085237165292103,
100
+ "policy_rollout_progress": 0.44975976781411603,
101
+ "policy_rollout_success_rate": 0.26262626262626265,
102
+ "restore_max_error": 3.948807716369629e-07
103
+ }
104
+ }
105
+ },
106
+ {
107
+ "seed": 1,
108
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_scale0p35_type_success010.json",
109
+ "num_groups": 575,
110
+ "selection_mode": "retrieval_residual",
111
+ "num_candidates": 5,
112
+ "candidate_sigma": 0.0,
113
+ "field_optim_steps": 0,
114
+ "field_optim_step_size": 0.0,
115
+ "field_optim_trust_radius": 0.0,
116
+ "field_optim_l2_penalty": 0.0,
117
+ "retrieval_neighbors": 1,
118
+ "retrieval_metric": "raw",
119
+ "retrieval_type_min_success": 0.1,
120
+ "retrieval_residual_scale": 0.35,
121
+ "retrieval_residual_anchor": "expert",
122
+ "policy_rollout_success_rate": 0.3234782608695652,
123
+ "policy_rollout_progress": 0.5480941037586687,
124
+ "oracle_success_rate": 0.8695652173913043,
125
+ "action_mse_to_best": 0.3963683258790685,
126
+ "best_policy_val": {
127
+ "bc_loss": 0.1315989200439718,
128
+ "field_effect_loss": 0.010110370814800262,
129
+ "field_potential_loss": 0.31546536915832096,
130
+ "field_preference_loss": 0.4960637440284093,
131
+ "lattice_edges": 3833.3333333333335,
132
+ "progress_mae": 0.19023226698239645,
133
+ "rank_acc": 0.8255469501018524,
134
+ "rank_loss": 0.4960531195004781,
135
+ "regret_mae": 0.42156532737943864,
136
+ "success_accuracy": 0.8807473679383596,
137
+ "total_loss": 1.561984618504842
138
+ },
139
+ "per_task": {
140
+ "LiftPegUpright-v1": {
141
+ "action_mse_to_best": 0.33575865034574426,
142
+ "expert_success_rate": 0.8584070796460177,
143
+ "num_groups": 113,
144
+ "oracle_success_rate": 0.9380530973451328,
145
+ "policy_expert_regret": 1.006511115131125,
146
+ "policy_oracle_regret": 1.0859404535947648,
147
+ "policy_rollout_progress": 0.6225796376181915,
148
+ "policy_rollout_success_rate": 0.21238938053097345,
149
+ "restore_max_error": 4.76837158203125e-07
150
+ },
151
+ "PickCube-v1": {
152
+ "action_mse_to_best": 0.3271206602170739,
153
+ "expert_success_rate": 0.9402173913043478,
154
+ "num_groups": 184,
155
+ "oracle_success_rate": 0.9456521739130435,
156
+ "policy_expert_regret": 1.1060857616780002,
157
+ "policy_oracle_regret": 1.1061097661529302,
158
+ "policy_rollout_progress": 0.5576142925192076,
159
+ "policy_rollout_success_rate": 0.2826086956521739,
160
+ "restore_max_error": 4.76837158203125e-07
161
+ },
162
+ "PullCube-v1": {
163
+ "action_mse_to_best": 0.5733223482966423,
164
+ "expert_success_rate": 0.25,
165
+ "num_groups": 76,
166
+ "oracle_success_rate": 0.40789473684210525,
167
+ "policy_expert_regret": 0.3021056252856445,
168
+ "policy_oracle_regret": 0.5768495806719569,
169
+ "policy_rollout_progress": 0.25590224674057316,
170
+ "policy_rollout_success_rate": 0.14473684210526316,
171
+ "restore_max_error": 4.76837158203125e-07
172
+ },
173
+ "PushCube-v1": {
174
+ "action_mse_to_best": 0.37514687879933967,
175
+ "expert_success_rate": 0.8198198198198198,
176
+ "num_groups": 111,
177
+ "oracle_success_rate": 1.0,
178
+ "policy_expert_regret": 0.36506524561224757,
179
+ "policy_oracle_regret": 0.44854555473671304,
180
+ "policy_rollout_progress": 0.7856886794975212,
181
+ "policy_rollout_success_rate": 0.7657657657657657,
182
+ "restore_max_error": 4.76837158203125e-07
183
+ },
184
+ "StackCube-v1": {
185
+ "action_mse_to_best": 0.48974787246354007,
186
+ "expert_success_rate": 0.6923076923076923,
187
+ "num_groups": 91,
188
+ "oracle_success_rate": 0.8571428571428571,
189
+ "policy_expert_regret": 1.1278909069823695,
190
+ "policy_oracle_regret": 1.3000386611118422,
191
+ "policy_rollout_progress": 0.39056666604765167,
192
+ "policy_rollout_success_rate": 0.15384615384615385,
193
+ "restore_max_error": 3.948807716369629e-07
194
+ }
195
+ }
196
+ },
197
+ {
198
+ "seed": 2,
199
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_scale0p35_type_success010.json",
200
+ "num_groups": 575,
201
+ "selection_mode": "retrieval_residual",
202
+ "num_candidates": 5,
203
+ "candidate_sigma": 0.0,
204
+ "field_optim_steps": 0,
205
+ "field_optim_step_size": 0.0,
206
+ "field_optim_trust_radius": 0.0,
207
+ "field_optim_l2_penalty": 0.0,
208
+ "retrieval_neighbors": 1,
209
+ "retrieval_metric": "raw",
210
+ "retrieval_type_min_success": 0.1,
211
+ "retrieval_residual_scale": 0.35,
212
+ "retrieval_residual_anchor": "expert",
213
+ "policy_rollout_success_rate": 0.34782608695652173,
214
+ "policy_rollout_progress": 0.5638281834157912,
215
+ "oracle_success_rate": 0.8765217391304347,
216
+ "action_mse_to_best": 0.42186426939442756,
217
+ "best_policy_val": {
218
+ "bc_loss": 0.11367896075050037,
219
+ "field_effect_loss": 0.009670218582161598,
220
+ "field_potential_loss": 0.2641640139950646,
221
+ "field_preference_loss": 0.5130490180518892,
222
+ "lattice_edges": 3833.3333333333335,
223
+ "progress_mae": 0.2021729110015763,
224
+ "rank_acc": 0.8333857821093665,
225
+ "rank_loss": 0.5130119257503085,
226
+ "regret_mae": 0.3958987047274907,
227
+ "success_accuracy": 0.8680730561415354,
228
+ "total_loss": 1.4394984311527677
229
+ },
230
+ "per_task": {
231
+ "LiftPegUpright-v1": {
232
+ "action_mse_to_best": 0.35727682683985523,
233
+ "expert_success_rate": 0.8229166666666666,
234
+ "num_groups": 96,
235
+ "oracle_success_rate": 0.9270833333333334,
236
+ "policy_expert_regret": 0.8651364526400963,
237
+ "policy_oracle_regret": 0.9713857987274727,
238
+ "policy_rollout_progress": 0.6393060578654209,
239
+ "policy_rollout_success_rate": 0.3229166666666667,
240
+ "restore_max_error": 3.5762786865234375e-07
241
+ },
242
+ "PickCube-v1": {
243
+ "action_mse_to_best": 0.3418835209061702,
244
+ "expert_success_rate": 0.9444444444444444,
245
+ "num_groups": 198,
246
+ "oracle_success_rate": 0.9595959595959596,
247
+ "policy_expert_regret": 1.054949007693189,
248
+ "policy_oracle_regret": 1.0620324582218974,
249
+ "policy_rollout_progress": 0.5998374074038983,
250
+ "policy_rollout_success_rate": 0.29292929292929293,
251
+ "restore_max_error": 4.76837158203125e-07
252
+ },
253
+ "PullCube-v1": {
254
+ "action_mse_to_best": 0.636862284772926,
255
+ "expert_success_rate": 0.24444444444444444,
256
+ "num_groups": 90,
257
+ "oracle_success_rate": 0.4666666666666667,
258
+ "policy_expert_regret": 0.29507516456974875,
259
+ "policy_oracle_regret": 0.5250848706397745,
260
+ "policy_rollout_progress": 0.31659532967540954,
261
+ "policy_rollout_success_rate": 0.2111111111111111,
262
+ "restore_max_error": 4.0978193283081055e-07
263
+ },
264
+ "PushCube-v1": {
265
+ "action_mse_to_best": 0.39332203817839667,
266
+ "expert_success_rate": 0.8514851485148515,
267
+ "num_groups": 101,
268
+ "oracle_success_rate": 1.0,
269
+ "policy_expert_regret": 0.3990432453332561,
270
+ "policy_oracle_regret": 0.472679869373246,
271
+ "policy_rollout_progress": 0.7748448831020015,
272
+ "policy_rollout_success_rate": 0.7524752475247525,
273
+ "restore_max_error": 4.76837158203125e-07
274
+ },
275
+ "StackCube-v1": {
276
+ "action_mse_to_best": 0.4837478988907403,
277
+ "expert_success_rate": 0.7666666666666667,
278
+ "num_groups": 90,
279
+ "oracle_success_rate": 0.9111111111111111,
280
+ "policy_expert_regret": 1.1536250283320746,
281
+ "policy_oracle_regret": 1.2897137719723912,
282
+ "policy_rollout_progress": 0.4145233819882075,
283
+ "policy_rollout_success_rate": 0.17777777777777778,
284
+ "restore_max_error": 4.76837158203125e-07
285
+ }
286
+ }
287
+ }
288
+ ]
289
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success010_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_scale0p35_type_success010.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: 33.33% +/- 1.28%
11
+ Gain vs h=16 rank checkpoint: +3.59%
12
+ Mean progress: 55.11%
13
+ Mean action MSE to best: 0.403
14
+
15
+ | seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 5 | 1 | raw | expert | 0.10 | 0.35 | 0.00 | 0 | 0.00 | 32.87% | 54.14% | 85.74% | 0.389 |
18
+ | 1 | retrieval_residual | 5 | 1 | raw | expert | 0.10 | 0.35 | 0.00 | 0 | 0.00 | 32.35% | 54.81% | 86.96% | 0.396 |
19
+ | 2 | retrieval_residual | 5 | 1 | raw | expert | 0.10 | 0.35 | 0.00 | 0 | 0.00 | 34.78% | 56.38% | 87.65% | 0.422 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success025_summary.json ADDED
@@ -0,0 +1,289 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_scale0p35_type_success025.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3327536231884058,
9
+ "std_success": 0.01458521231931493,
10
+ "mean_progress": 0.5508019064297084,
11
+ "mean_action_mse_to_best": 0.4013589472123894,
12
+ "gain_vs_h4": 0.03605362318840577,
13
+ "gain_vs_h16_rank_checkpoint": 0.03536231884057972,
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_scale0p35_type_success025.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 5,
21
+ "candidate_sigma": 0.0,
22
+ "field_optim_steps": 0,
23
+ "field_optim_step_size": 0.0,
24
+ "field_optim_trust_radius": 0.0,
25
+ "field_optim_l2_penalty": 0.0,
26
+ "retrieval_neighbors": 1,
27
+ "retrieval_metric": "raw",
28
+ "retrieval_type_min_success": 0.25,
29
+ "retrieval_residual_scale": 0.35,
30
+ "retrieval_residual_anchor": "expert",
31
+ "policy_rollout_success_rate": 0.32521739130434785,
32
+ "policy_rollout_progress": 0.5380419329922322,
33
+ "oracle_success_rate": 0.8573913043478261,
34
+ "action_mse_to_best": 0.3890486577226092,
35
+ "best_policy_val": {
36
+ "bc_loss": 0.13721593966086706,
37
+ "field_effect_loss": 0.009290305380192068,
38
+ "field_potential_loss": 0.2666468388504452,
39
+ "field_preference_loss": 0.5130573478009965,
40
+ "lattice_edges": 3833.3333333333335,
41
+ "progress_mae": 0.1933159919248687,
42
+ "rank_acc": 0.8265031774838766,
43
+ "rank_loss": 0.5130523675017886,
44
+ "regret_mae": 0.3756548762321472,
45
+ "success_accuracy": 0.8773836526605818,
46
+ "total_loss": 1.5581054819954767
47
+ },
48
+ "per_task": {
49
+ "LiftPegUpright-v1": {
50
+ "action_mse_to_best": 0.32533772717489257,
51
+ "expert_success_rate": 0.8865979381443299,
52
+ "num_groups": 97,
53
+ "oracle_success_rate": 0.9278350515463918,
54
+ "policy_expert_regret": 1.0923947420009632,
55
+ "policy_oracle_regret": 1.118810219709406,
56
+ "policy_rollout_progress": 0.567711429958491,
57
+ "policy_rollout_success_rate": 0.21649484536082475,
58
+ "restore_max_error": 4.76837158203125e-07
59
+ },
60
+ "PickCube-v1": {
61
+ "action_mse_to_best": 0.3092051282023581,
62
+ "expert_success_rate": 0.9375,
63
+ "num_groups": 208,
64
+ "oracle_success_rate": 0.9471153846153846,
65
+ "policy_expert_regret": 1.025236448774544,
66
+ "policy_oracle_regret": 1.0386593162010496,
67
+ "policy_rollout_progress": 0.5837433214227741,
68
+ "policy_rollout_success_rate": 0.30288461538461536,
69
+ "restore_max_error": 4.76837158203125e-07
70
+ },
71
+ "PullCube-v1": {
72
+ "action_mse_to_best": 0.607338386890176,
73
+ "expert_success_rate": 0.19480519480519481,
74
+ "num_groups": 77,
75
+ "oracle_success_rate": 0.36363636363636365,
76
+ "policy_expert_regret": 0.22117534107165773,
77
+ "policy_oracle_regret": 0.4382834872261125,
78
+ "policy_rollout_progress": 0.31418416151692247,
79
+ "policy_rollout_success_rate": 0.2077922077922078,
80
+ "restore_max_error": 4.76837158203125e-07
81
+ },
82
+ "PushCube-v1": {
83
+ "action_mse_to_best": 0.3701509776663907,
84
+ "expert_success_rate": 0.8617021276595744,
85
+ "num_groups": 94,
86
+ "oracle_success_rate": 0.9893617021276596,
87
+ "policy_expert_regret": 0.5719673947450963,
88
+ "policy_oracle_regret": 0.668413888742315,
89
+ "policy_rollout_progress": 0.682649941044919,
90
+ "policy_rollout_success_rate": 0.648936170212766,
91
+ "restore_max_error": 4.76837158203125e-07
92
+ },
93
+ "StackCube-v1": {
94
+ "action_mse_to_best": 0.4673869120935448,
95
+ "expert_success_rate": 0.6767676767676768,
96
+ "num_groups": 99,
97
+ "oracle_success_rate": 0.8585858585858586,
98
+ "policy_expert_regret": 0.9574467485601251,
99
+ "policy_oracle_regret": 1.1085237165292103,
100
+ "policy_rollout_progress": 0.44975976781411603,
101
+ "policy_rollout_success_rate": 0.26262626262626265,
102
+ "restore_max_error": 3.948807716369629e-07
103
+ }
104
+ }
105
+ },
106
+ {
107
+ "seed": 1,
108
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_scale0p35_type_success025.json",
109
+ "num_groups": 575,
110
+ "selection_mode": "retrieval_residual",
111
+ "num_candidates": 5,
112
+ "candidate_sigma": 0.0,
113
+ "field_optim_steps": 0,
114
+ "field_optim_step_size": 0.0,
115
+ "field_optim_trust_radius": 0.0,
116
+ "field_optim_l2_penalty": 0.0,
117
+ "retrieval_neighbors": 1,
118
+ "retrieval_metric": "raw",
119
+ "retrieval_type_min_success": 0.25,
120
+ "retrieval_residual_scale": 0.35,
121
+ "retrieval_residual_anchor": "expert",
122
+ "policy_rollout_success_rate": 0.3234782608695652,
123
+ "policy_rollout_progress": 0.5486927327529891,
124
+ "oracle_success_rate": 0.8695652173913043,
125
+ "action_mse_to_best": 0.39492910746970905,
126
+ "best_policy_val": {
127
+ "bc_loss": 0.1315989200439718,
128
+ "field_effect_loss": 0.010110370814800262,
129
+ "field_potential_loss": 0.31546536915832096,
130
+ "field_preference_loss": 0.4960637440284093,
131
+ "lattice_edges": 3833.3333333333335,
132
+ "progress_mae": 0.19023226698239645,
133
+ "rank_acc": 0.8255469501018524,
134
+ "rank_loss": 0.4960531195004781,
135
+ "regret_mae": 0.42156532737943864,
136
+ "success_accuracy": 0.8807473679383596,
137
+ "total_loss": 1.561984618504842
138
+ },
139
+ "per_task": {
140
+ "LiftPegUpright-v1": {
141
+ "action_mse_to_best": 0.32961989826300236,
142
+ "expert_success_rate": 0.8584070796460177,
143
+ "num_groups": 113,
144
+ "oracle_success_rate": 0.9380530973451328,
145
+ "policy_expert_regret": 0.9989857966393496,
146
+ "policy_oracle_regret": 1.0781879517356907,
147
+ "policy_rollout_progress": 0.6327733315725242,
148
+ "policy_rollout_success_rate": 0.22123893805309736,
149
+ "restore_max_error": 4.76837158203125e-07
150
+ },
151
+ "PickCube-v1": {
152
+ "action_mse_to_best": 0.3271206602170739,
153
+ "expert_success_rate": 0.9402173913043478,
154
+ "num_groups": 184,
155
+ "oracle_success_rate": 0.9456521739130435,
156
+ "policy_expert_regret": 1.1060857616780002,
157
+ "policy_oracle_regret": 1.1061097661529302,
158
+ "policy_rollout_progress": 0.5576142925192076,
159
+ "policy_rollout_success_rate": 0.2826086956521739,
160
+ "restore_max_error": 4.76837158203125e-07
161
+ },
162
+ "PullCube-v1": {
163
+ "action_mse_to_best": 0.5715608798751706,
164
+ "expert_success_rate": 0.25,
165
+ "num_groups": 76,
166
+ "oracle_success_rate": 0.40789473684210525,
167
+ "policy_expert_regret": 0.3506602184405844,
168
+ "policy_oracle_regret": 0.619953330042527,
169
+ "policy_rollout_progress": 0.24527493430497616,
170
+ "policy_rollout_success_rate": 0.13157894736842105,
171
+ "restore_max_error": 4.76837158203125e-07
172
+ },
173
+ "PushCube-v1": {
174
+ "action_mse_to_best": 0.37514687879933967,
175
+ "expert_success_rate": 0.8198198198198198,
176
+ "num_groups": 111,
177
+ "oracle_success_rate": 1.0,
178
+ "policy_expert_regret": 0.36506524561224757,
179
+ "policy_oracle_regret": 0.44854555473671304,
180
+ "policy_rollout_progress": 0.7856886794975212,
181
+ "policy_rollout_success_rate": 0.7657657657657657,
182
+ "restore_max_error": 4.76837158203125e-07
183
+ },
184
+ "StackCube-v1": {
185
+ "action_mse_to_best": 0.48974787246354007,
186
+ "expert_success_rate": 0.6923076923076923,
187
+ "num_groups": 91,
188
+ "oracle_success_rate": 0.8571428571428571,
189
+ "policy_expert_regret": 1.1278909069823695,
190
+ "policy_oracle_regret": 1.3000386611118422,
191
+ "policy_rollout_progress": 0.39056666604765167,
192
+ "policy_rollout_success_rate": 0.15384615384615385,
193
+ "restore_max_error": 3.948807716369629e-07
194
+ }
195
+ }
196
+ },
197
+ {
198
+ "seed": 2,
199
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_scale0p35_type_success025.json",
200
+ "num_groups": 575,
201
+ "selection_mode": "retrieval_residual",
202
+ "num_candidates": 5,
203
+ "candidate_sigma": 0.0,
204
+ "field_optim_steps": 0,
205
+ "field_optim_step_size": 0.0,
206
+ "field_optim_trust_radius": 0.0,
207
+ "field_optim_l2_penalty": 0.0,
208
+ "retrieval_neighbors": 1,
209
+ "retrieval_metric": "raw",
210
+ "retrieval_type_min_success": 0.25,
211
+ "retrieval_residual_scale": 0.35,
212
+ "retrieval_residual_anchor": "expert",
213
+ "policy_rollout_success_rate": 0.34956521739130436,
214
+ "policy_rollout_progress": 0.5656710535439039,
215
+ "oracle_success_rate": 0.8765217391304347,
216
+ "action_mse_to_best": 0.42009907644485,
217
+ "best_policy_val": {
218
+ "bc_loss": 0.11367896075050037,
219
+ "field_effect_loss": 0.009670218582161598,
220
+ "field_potential_loss": 0.2641640139950646,
221
+ "field_preference_loss": 0.5130490180518892,
222
+ "lattice_edges": 3833.3333333333335,
223
+ "progress_mae": 0.2021729110015763,
224
+ "rank_acc": 0.8333857821093665,
225
+ "rank_loss": 0.5130119257503085,
226
+ "regret_mae": 0.3958987047274907,
227
+ "success_accuracy": 0.8680730561415354,
228
+ "total_loss": 1.4394984311527677
229
+ },
230
+ "per_task": {
231
+ "LiftPegUpright-v1": {
232
+ "action_mse_to_best": 0.35361564260771655,
233
+ "expert_success_rate": 0.8229166666666666,
234
+ "num_groups": 96,
235
+ "oracle_success_rate": 0.9270833333333334,
236
+ "policy_expert_regret": 0.8470402403424183,
237
+ "policy_oracle_regret": 0.9533149556567272,
238
+ "policy_rollout_progress": 0.6469602342694998,
239
+ "policy_rollout_success_rate": 0.3333333333333333,
240
+ "restore_max_error": 3.5762786865234375e-07
241
+ },
242
+ "PickCube-v1": {
243
+ "action_mse_to_best": 0.3418835209061702,
244
+ "expert_success_rate": 0.9444444444444444,
245
+ "num_groups": 198,
246
+ "oracle_success_rate": 0.9595959595959596,
247
+ "policy_expert_regret": 1.054949007693189,
248
+ "policy_oracle_regret": 1.0620324582218974,
249
+ "policy_rollout_progress": 0.5998374074038983,
250
+ "policy_rollout_success_rate": 0.29292929292929293,
251
+ "restore_max_error": 4.76837158203125e-07
252
+ },
253
+ "PullCube-v1": {
254
+ "action_mse_to_best": 0.6294899263315731,
255
+ "expert_success_rate": 0.24444444444444444,
256
+ "num_groups": 90,
257
+ "oracle_success_rate": 0.4666666666666667,
258
+ "policy_expert_regret": 0.29368426450699997,
259
+ "policy_oracle_regret": 0.5414197320808347,
260
+ "policy_rollout_progress": 0.3202047673295561,
261
+ "policy_rollout_success_rate": 0.2111111111111111,
262
+ "restore_max_error": 4.0978193283081055e-07
263
+ },
264
+ "PushCube-v1": {
265
+ "action_mse_to_best": 0.39332203817839667,
266
+ "expert_success_rate": 0.8514851485148515,
267
+ "num_groups": 101,
268
+ "oracle_success_rate": 1.0,
269
+ "policy_expert_regret": 0.3990432453332561,
270
+ "policy_oracle_regret": 0.472679869373246,
271
+ "policy_rollout_progress": 0.7748448831020015,
272
+ "policy_rollout_success_rate": 0.7524752475247525,
273
+ "restore_max_error": 4.76837158203125e-07
274
+ },
275
+ "StackCube-v1": {
276
+ "action_mse_to_best": 0.4837478988907403,
277
+ "expert_success_rate": 0.7666666666666667,
278
+ "num_groups": 90,
279
+ "oracle_success_rate": 0.9111111111111111,
280
+ "policy_expert_regret": 1.1536250283320746,
281
+ "policy_oracle_regret": 1.2897137719723912,
282
+ "policy_rollout_progress": 0.4145233819882075,
283
+ "policy_rollout_success_rate": 0.17777777777777778,
284
+ "restore_max_error": 4.76837158203125e-07
285
+ }
286
+ }
287
+ }
288
+ ]
289
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success025_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_scale0p35_type_success025.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: 33.28% +/- 1.46%
11
+ Gain vs h=16 rank checkpoint: +3.54%
12
+ Mean progress: 55.08%
13
+ Mean action MSE to best: 0.401
14
+
15
+ | seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 5 | 1 | raw | expert | 0.25 | 0.35 | 0.00 | 0 | 0.00 | 32.52% | 53.80% | 85.74% | 0.389 |
18
+ | 1 | retrieval_residual | 5 | 1 | raw | expert | 0.25 | 0.35 | 0.00 | 0 | 0.00 | 32.35% | 54.87% | 86.96% | 0.395 |
19
+ | 2 | retrieval_residual | 5 | 1 | raw | expert | 0.25 | 0.35 | 0.00 | 0 | 0.00 | 34.96% | 56.57% | 87.65% | 0.420 |
scripts/eval_maniskill_policy_rollout.py CHANGED
@@ -73,6 +73,12 @@ def main(argv: list[str] | None = None) -> int:
73
  default=0,
74
  help="Base RNG seed for candidate perturbations.",
75
  )
 
 
 
 
 
 
76
  parser.add_argument(
77
  "--field-optim-steps",
78
  type=int,
@@ -153,6 +159,7 @@ def main(argv: list[str] | None = None) -> int:
153
  num_candidates=args.num_candidates,
154
  candidate_sigma=args.candidate_sigma,
155
  selection_seed=args.selection_seed,
 
156
  field_optim_steps=args.field_optim_steps,
157
  field_optim_step_size=args.field_optim_step_size,
158
  field_optim_trust_radius=args.field_optim_trust_radius,
 
73
  default=0,
74
  help="Base RNG seed for candidate perturbations.",
75
  )
76
+ parser.add_argument(
77
+ "--selection-margin",
78
+ type=float,
79
+ default=0.0,
80
+ help="Minimum field-potential advantage over candidate 0 required to leave it.",
81
+ )
82
  parser.add_argument(
83
  "--field-optim-steps",
84
  type=int,
 
159
  num_candidates=args.num_candidates,
160
  candidate_sigma=args.candidate_sigma,
161
  selection_seed=args.selection_seed,
162
+ selection_margin=args.selection_margin,
163
  field_optim_steps=args.field_optim_steps,
164
  field_optim_step_size=args.field_optim_step_size,
165
  field_optim_trust_radius=args.field_optim_trust_radius,
scripts/slurm/eval_maniskill_policy_rollout.sbatch CHANGED
@@ -44,6 +44,7 @@ SELECTION_MODE="${SELECTION_MODE:-policy}"
44
  NUM_CANDIDATES="${NUM_CANDIDATES:-1}"
45
  CANDIDATE_SIGMA="${CANDIDATE_SIGMA:-0.2}"
46
  SELECTION_SEED="${SELECTION_SEED:-0}"
 
47
  FIELD_OPTIM_STEPS="${FIELD_OPTIM_STEPS:-0}"
48
  FIELD_OPTIM_STEP_SIZE="${FIELD_OPTIM_STEP_SIZE:-0.05}"
49
  FIELD_OPTIM_TRUST_RADIUS="${FIELD_OPTIM_TRUST_RADIUS:-0.5}"
@@ -96,6 +97,7 @@ apptainer exec --nv \
96
  --num-candidates "$NUM_CANDIDATES" \
97
  --candidate-sigma "$CANDIDATE_SIGMA" \
98
  --selection-seed "$SELECTION_SEED" \
 
99
  --field-optim-steps "$FIELD_OPTIM_STEPS" \
100
  --field-optim-step-size "$FIELD_OPTIM_STEP_SIZE" \
101
  --field-optim-trust-radius "$FIELD_OPTIM_TRUST_RADIUS" \
 
44
  NUM_CANDIDATES="${NUM_CANDIDATES:-1}"
45
  CANDIDATE_SIGMA="${CANDIDATE_SIGMA:-0.2}"
46
  SELECTION_SEED="${SELECTION_SEED:-0}"
47
+ SELECTION_MARGIN="${SELECTION_MARGIN:-0.0}"
48
  FIELD_OPTIM_STEPS="${FIELD_OPTIM_STEPS:-0}"
49
  FIELD_OPTIM_STEP_SIZE="${FIELD_OPTIM_STEP_SIZE:-0.05}"
50
  FIELD_OPTIM_TRUST_RADIUS="${FIELD_OPTIM_TRUST_RADIUS:-0.5}"
 
97
  --num-candidates "$NUM_CANDIDATES" \
98
  --candidate-sigma "$CANDIDATE_SIGMA" \
99
  --selection-seed "$SELECTION_SEED" \
100
+ --selection-margin "$SELECTION_MARGIN" \
101
  --field-optim-steps "$FIELD_OPTIM_STEPS" \
102
  --field-optim-step-size "$FIELD_OPTIM_STEP_SIZE" \
103
  --field-optim-trust-radius "$FIELD_OPTIM_TRUST_RADIUS" \
scripts/slurm/summarize_h16_policy_ckpt.sbatch CHANGED
@@ -55,6 +55,7 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
55
  "selection_mode": data.get("selection_mode"),
56
  "num_candidates": data.get("num_candidates"),
57
  "candidate_sigma": data.get("candidate_sigma"),
 
58
  "field_optim_steps": data.get("field_optim_steps", 0),
59
  "field_optim_step_size": data.get("field_optim_step_size", 0.0),
60
  "field_optim_trust_radius": data.get("field_optim_trust_radius", 0.0),
@@ -124,12 +125,12 @@ lines = [
124
  f"Mean progress: {summary['mean_progress']:.2%}",
125
  f"Mean action MSE to best: {summary['mean_action_mse_to_best']:.3f}",
126
  "",
127
- "| seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |",
128
- "|---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|",
129
  ]
130
  for row in rows:
131
  lines.append(
132
- "| {seed} | {mode} | {k} | {retrieval} | {metric} | {anchor} | {min_success:.2f} | {scale:.2f} | {sigma:.2f} | {steps} | {trust:.2f} | "
133
  "{success:.2%} | {progress:.2%} | {oracle:.2%} | {mse:.3f} |".format(
134
  seed=row["seed"],
135
  mode=row.get("selection_mode") or "policy",
@@ -139,6 +140,7 @@ for row in rows:
139
  anchor=row.get("retrieval_residual_anchor") or "none",
140
  min_success=row.get("retrieval_type_min_success") or 0.0,
141
  scale=row.get("retrieval_residual_scale") or 0.0,
 
142
  sigma=row.get("candidate_sigma") or 0.0,
143
  steps=row.get("field_optim_steps") or 0,
144
  trust=row.get("field_optim_trust_radius") or 0.0,
 
55
  "selection_mode": data.get("selection_mode"),
56
  "num_candidates": data.get("num_candidates"),
57
  "candidate_sigma": data.get("candidate_sigma"),
58
+ "selection_margin": data.get("selection_margin", 0.0),
59
  "field_optim_steps": data.get("field_optim_steps", 0),
60
  "field_optim_step_size": data.get("field_optim_step_size", 0.0),
61
  "field_optim_trust_radius": data.get("field_optim_trust_radius", 0.0),
 
125
  f"Mean progress: {summary['mean_progress']:.2%}",
126
  f"Mean action MSE to best: {summary['mean_action_mse_to_best']:.3f}",
127
  "",
128
+ "| seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |",
129
+ "|---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|",
130
  ]
131
  for row in rows:
132
  lines.append(
133
+ "| {seed} | {mode} | {k} | {retrieval} | {metric} | {anchor} | {min_success:.2f} | {scale:.2f} | {margin:.3f} | {sigma:.2f} | {steps} | {trust:.2f} | "
134
  "{success:.2%} | {progress:.2%} | {oracle:.2%} | {mse:.3f} |".format(
135
  seed=row["seed"],
136
  mode=row.get("selection_mode") or "policy",
 
140
  anchor=row.get("retrieval_residual_anchor") or "none",
141
  min_success=row.get("retrieval_type_min_success") or 0.0,
142
  scale=row.get("retrieval_residual_scale") or 0.0,
143
+ margin=row.get("selection_margin") or 0.0,
144
  sigma=row.get("candidate_sigma") or 0.0,
145
  steps=row.get("field_optim_steps") or 0,
146
  trust=row.get("field_optim_trust_radius") or 0.0,