Hugging Face's logo

anhtld
/
vla

Model card Files
xet
 Community
anhtld commited on
Commit
dfe84cf
·
verified ·
1 Parent(s): 20f6dfd

Auto-sync: 2026-06-28 01:47:01 (part 3)

Browse files
Files changed (21) hide show
  1. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k32_sigma0p35_summary.json +280 -0
  2. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k32_sigma0p35_summary.md +19 -0
  3. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k64_sigma0p50_summary.json +280 -0
  4. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k64_sigma0p50_summary.md +19 -0
  5. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_summary.json +280 -0
  6. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_summary.md +19 -0
  7. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_summary.json +280 -0
  8. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_summary.md +19 -0
  9. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p75_summary.json +280 -0
  10. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p75_summary.md +19 -0
  11. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale1p25_summary.json +280 -0
  12. results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale1p25_summary.md +19 -0
  13. results/paper_core_results.md +22 -12
  14. results/paper_story_memo.md +42 -50
  15. results/paper_table_status.json +134 -1
  16. results/paper_table_status.md +8 -1
  17. scripts/build_paper_table_status.py +68 -1
  18. scripts/eval_maniskill_policy_rollout.py +8 -0
  19. scripts/slurm/eval_maniskill_policy_rollout.sbatch +2 -0
  20. scripts/slurm/summarize_h16_policy_ckpt.sbatch +5 -3
  21. tests/test_maniskill_policy_rollout.py +80 -0
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k32_sigma0p35_summary.json ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_hybrid_k32_sigma0p35.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3130434782608696,
9
+ "std_success": 0.013803919883815262,
10
+ "mean_progress": 0.5420378264707878,
11
+ "mean_action_mse_to_best": 0.5544718223749457,
12
+ "gain_vs_h4": 0.01634347826086957,
13
+ "gain_vs_h16_rank_checkpoint": 0.01565217391304352,
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_hybrid_k32_sigma0p35.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 47,
21
+ "candidate_sigma": 0.35,
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_residual_scale": 1.0,
28
+ "policy_rollout_success_rate": 0.3026086956521739,
29
+ "policy_rollout_progress": 0.5266784923049904,
30
+ "oracle_success_rate": 0.8573913043478261,
31
+ "action_mse_to_best": 0.6270939950713806,
32
+ "best_policy_val": {
33
+ "bc_loss": 0.13721593966086706,
34
+ "field_effect_loss": 0.009290305380192068,
35
+ "field_potential_loss": 0.2666468388504452,
36
+ "field_preference_loss": 0.5130573478009965,
37
+ "lattice_edges": 3833.3333333333335,
38
+ "progress_mae": 0.1933159919248687,
39
+ "rank_acc": 0.8265031774838766,
40
+ "rank_loss": 0.5130523675017886,
41
+ "regret_mae": 0.3756548762321472,
42
+ "success_accuracy": 0.8773836526605818,
43
+ "total_loss": 1.5581054819954767
44
+ },
45
+ "per_task": {
46
+ "LiftPegUpright-v1": {
47
+ "action_mse_to_best": 1.1160289499495832,
48
+ "expert_success_rate": 0.8865979381443299,
49
+ "num_groups": 97,
50
+ "oracle_success_rate": 0.9278350515463918,
51
+ "policy_expert_regret": 1.2770741686378557,
52
+ "policy_oracle_regret": 1.3166886242394595,
53
+ "policy_rollout_progress": 0.4729258089335923,
54
+ "policy_rollout_success_rate": 0.1134020618556701,
55
+ "restore_max_error": 4.76837158203125e-07
56
+ },
57
+ "PickCube-v1": {
58
+ "action_mse_to_best": 0.3464970012820021,
59
+ "expert_success_rate": 0.9375,
60
+ "num_groups": 208,
61
+ "oracle_success_rate": 0.9471153846153846,
62
+ "policy_expert_regret": 0.9952310958984666,
63
+ "policy_oracle_regret": 1.008653963324972,
64
+ "policy_rollout_progress": 0.5964586248513885,
65
+ "policy_rollout_success_rate": 0.3269230769230769,
66
+ "restore_max_error": 4.76837158203125e-07
67
+ },
68
+ "PullCube-v1": {
69
+ "action_mse_to_best": 0.6548769665615899,
70
+ "expert_success_rate": 0.19480519480519481,
71
+ "num_groups": 77,
72
+ "oracle_success_rate": 0.36363636363636365,
73
+ "policy_expert_regret": 0.19580742415923674,
74
+ "policy_oracle_regret": 0.38972585405489163,
75
+ "policy_rollout_progress": 0.3309086969624237,
76
+ "policy_rollout_success_rate": 0.22077922077922077,
77
+ "restore_max_error": 4.76837158203125e-07
78
+ },
79
+ "PushCube-v1": {
80
+ "action_mse_to_best": 0.40216122555764433,
81
+ "expert_success_rate": 0.8617021276595744,
82
+ "num_groups": 94,
83
+ "oracle_success_rate": 0.9893617021276596,
84
+ "policy_expert_regret": 0.5730919750764015,
85
+ "policy_oracle_regret": 0.6506427363195317,
86
+ "policy_rollout_progress": 0.6897827955953618,
87
+ "policy_rollout_success_rate": 0.6595744680851063,
88
+ "restore_max_error": 4.76837158203125e-07
89
+ },
90
+ "StackCube-v1": {
91
+ "action_mse_to_best": 0.9295371830971404,
92
+ "expert_success_rate": 0.6767676767676768,
93
+ "num_groups": 99,
94
+ "oracle_success_rate": 0.8585858585858586,
95
+ "policy_expert_regret": 1.0469527790943782,
96
+ "policy_oracle_regret": 1.2291584016397745,
97
+ "policy_rollout_progress": 0.4301351837136529,
98
+ "policy_rollout_success_rate": 0.16161616161616163,
99
+ "restore_max_error": 3.948807716369629e-07
100
+ }
101
+ }
102
+ },
103
+ {
104
+ "seed": 1,
105
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_hybrid_k32_sigma0p35.json",
106
+ "num_groups": 575,
107
+ "selection_mode": "retrieval_residual",
108
+ "num_candidates": 47,
109
+ "candidate_sigma": 0.35,
110
+ "field_optim_steps": 0,
111
+ "field_optim_step_size": 0.0,
112
+ "field_optim_trust_radius": 0.0,
113
+ "field_optim_l2_penalty": 0.0,
114
+ "retrieval_neighbors": 1,
115
+ "retrieval_residual_scale": 1.0,
116
+ "policy_rollout_success_rate": 0.30782608695652175,
117
+ "policy_rollout_progress": 0.5425913009938339,
118
+ "oracle_success_rate": 0.8695652173913043,
119
+ "action_mse_to_best": 0.5018047516209924,
120
+ "best_policy_val": {
121
+ "bc_loss": 0.1315989200439718,
122
+ "field_effect_loss": 0.010110370814800262,
123
+ "field_potential_loss": 0.31546536915832096,
124
+ "field_preference_loss": 0.4960637440284093,
125
+ "lattice_edges": 3833.3333333333335,
126
+ "progress_mae": 0.19023226698239645,
127
+ "rank_acc": 0.8255469501018524,
128
+ "rank_loss": 0.4960531195004781,
129
+ "regret_mae": 0.42156532737943864,
130
+ "success_accuracy": 0.8807473679383596,
131
+ "total_loss": 1.561984618504842
132
+ },
133
+ "per_task": {
134
+ "LiftPegUpright-v1": {
135
+ "action_mse_to_best": 0.4475946618837461,
136
+ "expert_success_rate": 0.8584070796460177,
137
+ "num_groups": 113,
138
+ "oracle_success_rate": 0.9380530973451328,
139
+ "policy_expert_regret": 1.0465806051429394,
140
+ "policy_oracle_regret": 1.1370370254843636,
141
+ "policy_rollout_progress": 0.600472930390223,
142
+ "policy_rollout_success_rate": 0.19469026548672566,
143
+ "restore_max_error": 4.76837158203125e-07
144
+ },
145
+ "PickCube-v1": {
146
+ "action_mse_to_best": 0.43887720265142294,
147
+ "expert_success_rate": 0.9402173913043478,
148
+ "num_groups": 184,
149
+ "oracle_success_rate": 0.9456521739130435,
150
+ "policy_expert_regret": 1.1457417606497589,
151
+ "policy_oracle_regret": 1.145765765124689,
152
+ "policy_rollout_progress": 0.5461787912670685,
153
+ "policy_rollout_success_rate": 0.2391304347826087,
154
+ "restore_max_error": 4.76837158203125e-07
155
+ },
156
+ "PullCube-v1": {
157
+ "action_mse_to_best": 0.6761482136422082,
158
+ "expert_success_rate": 0.25,
159
+ "num_groups": 76,
160
+ "oracle_success_rate": 0.40789473684210525,
161
+ "policy_expert_regret": 0.3213816526419434,
162
+ "policy_oracle_regret": 0.5480846182418693,
163
+ "policy_rollout_progress": 0.27930643545512696,
164
+ "policy_rollout_success_rate": 0.17105263157894737,
165
+ "restore_max_error": 4.76837158203125e-07
166
+ },
167
+ "PushCube-v1": {
168
+ "action_mse_to_best": 0.43849160747082383,
169
+ "expert_success_rate": 0.8198198198198198,
170
+ "num_groups": 111,
171
+ "oracle_success_rate": 1.0,
172
+ "policy_expert_regret": 0.4024832792378761,
173
+ "policy_oracle_regret": 0.4686909187752921,
174
+ "policy_rollout_progress": 0.7745523244679511,
175
+ "policy_rollout_success_rate": 0.7567567567567568,
176
+ "restore_max_error": 4.76837158203125e-07
177
+ },
178
+ "StackCube-v1": {
179
+ "action_mse_to_best": 0.6279812904975408,
180
+ "expert_success_rate": 0.6923076923076923,
181
+ "num_groups": 91,
182
+ "oracle_success_rate": 0.8571428571428571,
183
+ "policy_expert_regret": 1.1180502957367635,
184
+ "policy_oracle_regret": 1.2901980498662362,
185
+ "policy_rollout_progress": 0.40040727729325765,
186
+ "policy_rollout_success_rate": 0.15384615384615385,
187
+ "restore_max_error": 3.948807716369629e-07
188
+ }
189
+ }
190
+ },
191
+ {
192
+ "seed": 2,
193
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_hybrid_k32_sigma0p35.json",
194
+ "num_groups": 575,
195
+ "selection_mode": "retrieval_residual",
196
+ "num_candidates": 47,
197
+ "candidate_sigma": 0.35,
198
+ "field_optim_steps": 0,
199
+ "field_optim_step_size": 0.0,
200
+ "field_optim_trust_radius": 0.0,
201
+ "field_optim_l2_penalty": 0.0,
202
+ "retrieval_neighbors": 1,
203
+ "retrieval_residual_scale": 1.0,
204
+ "policy_rollout_success_rate": 0.32869565217391306,
205
+ "policy_rollout_progress": 0.556843686113539,
206
+ "oracle_success_rate": 0.8765217391304347,
207
+ "action_mse_to_best": 0.5345167204324642,
208
+ "best_policy_val": {
209
+ "bc_loss": 0.11367896075050037,
210
+ "field_effect_loss": 0.009670218582161598,
211
+ "field_potential_loss": 0.2641640139950646,
212
+ "field_preference_loss": 0.5130490180518892,
213
+ "lattice_edges": 3833.3333333333335,
214
+ "progress_mae": 0.2021729110015763,
215
+ "rank_acc": 0.8333857821093665,
216
+ "rank_loss": 0.5130119257503085,
217
+ "regret_mae": 0.3958987047274907,
218
+ "success_accuracy": 0.8680730561415354,
219
+ "total_loss": 1.4394984311527677
220
+ },
221
+ "per_task": {
222
+ "LiftPegUpright-v1": {
223
+ "action_mse_to_best": 0.4214031276496826,
224
+ "expert_success_rate": 0.8229166666666666,
225
+ "num_groups": 96,
226
+ "oracle_success_rate": 0.9270833333333334,
227
+ "policy_expert_regret": 0.902202141781648,
228
+ "policy_oracle_regret": 1.0084768570959568,
229
+ "policy_rollout_progress": 0.6230483328302702,
230
+ "policy_rollout_success_rate": 0.3020833333333333,
231
+ "restore_max_error": 3.5762786865234375e-07
232
+ },
233
+ "PickCube-v1": {
234
+ "action_mse_to_best": 0.4334832938820726,
235
+ "expert_success_rate": 0.9444444444444444,
236
+ "num_groups": 198,
237
+ "oracle_success_rate": 0.9595959595959596,
238
+ "policy_expert_regret": 1.093938429745836,
239
+ "policy_oracle_regret": 1.115265645585352,
240
+ "policy_rollout_progress": 0.587008260444484,
241
+ "policy_rollout_success_rate": 0.25252525252525254,
242
+ "restore_max_error": 4.76837158203125e-07
243
+ },
244
+ "PullCube-v1": {
245
+ "action_mse_to_best": 0.7262630649738842,
246
+ "expert_success_rate": 0.24444444444444444,
247
+ "num_groups": 90,
248
+ "oracle_success_rate": 0.4666666666666667,
249
+ "policy_expert_regret": 0.3447766973740525,
250
+ "policy_oracle_regret": 0.592945381005605,
251
+ "policy_rollout_progress": 0.3033671165506045,
252
+ "policy_rollout_success_rate": 0.2,
253
+ "restore_max_error": 4.0978193283081055e-07
254
+ },
255
+ "PushCube-v1": {
256
+ "action_mse_to_best": 0.4251232282212465,
257
+ "expert_success_rate": 0.8514851485148515,
258
+ "num_groups": 101,
259
+ "oracle_success_rate": 1.0,
260
+ "policy_expert_regret": 0.40312459238685006,
261
+ "policy_oracle_regret": 0.45788167033455157,
262
+ "policy_rollout_progress": 0.779742092041686,
263
+ "policy_rollout_success_rate": 0.7623762376237624,
264
+ "restore_max_error": 4.76837158203125e-07
265
+ },
266
+ "StackCube-v1": {
267
+ "action_mse_to_best": 0.8084622211961283,
268
+ "expert_success_rate": 0.7666666666666667,
269
+ "num_groups": 90,
270
+ "oracle_success_rate": 0.9111111111111111,
271
+ "policy_expert_regret": 1.1430777420600255,
272
+ "policy_oracle_regret": 1.2921499071849718,
273
+ "policy_rollout_progress": 0.42319835788673826,
274
+ "policy_rollout_success_rate": 0.16666666666666666,
275
+ "restore_max_error": 4.76837158203125e-07
276
+ }
277
+ }
278
+ }
279
+ ]
280
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k32_sigma0p35_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_hybrid_k32_sigma0p35.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: 31.30% +/- 1.38%
11
+ Gain vs h=16 rank checkpoint: +1.57%
12
+ Mean progress: 54.20%
13
+ Mean action MSE to best: 0.554
14
+
15
+ | seed | mode | k | retrieval K | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 47 | 1 | 1.00 | 0.35 | 0 | 0.00 | 30.26% | 52.67% | 85.74% | 0.627 |
18
+ | 1 | retrieval_residual | 47 | 1 | 1.00 | 0.35 | 0 | 0.00 | 30.78% | 54.26% | 86.96% | 0.502 |
19
+ | 2 | retrieval_residual | 47 | 1 | 1.00 | 0.35 | 0 | 0.00 | 32.87% | 55.68% | 87.65% | 0.535 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k64_sigma0p50_summary.json ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_hybrid_k64_sigma0p50.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3089855072463768,
9
+ "std_success": 0.011579701075616322,
10
+ "mean_progress": 0.5389061679982174,
11
+ "mean_action_mse_to_best": 0.5617759731394387,
12
+ "gain_vs_h4": 0.012285507246376781,
13
+ "gain_vs_h16_rank_checkpoint": 0.011594202898550732,
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_hybrid_k64_sigma0p50.json",
18
+ "num_groups": 575,
19
+ "selection_mode": "retrieval_residual",
20
+ "num_candidates": 79,
21
+ "candidate_sigma": 0.5,
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_residual_scale": 1.0,
28
+ "policy_rollout_success_rate": 0.2991304347826087,
29
+ "policy_rollout_progress": 0.5222346339698719,
30
+ "oracle_success_rate": 0.8573913043478261,
31
+ "action_mse_to_best": 0.6354098994126949,
32
+ "best_policy_val": {
33
+ "bc_loss": 0.13721593966086706,
34
+ "field_effect_loss": 0.009290305380192068,
35
+ "field_potential_loss": 0.2666468388504452,
36
+ "field_preference_loss": 0.5130573478009965,
37
+ "lattice_edges": 3833.3333333333335,
38
+ "progress_mae": 0.1933159919248687,
39
+ "rank_acc": 0.8265031774838766,
40
+ "rank_loss": 0.5130523675017886,
41
+ "regret_mae": 0.3756548762321472,
42
+ "success_accuracy": 0.8773836526605818,
43
+ "total_loss": 1.5581054819954767
44
+ },
45
+ "per_task": {
46
+ "LiftPegUpright-v1": {
47
+ "action_mse_to_best": 1.11691102218939,
48
+ "expert_success_rate": 0.8865979381443299,
49
+ "num_groups": 97,
50
+ "oracle_success_rate": 0.9278350515463918,
51
+ "policy_expert_regret": 1.2635736723536068,
52
+ "policy_oracle_regret": 1.3031881279552107,
53
+ "policy_rollout_progress": 0.47611702686732577,
54
+ "policy_rollout_success_rate": 0.12371134020618557,
55
+ "restore_max_error": 4.76837158203125e-07
56
+ },
57
+ "PickCube-v1": {
58
+ "action_mse_to_best": 0.3567841154144844,
59
+ "expert_success_rate": 0.9375,
60
+ "num_groups": 208,
61
+ "oracle_success_rate": 0.9471153846153846,
62
+ "policy_expert_regret": 0.9926363931825528,
63
+ "policy_oracle_regret": 1.0060592606090581,
64
+ "policy_rollout_progress": 0.5990533275673022,
65
+ "policy_rollout_success_rate": 0.3269230769230769,
66
+ "restore_max_error": 4.76837158203125e-07
67
+ },
68
+ "PullCube-v1": {
69
+ "action_mse_to_best": 0.685494723451602,
70
+ "expert_success_rate": 0.19480519480519481,
71
+ "num_groups": 77,
72
+ "oracle_success_rate": 0.36363636363636365,
73
+ "policy_expert_regret": 0.2233808282278962,
74
+ "policy_oracle_regret": 0.41934984407157866,
75
+ "policy_rollout_progress": 0.3121227929005762,
76
+ "policy_rollout_success_rate": 0.2077922077922078,
77
+ "restore_max_error": 4.76837158203125e-07
78
+ },
79
+ "PushCube-v1": {
80
+ "action_mse_to_best": 0.40427608692899664,
81
+ "expert_success_rate": 0.8617021276595744,
82
+ "num_groups": 94,
83
+ "oracle_success_rate": 0.9893617021276596,
84
+ "policy_expert_regret": 0.5944945258980102,
85
+ "policy_oracle_regret": 0.6927485898771184,
86
+ "policy_rollout_progress": 0.668953537782456,
87
+ "policy_rollout_success_rate": 0.6382978723404256,
88
+ "restore_max_error": 4.76837158203125e-07
89
+ },
90
+ "StackCube-v1": {
91
+ "action_mse_to_best": 0.9295371830971404,
92
+ "expert_success_rate": 0.6767676767676768,
93
+ "num_groups": 99,
94
+ "oracle_success_rate": 0.8585858585858586,
95
+ "policy_expert_regret": 1.0469527790943782,
96
+ "policy_oracle_regret": 1.2291584016397745,
97
+ "policy_rollout_progress": 0.4301351837136529,
98
+ "policy_rollout_success_rate": 0.16161616161616163,
99
+ "restore_max_error": 3.948807716369629e-07
100
+ }
101
+ }
102
+ },
103
+ {
104
+ "seed": 1,
105
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_hybrid_k64_sigma0p50.json",
106
+ "num_groups": 575,
107
+ "selection_mode": "retrieval_residual",
108
+ "num_candidates": 79,
109
+ "candidate_sigma": 0.5,
110
+ "field_optim_steps": 0,
111
+ "field_optim_step_size": 0.0,
112
+ "field_optim_trust_radius": 0.0,
113
+ "field_optim_l2_penalty": 0.0,
114
+ "retrieval_neighbors": 1,
115
+ "retrieval_residual_scale": 1.0,
116
+ "policy_rollout_success_rate": 0.3060869565217391,
117
+ "policy_rollout_progress": 0.5427700769899937,
118
+ "oracle_success_rate": 0.8695652173913043,
119
+ "action_mse_to_best": 0.5044497016493393,
120
+ "best_policy_val": {
121
+ "bc_loss": 0.1315989200439718,
122
+ "field_effect_loss": 0.010110370814800262,
123
+ "field_potential_loss": 0.31546536915832096,
124
+ "field_preference_loss": 0.4960637440284093,
125
+ "lattice_edges": 3833.3333333333335,
126
+ "progress_mae": 0.19023226698239645,
127
+ "rank_acc": 0.8255469501018524,
128
+ "rank_loss": 0.4960531195004781,
129
+ "regret_mae": 0.42156532737943864,
130
+ "success_accuracy": 0.8807473679383596,
131
+ "total_loss": 1.561984618504842
132
+ },
133
+ "per_task": {
134
+ "LiftPegUpright-v1": {
135
+ "action_mse_to_best": 0.44594824277029366,
136
+ "expert_success_rate": 0.8584070796460177,
137
+ "num_groups": 113,
138
+ "oracle_success_rate": 0.9380530973451328,
139
+ "policy_expert_regret": 1.05409726619193,
140
+ "policy_oracle_regret": 1.144553686533354,
141
+ "policy_rollout_progress": 0.6018058268633564,
142
+ "policy_rollout_success_rate": 0.18584070796460178,
143
+ "restore_max_error": 4.76837158203125e-07
144
+ },
145
+ "PickCube-v1": {
146
+ "action_mse_to_best": 0.4363012742251158,
147
+ "expert_success_rate": 0.9402173913043478,
148
+ "num_groups": 184,
149
+ "oracle_success_rate": 0.9456521739130435,
150
+ "policy_expert_regret": 1.1475759681855815,
151
+ "policy_oracle_regret": 1.1475999726605115,
152
+ "policy_rollout_progress": 0.544344583731246,
153
+ "policy_rollout_success_rate": 0.2391304347826087,
154
+ "restore_max_error": 4.76837158203125e-07
155
+ },
156
+ "PullCube-v1": {
157
+ "action_mse_to_best": 0.690646292543725,
158
+ "expert_success_rate": 0.25,
159
+ "num_groups": 76,
160
+ "oracle_success_rate": 0.40789473684210525,
161
+ "policy_expert_regret": 0.2980639918435591,
162
+ "policy_oracle_regret": 0.5446286118219588,
163
+ "policy_rollout_progress": 0.2806023089633931,
164
+ "policy_rollout_success_rate": 0.17105263157894737,
165
+ "restore_max_error": 4.76837158203125e-07
166
+ },
167
+ "PushCube-v1": {
168
+ "action_mse_to_best": 0.4481201197717104,
169
+ "expert_success_rate": 0.8198198198198198,
170
+ "num_groups": 111,
171
+ "oracle_success_rate": 1.0,
172
+ "policy_expert_regret": 0.4013105736927943,
173
+ "policy_oracle_regret": 0.46751821323021037,
174
+ "policy_rollout_progress": 0.7757250300130328,
175
+ "policy_rollout_success_rate": 0.7567567567567568,
176
+ "restore_max_error": 4.76837158203125e-07
177
+ },
178
+ "StackCube-v1": {
179
+ "action_mse_to_best": 0.6280938574716777,
180
+ "expert_success_rate": 0.6923076923076923,
181
+ "num_groups": 91,
182
+ "oracle_success_rate": 0.8571428571428571,
183
+ "policy_expert_regret": 1.1173797860577865,
184
+ "policy_oracle_regret": 1.2895275401872592,
185
+ "policy_rollout_progress": 0.40107778697223456,
186
+ "policy_rollout_success_rate": 0.15384615384615385,
187
+ "restore_max_error": 3.948807716369629e-07
188
+ }
189
+ }
190
+ },
191
+ {
192
+ "seed": 2,
193
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_hybrid_k64_sigma0p50.json",
194
+ "num_groups": 575,
195
+ "selection_mode": "retrieval_residual",
196
+ "num_candidates": 79,
197
+ "candidate_sigma": 0.5,
198
+ "field_optim_steps": 0,
199
+ "field_optim_step_size": 0.0,
200
+ "field_optim_trust_radius": 0.0,
201
+ "field_optim_l2_penalty": 0.0,
202
+ "retrieval_neighbors": 1,
203
+ "retrieval_residual_scale": 1.0,
204
+ "policy_rollout_success_rate": 0.3217391304347826,
205
+ "policy_rollout_progress": 0.5517137930347867,
206
+ "oracle_success_rate": 0.8765217391304347,
207
+ "action_mse_to_best": 0.5454683183562821,
208
+ "best_policy_val": {
209
+ "bc_loss": 0.11367896075050037,
210
+ "field_effect_loss": 0.009670218582161598,
211
+ "field_potential_loss": 0.2641640139950646,
212
+ "field_preference_loss": 0.5130490180518892,
213
+ "lattice_edges": 3833.3333333333335,
214
+ "progress_mae": 0.2021729110015763,
215
+ "rank_acc": 0.8333857821093665,
216
+ "rank_loss": 0.5130119257503085,
217
+ "regret_mae": 0.3958987047274907,
218
+ "success_accuracy": 0.8680730561415354,
219
+ "total_loss": 1.4394984311527677
220
+ },
221
+ "per_task": {
222
+ "LiftPegUpright-v1": {
223
+ "action_mse_to_best": 0.42028625178015017,
224
+ "expert_success_rate": 0.8229166666666666,
225
+ "num_groups": 96,
226
+ "oracle_success_rate": 0.9270833333333334,
227
+ "policy_expert_regret": 0.902202141781648,
228
+ "policy_oracle_regret": 1.0084768570959568,
229
+ "policy_rollout_progress": 0.6230483328302702,
230
+ "policy_rollout_success_rate": 0.3020833333333333,
231
+ "restore_max_error": 3.5762786865234375e-07
232
+ },
233
+ "PickCube-v1": {
234
+ "action_mse_to_best": 0.4327403934804177,
235
+ "expert_success_rate": 0.9444444444444444,
236
+ "num_groups": 198,
237
+ "oracle_success_rate": 0.9595959595959596,
238
+ "policy_expert_regret": 1.090324743152267,
239
+ "policy_oracle_regret": 1.1116519589917828,
240
+ "policy_rollout_progress": 0.5906219470380533,
241
+ "policy_rollout_success_rate": 0.25252525252525254,
242
+ "restore_max_error": 4.76837158203125e-07
243
+ },
244
+ "PullCube-v1": {
245
+ "action_mse_to_best": 0.7821945041418076,
246
+ "expert_success_rate": 0.24444444444444444,
247
+ "num_groups": 90,
248
+ "oracle_success_rate": 0.4666666666666667,
249
+ "policy_expert_regret": 0.39066616714828545,
250
+ "policy_oracle_regret": 0.658250150974426,
251
+ "policy_rollout_progress": 0.2611693287268281,
252
+ "policy_rollout_success_rate": 0.15555555555555556,
253
+ "restore_max_error": 4.0978193283081055e-07
254
+ },
255
+ "PushCube-v1": {
256
+ "action_mse_to_best": 0.42668968136652863,
257
+ "expert_success_rate": 0.8514851485148515,
258
+ "num_groups": 101,
259
+ "oracle_success_rate": 1.0,
260
+ "policy_expert_regret": 0.40216707327578327,
261
+ "policy_oracle_regret": 0.45692415122348484,
262
+ "policy_rollout_progress": 0.7806996111527528,
263
+ "policy_rollout_success_rate": 0.7623762376237624,
264
+ "restore_max_error": 4.76837158203125e-07
265
+ },
266
+ "StackCube-v1": {
267
+ "action_mse_to_best": 0.8235671309340331,
268
+ "expert_success_rate": 0.7666666666666667,
269
+ "num_groups": 90,
270
+ "oracle_success_rate": 0.9111111111111111,
271
+ "policy_expert_regret": 1.142678930858771,
272
+ "policy_oracle_regret": 1.291751095983717,
273
+ "policy_rollout_progress": 0.4235971690879928,
274
+ "policy_rollout_success_rate": 0.16666666666666666,
275
+ "restore_max_error": 4.76837158203125e-07
276
+ }
277
+ }
278
+ }
279
+ ]
280
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k64_sigma0p50_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_hybrid_k64_sigma0p50.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: 30.90% +/- 1.16%
11
+ Gain vs h=16 rank checkpoint: +1.16%
12
+ Mean progress: 53.89%
13
+ Mean action MSE to best: 0.562
14
+
15
+ | seed | mode | k | retrieval K | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 79 | 1 | 1.00 | 0.50 | 0 | 0.00 | 29.91% | 52.22% | 85.74% | 0.635 |
18
+ | 1 | retrieval_residual | 79 | 1 | 1.00 | 0.50 | 0 | 0.00 | 30.61% | 54.28% | 86.96% | 0.504 |
19
+ | 2 | retrieval_residual | 79 | 1 | 1.00 | 0.50 | 0 | 0.00 | 32.17% | 55.17% | 87.65% | 0.545 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_summary.json ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_scale0p25.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3292753623188406,
9
+ "std_success": 0.015194599909659314,
10
+ "mean_progress": 0.5523768832874687,
11
+ "mean_action_mse_to_best": 0.40910649155253087,
12
+ "gain_vs_h4": 0.03257536231884056,
13
+ "gain_vs_h16_rank_checkpoint": 0.03188405797101451,
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_scale0p25.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_residual_scale": 0.25,
28
+ "policy_rollout_success_rate": 0.32521739130434785,
29
+ "policy_rollout_progress": 0.5427134724674018,
30
+ "oracle_success_rate": 0.8573913043478261,
31
+ "action_mse_to_best": 0.39587786793789786,
32
+ "best_policy_val": {
33
+ "bc_loss": 0.13721593966086706,
34
+ "field_effect_loss": 0.009290305380192068,
35
+ "field_potential_loss": 0.2666468388504452,
36
+ "field_preference_loss": 0.5130573478009965,
37
+ "lattice_edges": 3833.3333333333335,
38
+ "progress_mae": 0.1933159919248687,
39
+ "rank_acc": 0.8265031774838766,
40
+ "rank_loss": 0.5130523675017886,
41
+ "regret_mae": 0.3756548762321472,
42
+ "success_accuracy": 0.8773836526605818,
43
+ "total_loss": 1.5581054819954767
44
+ },
45
+ "per_task": {
46
+ "LiftPegUpright-v1": {
47
+ "action_mse_to_best": 0.32431400695470036,
48
+ "expert_success_rate": 0.8865979381443299,
49
+ "num_groups": 97,
50
+ "oracle_success_rate": 0.9278350515463918,
51
+ "policy_expert_regret": 1.0923672899142982,
52
+ "policy_oracle_regret": 1.1187827676227413,
53
+ "policy_rollout_progress": 0.567738882045156,
54
+ "policy_rollout_success_rate": 0.21649484536082475,
55
+ "restore_max_error": 4.76837158203125e-07
56
+ },
57
+ "PickCube-v1": {
58
+ "action_mse_to_best": 0.30881268131689954,
59
+ "expert_success_rate": 0.9375,
60
+ "num_groups": 208,
61
+ "oracle_success_rate": 0.9471153846153846,
62
+ "policy_expert_regret": 1.0182457184150384,
63
+ "policy_oracle_regret": 1.031668585841544,
64
+ "policy_rollout_progress": 0.5859263594745874,
65
+ "policy_rollout_success_rate": 0.3076923076923077,
66
+ "restore_max_error": 4.76837158203125e-07
67
+ },
68
+ "PullCube-v1": {
69
+ "action_mse_to_best": 0.623021765382259,
70
+ "expert_success_rate": 0.19480519480519481,
71
+ "num_groups": 77,
72
+ "oracle_success_rate": 0.36363636363636365,
73
+ "policy_expert_regret": 0.21665887307230527,
74
+ "policy_oracle_regret": 0.38748751849226365,
75
+ "policy_rollout_progress": 0.35096192981612373,
76
+ "policy_rollout_success_rate": 0.24675324675324675,
77
+ "restore_max_error": 4.76837158203125e-07
78
+ },
79
+ "PushCube-v1": {
80
+ "action_mse_to_best": 0.3740823645778793,
81
+ "expert_success_rate": 0.8617021276595744,
82
+ "num_groups": 94,
83
+ "oracle_success_rate": 0.9893617021276596,
84
+ "policy_expert_regret": 0.5533106022692741,
85
+ "policy_oracle_regret": 0.6498078486703812,
86
+ "policy_rollout_progress": 0.6906176832445124,
87
+ "policy_rollout_success_rate": 0.6595744680851063,
88
+ "restore_max_error": 4.76837158203125e-07
89
+ },
90
+ "StackCube-v1": {
91
+ "action_mse_to_best": 0.49294807546480435,
92
+ "expert_success_rate": 0.6767676767676768,
93
+ "num_groups": 99,
94
+ "oracle_success_rate": 0.8585858585858586,
95
+ "policy_expert_regret": 0.9980843377233756,
96
+ "policy_oracle_regret": 1.1726798326379122,
97
+ "policy_rollout_progress": 0.4361087022104649,
98
+ "policy_rollout_success_rate": 0.21212121212121213,
99
+ "restore_max_error": 3.948807716369629e-07
100
+ }
101
+ }
102
+ },
103
+ {
104
+ "seed": 1,
105
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_scale0p25.json",
106
+ "num_groups": 575,
107
+ "selection_mode": "retrieval_residual",
108
+ "num_candidates": 16,
109
+ "candidate_sigma": 0.0,
110
+ "field_optim_steps": 0,
111
+ "field_optim_step_size": 0.0,
112
+ "field_optim_trust_radius": 0.0,
113
+ "field_optim_l2_penalty": 0.0,
114
+ "retrieval_neighbors": 1,
115
+ "retrieval_residual_scale": 0.25,
116
+ "policy_rollout_success_rate": 0.3165217391304348,
117
+ "policy_rollout_progress": 0.5491900552825435,
118
+ "oracle_success_rate": 0.8695652173913043,
119
+ "action_mse_to_best": 0.4010985842379539,
120
+ "best_policy_val": {
121
+ "bc_loss": 0.1315989200439718,
122
+ "field_effect_loss": 0.010110370814800262,
123
+ "field_potential_loss": 0.31546536915832096,
124
+ "field_preference_loss": 0.4960637440284093,
125
+ "lattice_edges": 3833.3333333333335,
126
+ "progress_mae": 0.19023226698239645,
127
+ "rank_acc": 0.8255469501018524,
128
+ "rank_loss": 0.4960531195004781,
129
+ "regret_mae": 0.42156532737943864,
130
+ "success_accuracy": 0.8807473679383596,
131
+ "total_loss": 1.561984618504842
132
+ },
133
+ "per_task": {
134
+ "LiftPegUpright-v1": {
135
+ "action_mse_to_best": 0.3345377698656073,
136
+ "expert_success_rate": 0.8584070796460177,
137
+ "num_groups": 113,
138
+ "oracle_success_rate": 0.9380530973451328,
139
+ "policy_expert_regret": 1.006345159316485,
140
+ "policy_oracle_regret": 1.0856339500277443,
141
+ "policy_rollout_progress": 0.6228861411852119,
142
+ "policy_rollout_success_rate": 0.21238938053097345,
143
+ "restore_max_error": 4.76837158203125e-07
144
+ },
145
+ "PickCube-v1": {
146
+ "action_mse_to_best": 0.33102217217421404,
147
+ "expert_success_rate": 0.9402173913043478,
148
+ "num_groups": 184,
149
+ "oracle_success_rate": 0.9456521739130435,
150
+ "policy_expert_regret": 1.1590834624771758,
151
+ "policy_oracle_regret": 1.159107466952106,
152
+ "policy_rollout_progress": 0.538196094460664,
153
+ "policy_rollout_success_rate": 0.23369565217391305,
154
+ "restore_max_error": 4.76837158203125e-07
155
+ },
156
+ "PullCube-v1": {
157
+ "action_mse_to_best": 0.5778321055205244,
158
+ "expert_success_rate": 0.25,
159
+ "num_groups": 76,
160
+ "oracle_success_rate": 0.40789473684210525,
161
+ "policy_expert_regret": 0.33867529622818293,
162
+ "policy_oracle_regret": 0.5427592811909946,
163
+ "policy_rollout_progress": 0.2870815945789218,
164
+ "policy_rollout_success_rate": 0.17105263157894737,
165
+ "restore_max_error": 4.76837158203125e-07
166
+ },
167
+ "PushCube-v1": {
168
+ "action_mse_to_best": 0.37684367295052557,
169
+ "expert_success_rate": 0.8198198198198198,
170
+ "num_groups": 111,
171
+ "oracle_success_rate": 1.0,
172
+ "policy_expert_regret": 0.3626773151996973,
173
+ "policy_oracle_regret": 0.4123242026245272,
174
+ "policy_rollout_progress": 0.803892013591689,
175
+ "policy_rollout_success_rate": 0.7837837837837838,
176
+ "restore_max_error": 4.76837158203125e-07
177
+ },
178
+ "StackCube-v1": {
179
+ "action_mse_to_best": 0.5074280279657835,
180
+ "expert_success_rate": 0.6923076923076923,
181
+ "num_groups": 91,
182
+ "oracle_success_rate": 0.8571428571428571,
183
+ "policy_expert_regret": 1.1193383229630334,
184
+ "policy_oracle_regret": 1.291486077092506,
185
+ "policy_rollout_progress": 0.3881302390779768,
186
+ "policy_rollout_success_rate": 0.16483516483516483,
187
+ "restore_max_error": 3.948807716369629e-07
188
+ }
189
+ }
190
+ },
191
+ {
192
+ "seed": 2,
193
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_scale0p25.json",
194
+ "num_groups": 575,
195
+ "selection_mode": "retrieval_residual",
196
+ "num_candidates": 16,
197
+ "candidate_sigma": 0.0,
198
+ "field_optim_steps": 0,
199
+ "field_optim_step_size": 0.0,
200
+ "field_optim_trust_radius": 0.0,
201
+ "field_optim_l2_penalty": 0.0,
202
+ "retrieval_neighbors": 1,
203
+ "retrieval_residual_scale": 0.25,
204
+ "policy_rollout_success_rate": 0.34608695652173915,
205
+ "policy_rollout_progress": 0.5652271221124607,
206
+ "oracle_success_rate": 0.8765217391304347,
207
+ "action_mse_to_best": 0.43034302248174083,
208
+ "best_policy_val": {
209
+ "bc_loss": 0.11367896075050037,
210
+ "field_effect_loss": 0.009670218582161598,
211
+ "field_potential_loss": 0.2641640139950646,
212
+ "field_preference_loss": 0.5130490180518892,
213
+ "lattice_edges": 3833.3333333333335,
214
+ "progress_mae": 0.2021729110015763,
215
+ "rank_acc": 0.8333857821093665,
216
+ "rank_loss": 0.5130119257503085,
217
+ "regret_mae": 0.3958987047274907,
218
+ "success_accuracy": 0.8680730561415354,
219
+ "total_loss": 1.4394984311527677
220
+ },
221
+ "per_task": {
222
+ "LiftPegUpright-v1": {
223
+ "action_mse_to_best": 0.3580305190165139,
224
+ "expert_success_rate": 0.8229166666666666,
225
+ "num_groups": 96,
226
+ "oracle_success_rate": 0.9270833333333334,
227
+ "policy_expert_regret": 0.8474036318560442,
228
+ "policy_oracle_regret": 0.9536635428667068,
229
+ "policy_rollout_progress": 0.6466116470595201,
230
+ "policy_rollout_success_rate": 0.3333333333333333,
231
+ "restore_max_error": 3.5762786865234375e-07
232
+ },
233
+ "PickCube-v1": {
234
+ "action_mse_to_best": 0.34562643173367086,
235
+ "expert_success_rate": 0.9444444444444444,
236
+ "num_groups": 198,
237
+ "oracle_success_rate": 0.9595959595959596,
238
+ "policy_expert_regret": 1.0567232194083809,
239
+ "policy_oracle_regret": 1.0638066699370892,
240
+ "policy_rollout_progress": 0.5980631956887065,
241
+ "policy_rollout_success_rate": 0.29292929292929293,
242
+ "restore_max_error": 4.76837158203125e-07
243
+ },
244
+ "PullCube-v1": {
245
+ "action_mse_to_best": 0.6627736421922843,
246
+ "expert_success_rate": 0.24444444444444444,
247
+ "num_groups": 90,
248
+ "oracle_success_rate": 0.4666666666666667,
249
+ "policy_expert_regret": 0.2764986334575547,
250
+ "policy_oracle_regret": 0.5040211054599948,
251
+ "policy_rollout_progress": 0.3299036862535609,
252
+ "policy_rollout_success_rate": 0.2222222222222222,
253
+ "restore_max_error": 4.0978193283081055e-07
254
+ },
255
+ "PushCube-v1": {
256
+ "action_mse_to_best": 0.3916822287558329,
257
+ "expert_success_rate": 0.8514851485148515,
258
+ "num_groups": 101,
259
+ "oracle_success_rate": 1.0,
260
+ "policy_expert_regret": 0.4015513146572774,
261
+ "policy_oracle_regret": 0.4742009368273291,
262
+ "policy_rollout_progress": 0.7733238156479184,
263
+ "policy_rollout_success_rate": 0.7524752475247525,
264
+ "restore_max_error": 4.76837158203125e-07
265
+ },
266
+ "StackCube-v1": {
267
+ "action_mse_to_best": 0.50480824129449,
268
+ "expert_success_rate": 0.7666666666666667,
269
+ "num_groups": 90,
270
+ "oracle_success_rate": 0.9111111111111111,
271
+ "policy_expert_regret": 1.1974563411540455,
272
+ "policy_oracle_regret": 1.3337924069828457,
273
+ "policy_rollout_progress": 0.4079703023036321,
274
+ "policy_rollout_success_rate": 0.14444444444444443,
275
+ "restore_max_error": 4.76837158203125e-07
276
+ }
277
+ }
278
+ }
279
+ ]
280
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_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_scale0p25.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: 32.93% +/- 1.52%
11
+ Gain vs h=16 rank checkpoint: +3.19%
12
+ Mean progress: 55.24%
13
+ Mean action MSE to best: 0.409
14
+
15
+ | seed | mode | k | retrieval K | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 16 | 1 | 0.25 | 0.00 | 0 | 0.00 | 32.52% | 54.27% | 85.74% | 0.396 |
18
+ | 1 | retrieval_residual | 16 | 1 | 0.25 | 0.00 | 0 | 0.00 | 31.65% | 54.92% | 86.96% | 0.401 |
19
+ | 2 | retrieval_residual | 16 | 1 | 0.25 | 0.00 | 0 | 0.00 | 34.61% | 56.52% | 87.65% | 0.430 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_summary.json ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_scale0p50.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": 0.00821880978478714,
10
+ "mean_progress": 0.5528276873363749,
11
+ "mean_action_mse_to_best": 0.4331563813282528,
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_scale0p50.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_residual_scale": 0.5,
28
+ "policy_rollout_success_rate": 0.33043478260869563,
29
+ "policy_rollout_progress": 0.544471482436942,
30
+ "oracle_success_rate": 0.8573913043478261,
31
+ "action_mse_to_best": 0.4129274774707206,
32
+ "best_policy_val": {
33
+ "bc_loss": 0.13721593966086706,
34
+ "field_effect_loss": 0.009290305380192068,
35
+ "field_potential_loss": 0.2666468388504452,
36
+ "field_preference_loss": 0.5130573478009965,
37
+ "lattice_edges": 3833.3333333333335,
38
+ "progress_mae": 0.1933159919248687,
39
+ "rank_acc": 0.8265031774838766,
40
+ "rank_loss": 0.5130523675017886,
41
+ "regret_mae": 0.3756548762321472,
42
+ "success_accuracy": 0.8773836526605818,
43
+ "total_loss": 1.5581054819954767
44
+ },
45
+ "per_task": {
46
+ "LiftPegUpright-v1": {
47
+ "action_mse_to_best": 0.33314846369159434,
48
+ "expert_success_rate": 0.8865979381443299,
49
+ "num_groups": 97,
50
+ "oracle_success_rate": 0.9278350515463918,
51
+ "policy_expert_regret": 1.0792845781009222,
52
+ "policy_oracle_regret": 1.105700055809365,
53
+ "policy_rollout_progress": 0.5705123155080166,
54
+ "policy_rollout_success_rate": 0.2268041237113402,
55
+ "restore_max_error": 4.76837158203125e-07
56
+ },
57
+ "PickCube-v1": {
58
+ "action_mse_to_best": 0.3205748214744605,
59
+ "expert_success_rate": 0.9375,
60
+ "num_groups": 208,
61
+ "oracle_success_rate": 0.9471153846153846,
62
+ "policy_expert_regret": 1.0125951060147669,
63
+ "policy_oracle_regret": 1.0260179734412724,
64
+ "policy_rollout_progress": 0.5887099993504727,
65
+ "policy_rollout_success_rate": 0.3173076923076923,
66
+ "restore_max_error": 4.76837158203125e-07
67
+ },
68
+ "PullCube-v1": {
69
+ "action_mse_to_best": 0.6671761597518797,
70
+ "expert_success_rate": 0.19480519480519481,
71
+ "num_groups": 77,
72
+ "oracle_success_rate": 0.36363636363636365,
73
+ "policy_expert_regret": 0.24359342272018458,
74
+ "policy_oracle_regret": 0.4621977220204743,
75
+ "policy_rollout_progress": 0.3145493839371514,
76
+ "policy_rollout_success_rate": 0.2077922077922078,
77
+ "restore_max_error": 4.76837158203125e-07
78
+ },
79
+ "PushCube-v1": {
80
+ "action_mse_to_best": 0.3714648786773707,
81
+ "expert_success_rate": 0.8617021276595744,
82
+ "num_groups": 94,
83
+ "oracle_success_rate": 0.9893617021276596,
84
+ "policy_expert_regret": 0.5719942746961371,
85
+ "policy_oracle_regret": 0.6495450359392674,
86
+ "policy_rollout_progress": 0.6908804959756263,
87
+ "policy_rollout_success_rate": 0.6595744680851063,
88
+ "restore_max_error": 4.76837158203125e-07
89
+ },
90
+ "StackCube-v1": {
91
+ "action_mse_to_best": 0.5267482101446902,
92
+ "expert_success_rate": 0.6767676767676768,
93
+ "num_groups": 99,
94
+ "oracle_success_rate": 0.8585858585858586,
95
+ "policy_expert_regret": 0.9495171221217724,
96
+ "policy_oracle_regret": 1.112660668715082,
97
+ "policy_rollout_progress": 0.4658248358302646,
98
+ "policy_rollout_success_rate": 0.24242424242424243,
99
+ "restore_max_error": 3.948807716369629e-07
100
+ }
101
+ }
102
+ },
103
+ {
104
+ "seed": 1,
105
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_scale0p50.json",
106
+ "num_groups": 575,
107
+ "selection_mode": "retrieval_residual",
108
+ "num_candidates": 16,
109
+ "candidate_sigma": 0.0,
110
+ "field_optim_steps": 0,
111
+ "field_optim_step_size": 0.0,
112
+ "field_optim_trust_radius": 0.0,
113
+ "field_optim_l2_penalty": 0.0,
114
+ "retrieval_neighbors": 1,
115
+ "retrieval_residual_scale": 0.5,
116
+ "policy_rollout_success_rate": 0.3269565217391304,
117
+ "policy_rollout_progress": 0.5530163816378816,
118
+ "oracle_success_rate": 0.8695652173913043,
119
+ "action_mse_to_best": 0.42286321135642735,
120
+ "best_policy_val": {
121
+ "bc_loss": 0.1315989200439718,
122
+ "field_effect_loss": 0.010110370814800262,
123
+ "field_potential_loss": 0.31546536915832096,
124
+ "field_preference_loss": 0.4960637440284093,
125
+ "lattice_edges": 3833.3333333333335,
126
+ "progress_mae": 0.19023226698239645,
127
+ "rank_acc": 0.8255469501018524,
128
+ "rank_loss": 0.4960531195004781,
129
+ "regret_mae": 0.42156532737943864,
130
+ "success_accuracy": 0.8807473679383596,
131
+ "total_loss": 1.561984618504842
132
+ },
133
+ "per_task": {
134
+ "LiftPegUpright-v1": {
135
+ "action_mse_to_best": 0.33098896499723196,
136
+ "expert_success_rate": 0.8584070796460177,
137
+ "num_groups": 113,
138
+ "oracle_success_rate": 0.9380530973451328,
139
+ "policy_expert_regret": 0.998058971973647,
140
+ "policy_oracle_regret": 1.0891584641901793,
141
+ "policy_rollout_progress": 0.619361627022777,
142
+ "policy_rollout_success_rate": 0.21238938053097345,
143
+ "restore_max_error": 4.76837158203125e-07
144
+ },
145
+ "PickCube-v1": {
146
+ "action_mse_to_best": 0.3595517423692281,
147
+ "expert_success_rate": 0.9402173913043478,
148
+ "num_groups": 184,
149
+ "oracle_success_rate": 0.9456521739130435,
150
+ "policy_expert_regret": 1.1190804282618363,
151
+ "policy_oracle_regret": 1.1266173332291858,
152
+ "policy_rollout_progress": 0.5457061722864518,
153
+ "policy_rollout_success_rate": 0.266304347826087,
154
+ "restore_max_error": 4.76837158203125e-07
155
+ },
156
+ "PullCube-v1": {
157
+ "action_mse_to_best": 0.6149560018981758,
158
+ "expert_success_rate": 0.25,
159
+ "num_groups": 76,
160
+ "oracle_success_rate": 0.40789473684210525,
161
+ "policy_expert_regret": 0.27304727383154004,
162
+ "policy_oracle_regret": 0.5447679004680953,
163
+ "policy_rollout_progress": 0.27486991495090096,
164
+ "policy_rollout_success_rate": 0.15789473684210525,
165
+ "restore_max_error": 4.76837158203125e-07
166
+ },
167
+ "PushCube-v1": {
168
+ "action_mse_to_best": 0.39232181858372045,
169
+ "expert_success_rate": 0.8198198198198198,
170
+ "num_groups": 111,
171
+ "oracle_success_rate": 1.0,
172
+ "policy_expert_regret": 0.34771036054636983,
173
+ "policy_oracle_regret": 0.38119086757436527,
174
+ "policy_rollout_progress": 0.8170073306238329,
175
+ "policy_rollout_success_rate": 0.8018018018018018,
176
+ "restore_max_error": 4.76837158203125e-07
177
+ },
178
+ "StackCube-v1": {
179
+ "action_mse_to_best": 0.5417878558490794,
180
+ "expert_success_rate": 0.6923076923076923,
181
+ "num_groups": 91,
182
+ "oracle_success_rate": 0.8571428571428571,
183
+ "policy_expert_regret": 1.122757653777416,
184
+ "policy_oracle_regret": 1.2949054079068887,
185
+ "policy_rollout_progress": 0.3956999192526052,
186
+ "policy_rollout_success_rate": 0.15384615384615385,
187
+ "restore_max_error": 3.948807716369629e-07
188
+ }
189
+ }
190
+ },
191
+ {
192
+ "seed": 2,
193
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_scale0p50.json",
194
+ "num_groups": 575,
195
+ "selection_mode": "retrieval_residual",
196
+ "num_candidates": 16,
197
+ "candidate_sigma": 0.0,
198
+ "field_optim_steps": 0,
199
+ "field_optim_step_size": 0.0,
200
+ "field_optim_trust_radius": 0.0,
201
+ "field_optim_l2_penalty": 0.0,
202
+ "retrieval_neighbors": 1,
203
+ "retrieval_residual_scale": 0.5,
204
+ "policy_rollout_success_rate": 0.3426086956521739,
205
+ "policy_rollout_progress": 0.560995197934301,
206
+ "oracle_success_rate": 0.8765217391304347,
207
+ "action_mse_to_best": 0.46367845515761036,
208
+ "best_policy_val": {
209
+ "bc_loss": 0.11367896075050037,
210
+ "field_effect_loss": 0.009670218582161598,
211
+ "field_potential_loss": 0.2641640139950646,
212
+ "field_preference_loss": 0.5130490180518892,
213
+ "lattice_edges": 3833.3333333333335,
214
+ "progress_mae": 0.2021729110015763,
215
+ "rank_acc": 0.8333857821093665,
216
+ "rank_loss": 0.5130119257503085,
217
+ "regret_mae": 0.3958987047274907,
218
+ "success_accuracy": 0.8680730561415354,
219
+ "total_loss": 1.4394984311527677
220
+ },
221
+ "per_task": {
222
+ "LiftPegUpright-v1": {
223
+ "action_mse_to_best": 0.36153392906514153,
224
+ "expert_success_rate": 0.8229166666666666,
225
+ "num_groups": 96,
226
+ "oracle_success_rate": 0.9270833333333334,
227
+ "policy_expert_regret": 0.8829973929872116,
228
+ "policy_oracle_regret": 0.9760098507006963,
229
+ "policy_rollout_progress": 0.6346820058921973,
230
+ "policy_rollout_success_rate": 0.3229166666666667,
231
+ "restore_max_error": 3.5762786865234375e-07
232
+ },
233
+ "PickCube-v1": {
234
+ "action_mse_to_best": 0.37652896986239487,
235
+ "expert_success_rate": 0.9444444444444444,
236
+ "num_groups": 198,
237
+ "oracle_success_rate": 0.9595959595959596,
238
+ "policy_expert_regret": 1.0838303450457376,
239
+ "policy_oracle_regret": 1.090913795574446,
240
+ "policy_rollout_progress": 0.5840688492688868,
241
+ "policy_rollout_success_rate": 0.2727272727272727,
242
+ "restore_max_error": 4.76837158203125e-07
243
+ },
244
+ "PullCube-v1": {
245
+ "action_mse_to_best": 0.7301618857516183,
246
+ "expert_success_rate": 0.24444444444444444,
247
+ "num_groups": 90,
248
+ "oracle_success_rate": 0.4666666666666667,
249
+ "policy_expert_regret": 0.30010166329642135,
250
+ "policy_oracle_regret": 0.5302090961693062,
251
+ "policy_rollout_progress": 0.31182427391823797,
252
+ "policy_rollout_success_rate": 0.2111111111111111,
253
+ "restore_max_error": 4.0978193283081055e-07
254
+ },
255
+ "PushCube-v1": {
256
+ "action_mse_to_best": 0.4178345319215614,
257
+ "expert_success_rate": 0.8514851485148515,
258
+ "num_groups": 101,
259
+ "oracle_success_rate": 1.0,
260
+ "policy_expert_regret": 0.38219682946063505,
261
+ "policy_oracle_regret": 0.4192075215943969,
262
+ "policy_rollout_progress": 0.7986142605838209,
263
+ "policy_rollout_success_rate": 0.7821782178217822,
264
+ "restore_max_error": 4.76837158203125e-07
265
+ },
266
+ "StackCube-v1": {
267
+ "action_mse_to_best": 0.5493251227877206,
268
+ "expert_success_rate": 0.7666666666666667,
269
+ "num_groups": 90,
270
+ "oracle_success_rate": 0.9111111111111111,
271
+ "policy_expert_regret": 1.15032865587208,
272
+ "policy_oracle_regret": 1.3123159415192074,
273
+ "policy_rollout_progress": 0.41414343466361364,
274
+ "policy_rollout_success_rate": 0.15555555555555556,
275
+ "restore_max_error": 4.76837158203125e-07
276
+ }
277
+ }
278
+ }
279
+ ]
280
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_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_scale0p50.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% +/- 0.82%
11
+ Gain vs h=16 rank checkpoint: +3.59%
12
+ Mean progress: 55.28%
13
+ Mean action MSE to best: 0.433
14
+
15
+ | seed | mode | k | retrieval K | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 16 | 1 | 0.50 | 0.00 | 0 | 0.00 | 33.04% | 54.45% | 85.74% | 0.413 |
18
+ | 1 | retrieval_residual | 16 | 1 | 0.50 | 0.00 | 0 | 0.00 | 32.70% | 55.30% | 86.96% | 0.423 |
19
+ | 2 | retrieval_residual | 16 | 1 | 0.50 | 0.00 | 0 | 0.00 | 34.26% | 56.10% | 87.65% | 0.464 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p75_summary.json ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_scale0p75.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.3269565217391304,
9
+ "std_success": 0.015161387629706683,
10
+ "mean_progress": 0.5497141635662917,
11
+ "mean_action_mse_to_best": 0.5080524536079147,
12
+ "gain_vs_h4": 0.030256521739130404,
13
+ "gain_vs_h16_rank_checkpoint": 0.029565217391304355,
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_scale0p75.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_residual_scale": 0.75,
28
+ "policy_rollout_success_rate": 0.32,
29
+ "policy_rollout_progress": 0.5345233704787477,
30
+ "oracle_success_rate": 0.8573913043478261,
31
+ "action_mse_to_best": 0.5423873371468938,
32
+ "best_policy_val": {
33
+ "bc_loss": 0.13721593966086706,
34
+ "field_effect_loss": 0.009290305380192068,
35
+ "field_potential_loss": 0.2666468388504452,
36
+ "field_preference_loss": 0.5130573478009965,
37
+ "lattice_edges": 3833.3333333333335,
38
+ "progress_mae": 0.1933159919248687,
39
+ "rank_acc": 0.8265031774838766,
40
+ "rank_loss": 0.5130523675017886,
41
+ "regret_mae": 0.3756548762321472,
42
+ "success_accuracy": 0.8773836526605818,
43
+ "total_loss": 1.5581054819954767
44
+ },
45
+ "per_task": {
46
+ "LiftPegUpright-v1": {
47
+ "action_mse_to_best": 0.805569264125655,
48
+ "expert_success_rate": 0.8865979381443299,
49
+ "num_groups": 97,
50
+ "oracle_success_rate": 0.9278350515463918,
51
+ "policy_expert_regret": 1.2286027728282298,
52
+ "policy_oracle_regret": 1.2683443485461559,
53
+ "policy_rollout_progress": 0.4903422495753495,
54
+ "policy_rollout_success_rate": 0.14432989690721648,
55
+ "restore_max_error": 4.76837158203125e-07
56
+ },
57
+ "PickCube-v1": {
58
+ "action_mse_to_best": 0.35622545783049786,
59
+ "expert_success_rate": 0.9375,
60
+ "num_groups": 208,
61
+ "oracle_success_rate": 0.9471153846153846,
62
+ "policy_expert_regret": 1.0036513171463202,
63
+ "policy_oracle_regret": 1.0170741845728257,
64
+ "policy_rollout_progress": 0.5928460959112272,
65
+ "policy_rollout_success_rate": 0.32211538461538464,
66
+ "restore_max_error": 4.76837158203125e-07
67
+ },
68
+ "PullCube-v1": {
69
+ "action_mse_to_best": 0.6661438940020351,
70
+ "expert_success_rate": 0.19480519480519481,
71
+ "num_groups": 77,
72
+ "oracle_success_rate": 0.36363636363636365,
73
+ "policy_expert_regret": 0.21825521021579428,
74
+ "policy_oracle_regret": 0.38739257521909215,
75
+ "policy_rollout_progress": 0.3385705710494552,
76
+ "policy_rollout_success_rate": 0.23376623376623376,
77
+ "restore_max_error": 4.76837158203125e-07
78
+ },
79
+ "PushCube-v1": {
80
+ "action_mse_to_best": 0.4044005741226546,
81
+ "expert_success_rate": 0.8617021276595744,
82
+ "num_groups": 94,
83
+ "oracle_success_rate": 0.9893617021276596,
84
+ "policy_expert_regret": 0.5533750363169833,
85
+ "policy_oracle_regret": 0.6309257975601136,
86
+ "policy_rollout_progress": 0.6988614364824397,
87
+ "policy_rollout_success_rate": 0.6702127659574468,
88
+ "restore_max_error": 4.76837158203125e-07
89
+ },
90
+ "StackCube-v1": {
91
+ "action_mse_to_best": 0.7104128404529859,
92
+ "expert_success_rate": 0.6767676767676768,
93
+ "num_groups": 99,
94
+ "oracle_success_rate": 0.8585858585858586,
95
+ "policy_expert_regret": 0.9903282389195278,
96
+ "policy_oracle_regret": 1.149254069936396,
97
+ "policy_rollout_progress": 0.45164487744220577,
98
+ "policy_rollout_success_rate": 0.2222222222222222,
99
+ "restore_max_error": 3.948807716369629e-07
100
+ }
101
+ }
102
+ },
103
+ {
104
+ "seed": 1,
105
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_scale0p75.json",
106
+ "num_groups": 575,
107
+ "selection_mode": "retrieval_residual",
108
+ "num_candidates": 16,
109
+ "candidate_sigma": 0.0,
110
+ "field_optim_steps": 0,
111
+ "field_optim_step_size": 0.0,
112
+ "field_optim_trust_radius": 0.0,
113
+ "field_optim_l2_penalty": 0.0,
114
+ "retrieval_neighbors": 1,
115
+ "retrieval_residual_scale": 0.75,
116
+ "policy_rollout_success_rate": 0.3165217391304348,
117
+ "policy_rollout_progress": 0.5510243390516743,
118
+ "oracle_success_rate": 0.8695652173913043,
119
+ "action_mse_to_best": 0.4668293072898751,
120
+ "best_policy_val": {
121
+ "bc_loss": 0.1315989200439718,
122
+ "field_effect_loss": 0.010110370814800262,
123
+ "field_potential_loss": 0.31546536915832096,
124
+ "field_preference_loss": 0.4960637440284093,
125
+ "lattice_edges": 3833.3333333333335,
126
+ "progress_mae": 0.19023226698239645,
127
+ "rank_acc": 0.8255469501018524,
128
+ "rank_loss": 0.4960531195004781,
129
+ "regret_mae": 0.42156532737943864,
130
+ "success_accuracy": 0.8807473679383596,
131
+ "total_loss": 1.561984618504842
132
+ },
133
+ "per_task": {
134
+ "LiftPegUpright-v1": {
135
+ "action_mse_to_best": 0.3412654227371105,
136
+ "expert_success_rate": 0.8584070796460177,
137
+ "num_groups": 113,
138
+ "oracle_success_rate": 0.9380530973451328,
139
+ "policy_expert_regret": 1.0164116371785645,
140
+ "policy_oracle_regret": 1.097544903380681,
141
+ "policy_rollout_progress": 0.6222659374496579,
142
+ "policy_rollout_success_rate": 0.21238938053097345,
143
+ "restore_max_error": 4.76837158203125e-07
144
+ },
145
+ "PickCube-v1": {
146
+ "action_mse_to_best": 0.39500651939812564,
147
+ "expert_success_rate": 0.9402173913043478,
148
+ "num_groups": 184,
149
+ "oracle_success_rate": 0.9456521739130435,
150
+ "policy_expert_regret": 1.1404846262277872,
151
+ "policy_oracle_regret": 1.1405086307027175,
152
+ "policy_rollout_progress": 0.5460016471283187,
153
+ "policy_rollout_success_rate": 0.24456521739130435,
154
+ "restore_max_error": 4.76837158203125e-07
155
+ },
156
+ "PullCube-v1": {
157
+ "action_mse_to_best": 0.6672222055494785,
158
+ "expert_success_rate": 0.25,
159
+ "num_groups": 76,
160
+ "oracle_success_rate": 0.40789473684210525,
161
+ "policy_expert_regret": 0.29614157379163725,
162
+ "policy_oracle_regret": 0.5693492501925089,
163
+ "policy_rollout_progress": 0.2704730750747809,
164
+ "policy_rollout_success_rate": 0.15789473684210525,
165
+ "restore_max_error": 4.76837158203125e-07
166
+ },
167
+ "PushCube-v1": {
168
+ "action_mse_to_best": 0.45045896941745606,
169
+ "expert_success_rate": 0.8198198198198198,
170
+ "num_groups": 111,
171
+ "oracle_success_rate": 1.0,
172
+ "policy_expert_regret": 0.34760262380849133,
173
+ "policy_oracle_regret": 0.4138102633459074,
174
+ "policy_rollout_progress": 0.8024059528703088,
175
+ "policy_rollout_success_rate": 0.7837837837837838,
176
+ "restore_max_error": 4.76837158203125e-07
177
+ },
178
+ "StackCube-v1": {
179
+ "action_mse_to_best": 0.6205805068794679,
180
+ "expert_success_rate": 0.6923076923076923,
181
+ "num_groups": 91,
182
+ "oracle_success_rate": 0.8571428571428571,
183
+ "policy_expert_regret": 1.1180660732827343,
184
+ "policy_oracle_regret": 1.290213827412207,
185
+ "policy_rollout_progress": 0.40039149974728677,
186
+ "policy_rollout_success_rate": 0.15384615384615385,
187
+ "restore_max_error": 3.948807716369629e-07
188
+ }
189
+ }
190
+ },
191
+ {
192
+ "seed": 2,
193
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_scale0p75.json",
194
+ "num_groups": 575,
195
+ "selection_mode": "retrieval_residual",
196
+ "num_candidates": 16,
197
+ "candidate_sigma": 0.0,
198
+ "field_optim_steps": 0,
199
+ "field_optim_step_size": 0.0,
200
+ "field_optim_trust_radius": 0.0,
201
+ "field_optim_l2_penalty": 0.0,
202
+ "retrieval_neighbors": 1,
203
+ "retrieval_residual_scale": 0.75,
204
+ "policy_rollout_success_rate": 0.3443478260869565,
205
+ "policy_rollout_progress": 0.5635947811684531,
206
+ "oracle_success_rate": 0.8765217391304347,
207
+ "action_mse_to_best": 0.5149407163869751,
208
+ "best_policy_val": {
209
+ "bc_loss": 0.11367896075050037,
210
+ "field_effect_loss": 0.009670218582161598,
211
+ "field_potential_loss": 0.2641640139950646,
212
+ "field_preference_loss": 0.5130490180518892,
213
+ "lattice_edges": 3833.3333333333335,
214
+ "progress_mae": 0.2021729110015763,
215
+ "rank_acc": 0.8333857821093665,
216
+ "rank_loss": 0.5130119257503085,
217
+ "regret_mae": 0.3958987047274907,
218
+ "success_accuracy": 0.8680730561415354,
219
+ "total_loss": 1.4394984311527677
220
+ },
221
+ "per_task": {
222
+ "LiftPegUpright-v1": {
223
+ "action_mse_to_best": 0.4603082099104843,
224
+ "expert_success_rate": 0.8229166666666666,
225
+ "num_groups": 96,
226
+ "oracle_success_rate": 0.9270833333333334,
227
+ "policy_expert_regret": 0.924809723161161,
228
+ "policy_oracle_regret": 1.04467046353966,
229
+ "policy_rollout_progress": 0.6076880597199003,
230
+ "policy_rollout_success_rate": 0.28125,
231
+ "restore_max_error": 3.5762786865234375e-07
232
+ },
233
+ "PickCube-v1": {
234
+ "action_mse_to_best": 0.41822907164918655,
235
+ "expert_success_rate": 0.9444444444444444,
236
+ "num_groups": 198,
237
+ "oracle_success_rate": 0.9595959595959596,
238
+ "policy_expert_regret": 1.0595279995145068,
239
+ "policy_oracle_regret": 1.0666114500432153,
240
+ "policy_rollout_progress": 0.6003089206330855,
241
+ "policy_rollout_success_rate": 0.2878787878787879,
242
+ "restore_max_error": 4.76837158203125e-07
243
+ },
244
+ "PullCube-v1": {
245
+ "action_mse_to_best": 0.707032201356358,
246
+ "expert_success_rate": 0.24444444444444444,
247
+ "num_groups": 90,
248
+ "oracle_success_rate": 0.4666666666666667,
249
+ "policy_expert_regret": 0.2938250019835929,
250
+ "policy_oracle_regret": 0.5023684386267431,
251
+ "policy_rollout_progress": 0.34422175456873244,
252
+ "policy_rollout_success_rate": 0.24444444444444444,
253
+ "restore_max_error": 4.0978193283081055e-07
254
+ },
255
+ "PushCube-v1": {
256
+ "action_mse_to_best": 0.42296098563635703,
257
+ "expert_success_rate": 0.8514851485148515,
258
+ "num_groups": 101,
259
+ "oracle_success_rate": 1.0,
260
+ "policy_expert_regret": 0.40117800811139664,
261
+ "policy_oracle_regret": 0.4559350860590982,
262
+ "policy_rollout_progress": 0.7816886763171395,
263
+ "policy_rollout_success_rate": 0.7623762376237624,
264
+ "restore_max_error": 4.76837158203125e-07
265
+ },
266
+ "StackCube-v1": {
267
+ "action_mse_to_best": 0.6971112212580112,
268
+ "expert_success_rate": 0.7666666666666667,
269
+ "num_groups": 90,
270
+ "oracle_success_rate": 0.9111111111111111,
271
+ "policy_expert_regret": 1.159649470117357,
272
+ "policy_oracle_regret": 1.3087216352423032,
273
+ "policy_rollout_progress": 0.4104140548242463,
274
+ "policy_rollout_success_rate": 0.16666666666666666,
275
+ "restore_max_error": 4.76837158203125e-07
276
+ }
277
+ }
278
+ }
279
+ ]
280
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p75_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_scale0p75.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: 32.70% +/- 1.52%
11
+ Gain vs h=16 rank checkpoint: +2.96%
12
+ Mean progress: 54.97%
13
+ Mean action MSE to best: 0.508
14
+
15
+ | seed | mode | k | retrieval K | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 16 | 1 | 0.75 | 0.00 | 0 | 0.00 | 32.00% | 53.45% | 85.74% | 0.542 |
18
+ | 1 | retrieval_residual | 16 | 1 | 0.75 | 0.00 | 0 | 0.00 | 31.65% | 55.10% | 86.96% | 0.467 |
19
+ | 2 | retrieval_residual | 16 | 1 | 0.75 | 0.00 | 0 | 0.00 | 34.43% | 56.36% | 87.65% | 0.515 |
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale1p25_summary.json ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_scale1p25.json",
5
+ "num_completed": 3,
6
+ "baseline_h4_policy_success": 0.2967,
7
+ "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
8
+ "mean_success": 0.32521739130434785,
9
+ "std_success": 0.02465642935436144,
10
+ "mean_progress": 0.5510169452599898,
11
+ "mean_action_mse_to_best": 0.5572565441181803,
12
+ "gain_vs_h4": 0.028517391304347828,
13
+ "gain_vs_h16_rank_checkpoint": 0.02782608695652178,
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_scale1p25.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_residual_scale": 1.25,
28
+ "policy_rollout_success_rate": 0.3060869565217391,
29
+ "policy_rollout_progress": 0.5310814157022041,
30
+ "oracle_success_rate": 0.8573913043478261,
31
+ "action_mse_to_best": 0.6267042727638846,
32
+ "best_policy_val": {
33
+ "bc_loss": 0.13721593966086706,
34
+ "field_effect_loss": 0.009290305380192068,
35
+ "field_potential_loss": 0.2666468388504452,
36
+ "field_preference_loss": 0.5130573478009965,
37
+ "lattice_edges": 3833.3333333333335,
38
+ "progress_mae": 0.1933159919248687,
39
+ "rank_acc": 0.8265031774838766,
40
+ "rank_loss": 0.5130523675017886,
41
+ "regret_mae": 0.3756548762321472,
42
+ "success_accuracy": 0.8773836526605818,
43
+ "total_loss": 1.5581054819954767
44
+ },
45
+ "per_task": {
46
+ "LiftPegUpright-v1": {
47
+ "action_mse_to_best": 0.935821759639327,
48
+ "expert_success_rate": 0.8865979381443299,
49
+ "num_groups": 97,
50
+ "oracle_success_rate": 0.9278350515463918,
51
+ "policy_expert_regret": 1.2285800832755787,
52
+ "policy_oracle_regret": 1.2816902827663519,
53
+ "policy_rollout_progress": 0.4976148720561844,
54
+ "policy_rollout_success_rate": 0.12371134020618557,
55
+ "restore_max_error": 4.76837158203125e-07
56
+ },
57
+ "PickCube-v1": {
58
+ "action_mse_to_best": 0.3258183661561746,
59
+ "expert_success_rate": 0.9375,
60
+ "num_groups": 208,
61
+ "oracle_success_rate": 0.9471153846153846,
62
+ "policy_expert_regret": 0.9906795064811237,
63
+ "policy_oracle_regret": 1.0041023739076291,
64
+ "policy_rollout_progress": 0.6010102142687314,
65
+ "policy_rollout_success_rate": 0.3269230769230769,
66
+ "restore_max_error": 4.76837158203125e-07
67
+ },
68
+ "PullCube-v1": {
69
+ "action_mse_to_best": 0.6729415765443405,
70
+ "expert_success_rate": 0.19480519480519481,
71
+ "num_groups": 77,
72
+ "oracle_success_rate": 0.36363636363636365,
73
+ "policy_expert_regret": 0.24571258389140033,
74
+ "policy_oracle_regret": 0.4187104860993477,
75
+ "policy_rollout_progress": 0.3208115470176865,
76
+ "policy_rollout_success_rate": 0.22077922077922077,
77
+ "restore_max_error": 4.76837158203125e-07
78
+ },
79
+ "PushCube-v1": {
80
+ "action_mse_to_best": 0.4120625256303143,
81
+ "expert_success_rate": 0.8617021276595744,
82
+ "num_groups": 94,
83
+ "oracle_success_rate": 0.9893617021276596,
84
+ "policy_expert_regret": 0.5532618870443486,
85
+ "policy_oracle_regret": 0.630812648287479,
86
+ "policy_rollout_progress": 0.6989745857550743,
87
+ "policy_rollout_success_rate": 0.6702127659574468,
88
+ "restore_max_error": 4.76837158203125e-07
89
+ },
90
+ "StackCube-v1": {
91
+ "action_mse_to_best": 1.123834820106776,
92
+ "expert_success_rate": 0.6767676767676768,
93
+ "num_groups": 99,
94
+ "oracle_success_rate": 0.8585858585858586,
95
+ "policy_expert_regret": 1.0675142982391397,
96
+ "policy_oracle_regret": 1.238213315756634,
97
+ "policy_rollout_progress": 0.42108026959679346,
98
+ "policy_rollout_success_rate": 0.16161616161616163,
99
+ "restore_max_error": 3.948807716369629e-07
100
+ }
101
+ }
102
+ },
103
+ {
104
+ "seed": 1,
105
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_scale1p25.json",
106
+ "num_groups": 575,
107
+ "selection_mode": "retrieval_residual",
108
+ "num_candidates": 16,
109
+ "candidate_sigma": 0.0,
110
+ "field_optim_steps": 0,
111
+ "field_optim_step_size": 0.0,
112
+ "field_optim_trust_radius": 0.0,
113
+ "field_optim_l2_penalty": 0.0,
114
+ "retrieval_neighbors": 1,
115
+ "retrieval_residual_scale": 1.25,
116
+ "policy_rollout_success_rate": 0.3165217391304348,
117
+ "policy_rollout_progress": 0.5496013016751765,
118
+ "oracle_success_rate": 0.8695652173913043,
119
+ "action_mse_to_best": 0.5111655844132538,
120
+ "best_policy_val": {
121
+ "bc_loss": 0.1315989200439718,
122
+ "field_effect_loss": 0.010110370814800262,
123
+ "field_potential_loss": 0.31546536915832096,
124
+ "field_preference_loss": 0.4960637440284093,
125
+ "lattice_edges": 3833.3333333333335,
126
+ "progress_mae": 0.19023226698239645,
127
+ "rank_acc": 0.8255469501018524,
128
+ "rank_loss": 0.4960531195004781,
129
+ "regret_mae": 0.42156532737943864,
130
+ "success_accuracy": 0.8807473679383596,
131
+ "total_loss": 1.561984618504842
132
+ },
133
+ "per_task": {
134
+ "LiftPegUpright-v1": {
135
+ "action_mse_to_best": 0.4187577939386257,
136
+ "expert_success_rate": 0.8584070796460177,
137
+ "num_groups": 113,
138
+ "oracle_success_rate": 0.9380530973451328,
139
+ "policy_expert_regret": 1.0255064777040903,
140
+ "policy_oracle_regret": 1.1159628980455145,
141
+ "policy_rollout_progress": 0.6126975003069481,
142
+ "policy_rollout_success_rate": 0.20353982300884957,
143
+ "restore_max_error": 4.76837158203125e-07
144
+ },
145
+ "PickCube-v1": {
146
+ "action_mse_to_best": 0.44282226890082593,
147
+ "expert_success_rate": 0.9402173913043478,
148
+ "num_groups": 184,
149
+ "oracle_success_rate": 0.9456521739130435,
150
+ "policy_expert_regret": 1.1417508029743382,
151
+ "policy_oracle_regret": 1.1417748074492682,
152
+ "policy_rollout_progress": 0.5425411091874475,
153
+ "policy_rollout_success_rate": 0.2391304347826087,
154
+ "restore_max_error": 4.76837158203125e-07
155
+ },
156
+ "PullCube-v1": {
157
+ "action_mse_to_best": 0.6937027322618585,
158
+ "expert_success_rate": 0.25,
159
+ "num_groups": 76,
160
+ "oracle_success_rate": 0.40789473684210525,
161
+ "policy_expert_regret": 0.275648651243261,
162
+ "policy_oracle_regret": 0.5241066290365876,
163
+ "policy_rollout_progress": 0.2887220772467309,
164
+ "policy_rollout_success_rate": 0.18421052631578946,
165
+ "restore_max_error": 4.76837158203125e-07
166
+ },
167
+ "PushCube-v1": {
168
+ "action_mse_to_best": 0.45558100760922776,
169
+ "expert_success_rate": 0.8198198198198198,
170
+ "num_groups": 111,
171
+ "oracle_success_rate": 1.0,
172
+ "policy_expert_regret": 0.3488818111720386,
173
+ "policy_oracle_regret": 0.41508945070945463,
174
+ "policy_rollout_progress": 0.8011267655067615,
175
+ "policy_rollout_success_rate": 0.7837837837837838,
176
+ "restore_max_error": 4.76837158203125e-07
177
+ },
178
+ "StackCube-v1": {
179
+ "action_mse_to_best": 0.6794547620689967,
180
+ "expert_success_rate": 0.6923076923076923,
181
+ "num_groups": 91,
182
+ "oracle_success_rate": 0.8571428571428571,
183
+ "policy_expert_regret": 1.1218595730734395,
184
+ "policy_oracle_regret": 1.2940073272029122,
185
+ "policy_rollout_progress": 0.3965979999565816,
186
+ "policy_rollout_success_rate": 0.15384615384615385,
187
+ "restore_max_error": 3.948807716369629e-07
188
+ }
189
+ }
190
+ },
191
+ {
192
+ "seed": 2,
193
+ "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_scale1p25.json",
194
+ "num_groups": 575,
195
+ "selection_mode": "retrieval_residual",
196
+ "num_candidates": 16,
197
+ "candidate_sigma": 0.0,
198
+ "field_optim_steps": 0,
199
+ "field_optim_step_size": 0.0,
200
+ "field_optim_trust_radius": 0.0,
201
+ "field_optim_l2_penalty": 0.0,
202
+ "retrieval_neighbors": 1,
203
+ "retrieval_residual_scale": 1.25,
204
+ "policy_rollout_success_rate": 0.35304347826086957,
205
+ "policy_rollout_progress": 0.5723681184025886,
206
+ "oracle_success_rate": 0.8765217391304347,
207
+ "action_mse_to_best": 0.5338997751774024,
208
+ "best_policy_val": {
209
+ "bc_loss": 0.11367896075050037,
210
+ "field_effect_loss": 0.009670218582161598,
211
+ "field_potential_loss": 0.2641640139950646,
212
+ "field_preference_loss": 0.5130490180518892,
213
+ "lattice_edges": 3833.3333333333335,
214
+ "progress_mae": 0.2021729110015763,
215
+ "rank_acc": 0.8333857821093665,
216
+ "rank_loss": 0.5130119257503085,
217
+ "regret_mae": 0.3958987047274907,
218
+ "success_accuracy": 0.8680730561415354,
219
+ "total_loss": 1.4394984311527677
220
+ },
221
+ "per_task": {
222
+ "LiftPegUpright-v1": {
223
+ "action_mse_to_best": 0.3945636663168746,
224
+ "expert_success_rate": 0.8229166666666666,
225
+ "num_groups": 96,
226
+ "oracle_success_rate": 0.9270833333333334,
227
+ "policy_expert_regret": 0.8837322552377979,
228
+ "policy_oracle_regret": 0.9900069705521067,
229
+ "policy_rollout_progress": 0.6311015527074536,
230
+ "policy_rollout_success_rate": 0.3125,
231
+ "restore_max_error": 3.5762786865234375e-07
232
+ },
233
+ "PickCube-v1": {
234
+ "action_mse_to_best": 0.3886390179019384,
235
+ "expert_success_rate": 0.9444444444444444,
236
+ "num_groups": 198,
237
+ "oracle_success_rate": 0.9595959595959596,
238
+ "policy_expert_regret": 1.0242929172730355,
239
+ "policy_oracle_regret": 1.0384738881122133,
240
+ "policy_rollout_progress": 0.6154013064530011,
241
+ "policy_rollout_success_rate": 0.30808080808080807,
242
+ "restore_max_error": 4.76837158203125e-07
243
+ },
244
+ "PullCube-v1": {
245
+ "action_mse_to_best": 0.7077205310265223,
246
+ "expert_success_rate": 0.24444444444444444,
247
+ "num_groups": 90,
248
+ "oracle_success_rate": 0.4666666666666667,
249
+ "policy_expert_regret": 0.29214862477965653,
250
+ "policy_oracle_regret": 0.49812486110151644,
251
+ "policy_rollout_progress": 0.340054856111399,
252
+ "policy_rollout_success_rate": 0.23333333333333334,
253
+ "restore_max_error": 4.0978193283081055e-07
254
+ },
255
+ "PushCube-v1": {
256
+ "action_mse_to_best": 0.41758197951729936,
257
+ "expert_success_rate": 0.8514851485148515,
258
+ "num_groups": 101,
259
+ "oracle_success_rate": 1.0,
260
+ "policy_expert_regret": 0.38248350021272604,
261
+ "policy_oracle_regret": 0.4372405781604276,
262
+ "policy_rollout_progress": 0.7904821941168001,
263
+ "policy_rollout_success_rate": 0.7722772277227723,
264
+ "restore_max_error": 4.76837158203125e-07
265
+ },
266
+ "StackCube-v1": {
267
+ "action_mse_to_best": 0.9588122832485371,
268
+ "expert_success_rate": 0.7666666666666667,
269
+ "num_groups": 90,
270
+ "oracle_success_rate": 0.9111111111111111,
271
+ "policy_expert_regret": 1.1859114145239193,
272
+ "policy_oracle_regret": 1.3349835796488656,
273
+ "policy_rollout_progress": 0.4025869076450666,
274
+ "policy_rollout_success_rate": 0.14444444444444443,
275
+ "restore_max_error": 4.76837158203125e-07
276
+ }
277
+ }
278
+ }
279
+ ]
280
+ }
results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale1p25_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_scale1p25.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: 32.52% +/- 2.47%
11
+ Gain vs h=16 rank checkpoint: +2.78%
12
+ Mean progress: 55.10%
13
+ Mean action MSE to best: 0.557
14
+
15
+ | seed | mode | k | retrieval K | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
16
+ |---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
17
+ | 0 | retrieval_residual | 16 | 1 | 1.25 | 0.00 | 0 | 0.00 | 30.61% | 53.11% | 85.74% | 0.627 |
18
+ | 1 | retrieval_residual | 16 | 1 | 1.25 | 0.00 | 0 | 0.00 | 31.65% | 54.96% | 86.96% | 0.511 |
19
+ | 2 | retrieval_residual | 16 | 1 | 1.25 | 0.00 | 0 | 0.00 | 35.30% | 57.24% | 87.65% | 0.534 |
results/paper_core_results.md CHANGED
@@ -14,13 +14,21 @@ baseline is the h=16 rank-checkpoint online rollout (`29.74%`).
14
  | Near-miss distillation policy, BC x5 | No | No | 28.29% | -1.45 pp | Stronger BC still stays below policy baseline |
15
  | Near-miss proposal + field, best-policy ckpt | No | No | 26.32% | -3.42 pp | Field scoring around the BC-selected checkpoint is unstable |
16
  | Near-miss proposal + field, field ckpt | No | No | 30.14% | +0.41 pp | Clean proposal route begins to recover the mechanism |
17
- | Near-miss proposal + field, BC x5 field ckpt | No | No | 32.93% | +3.19 pp | Best deployment-clean result so far; still far below same-state lattice |
18
  | Trust-region field optimization | No | No | 25.39% | -4.35 pp | Differentiable field ascent is a negative diagnostic; the field is not a generic action optimizer |
19
  | Best non-expert proposal policy | No | No | 27.88% | -1.86 pp | Broadening BC targets beyond near-miss does not solve proposal generation |
20
  | Best non-expert proposal + field | No | No | 26.49% | -3.25 pp | The field still needs local counterfactual proposal geometry |
21
  | Field-selected no-expert policy, seed-0 train map | No | No | 26.84% | -2.90 pp | Distilling the field's no-expert teacher from one split does not improve direct rollout |
22
  | Field-selected no-expert policy + field, seed-0 train map | No | No | 27.65% | -2.09 pp | Field scoring around that student remains below baseline |
23
- | Train-state residual retrieval | No | No | 32.12% | +2.38 pp | Transferred counterfactual residuals are a positive clean bridge but do not beat the near-miss proposal policy |
 
 
 
 
 
 
 
 
24
  | KNN train-state residual retrieval | No | No | 29.91% | +0.17 pp | Adding more retrieved tangent neighborhoods dilutes the signal |
25
  | Train-state near-miss residual retrieval | No | No | 14.06% smoke | -15.68 pp | Restricting to transferred near-miss residuals failed in smoke; full jobs canceled |
26
  | Lattice, no expert/no near-miss | Yes | No | 25.57% | -4.17 pp | Non-local negatives do not help |
@@ -38,17 +46,19 @@ Suggested main-table rows:
38
  5. Trust-region field optimization
39
  6. Best non-expert proposal + field
40
  7. Field-selected no-expert policy + field, seed-0 train map
41
- 8. Train-state residual retrieval
42
- 9. Lattice, near-miss only
43
- 10. Lattice, no expert
44
- 11. Lattice, full
45
- 12. Oracle ceiling
 
 
46
 
47
  Suggested claim:
48
 
49
  > DoVLA-CIL is not a better behavior-cloning policy; it is a local counterfactual action
50
- > selection rule. A deployment-clean near-miss proposal policy plus the field gives a small
51
- > gain, and transferred counterfactual residuals nearly match it, while field-gradient ascent,
52
- > KNN residual retrieval, broader non-expert BC targets, and seed-0 field-teacher distillation
53
- > fail. The large effect appears only when the field is queried on same-state intervention
54
- > proposals, and the mechanism is isolated to near-miss counterfactuals.
 
14
  | Near-miss distillation policy, BC x5 | No | No | 28.29% | -1.45 pp | Stronger BC still stays below policy baseline |
15
  | Near-miss proposal + field, best-policy ckpt | No | No | 26.32% | -3.42 pp | Field scoring around the BC-selected checkpoint is unstable |
16
  | Near-miss proposal + field, field ckpt | No | No | 30.14% | +0.41 pp | Clean proposal route begins to recover the mechanism |
17
+ | Near-miss proposal + field, BC x5 field ckpt | No | No | 32.93% | +3.19 pp | Strong clean bridge; still far below same-state lattice |
18
  | Trust-region field optimization | No | No | 25.39% | -4.35 pp | Differentiable field ascent is a negative diagnostic; the field is not a generic action optimizer |
19
  | Best non-expert proposal policy | No | No | 27.88% | -1.86 pp | Broadening BC targets beyond near-miss does not solve proposal generation |
20
  | Best non-expert proposal + field | No | No | 26.49% | -3.25 pp | The field still needs local counterfactual proposal geometry |
21
  | Field-selected no-expert policy, seed-0 train map | No | No | 26.84% | -2.90 pp | Distilling the field's no-expert teacher from one split does not improve direct rollout |
22
  | Field-selected no-expert policy + field, seed-0 train map | No | No | 27.65% | -2.09 pp | Field scoring around that student remains below baseline |
23
+ | Field-selected no-expert policy, aligned allmap | No | No | 28.00% | -1.74 pp | Full train/val target coverage does not fix the field-teacher student |
24
+ | Field-selected no-expert policy + field, aligned allmap | No | No | 26.49% | -3.25 pp | Field scoring around the aligned student remains below baseline |
25
+ | Train-state residual retrieval | No | No | 32.12% | +2.38 pp | Transferred counterfactual residuals are a positive clean bridge |
26
+ | Train-state residual retrieval, scale 0.25 | No | No | 32.93% | +3.19 pp | Smaller tangent step ties the previous clean best |
27
+ | Train-state residual retrieval, scale 0.50 | No | No | 33.33% | +3.59 pp | Current best deployment-clean bridge; calibrated local tangent transport |
28
+ | Train-state residual retrieval, scale 0.75 | No | No | 32.70% | +2.96 pp | Larger tangent steps begin to lose success |
29
+ | Train-state residual retrieval, scale 1.25 | No | No | 32.52% | +2.78 pp | Further scale increase does not help |
30
+ | Residual+Gaussian hybrid, K32 sigma0.35 | No | No | 31.30% | +1.57 pp | Adding policy-centered Gaussian proposals dilutes residual transport |
31
+ | Residual+Gaussian hybrid, K64 sigma0.50 | No | No | 30.90% | +1.16 pp | Larger hybrid search is worse |
32
  | KNN train-state residual retrieval | No | No | 29.91% | +0.17 pp | Adding more retrieved tangent neighborhoods dilutes the signal |
33
  | Train-state near-miss residual retrieval | No | No | 14.06% smoke | -15.68 pp | Restricting to transferred near-miss residuals failed in smoke; full jobs canceled |
34
  | Lattice, no expert/no near-miss | Yes | No | 25.57% | -4.17 pp | Non-local negatives do not help |
 
46
  5. Trust-region field optimization
47
  6. Best non-expert proposal + field
48
  7. Field-selected no-expert policy + field, seed-0 train map
49
+ 8. Field-selected no-expert policy + field, aligned allmap
50
+ 9. Train-state residual retrieval, scale 0.50
51
+ 10. Residual+Gaussian hybrid, K32 sigma0.35
52
+ 11. Lattice, near-miss only
53
+ 12. Lattice, no expert
54
+ 13. Lattice, full
55
+ 14. Oracle ceiling
56
 
57
  Suggested claim:
58
 
59
  > DoVLA-CIL is not a better behavior-cloning policy; it is a local counterfactual action
60
+ > selection rule. Deployment-clean calibrated counterfactual residual transport gives the
61
+ > strongest clean gain so far, while field-gradient ascent, KNN residual retrieval, broader
62
+ > non-expert BC targets, field-teacher distillation, and residual+Gaussian hybrids fail. The
63
+ > large effect appears only when the field is queried on same-state intervention proposals,
64
+ > and the mechanism is isolated to local near-miss counterfactual geometry.
results/paper_story_memo.md CHANGED
@@ -16,14 +16,15 @@ when queried on proposal geometry that matches those local counterfactuals.
16
  | Same-state local counterfactual proposals are the mechanism | near-miss-only lattice is 55.94%; removing expert+near_miss drops to 25.57% | Strongly supported |
17
  | Conservative same-state result is large | no-expert lattice is 56.99% vs 29.74% policy | Main result |
18
  | Full lattice gives upper result | full lattice is 69.33%, oracle is 86.78% | Strong but label expert proposal clearly |
19
- | Deployment-clean proposal is currently a bottleneck | best clean proposal+field sweep is 32.93%, far below 56.99% | Supported |
20
  | Gradient-based field optimization does not solve the clean proposal gap | `field_optim` best observed result is 25.39% | Negative diagnostic |
21
  | A broader non-expert proposal target does not reduce the proposal gap | direct broad non-expert policy is 27.88%; with field scoring it is 26.49% | Negative diagnostic |
22
  | Counterfactual residuals transfer better than absolute retrieved actions | nearest residual retrieval is 32.12% vs absolute retrieval 28.93%; KNN4 residual drops to 29.91% | Supported as a clean bridge |
23
- | Residual magnitude may be the next clean bottleneck | scale sweep for nearest residual transport is pending | Pending |
24
- | Residual transport and Gaussian local proposals may be complementary | hybrid residual+Gaussian proposal jobs are pending | Pending |
25
  | Seed-0 train-split field-teacher distillation does not solve the proposal gap | direct student is 26.84%; with field scoring it is 27.65% | Negative diagnostic |
26
- | All-split field-teacher distillation may still help checkpointing/coverage | allmap training has 100% train/val target coverage; rollout eval is pending | Pending |
 
27
 
28
  ## Main Table Candidate
29
 
@@ -38,11 +39,14 @@ clean proposal result, the intended main rows are:
38
  5. Trust-region field optimization: 25.39%
39
  6. Broad non-expert proposal + field: 26.49%
40
  7. Field-selected no-expert proposal + field, seed-0 train map: 27.65%
41
- 8. Train-state residual retrieval: 32.12%
42
- 9. Lattice, near-miss only: 55.94%
43
- 10. Lattice, no expert: 56.99%
44
- 11. Lattice, full: 69.33%
45
- 12. Oracle ceiling: 86.78%
 
 
 
46
 
47
  ## Novelty Framing
48
 
@@ -64,56 +68,44 @@ test-time search. The cleaner novelty is:
64
  |---|---|---|
65
  | Same-state lattice is not deployment-clean | show no-expert lattice and near-miss-only mechanism; show retrieval/Gaussian failures | improve clean proposal route |
66
  | Full lattice includes expert proposal | label as upper deployment/ceiling, not main conservative result | keep no-expert row as main |
67
- | Gains are from candidate leakage, not learning | selection never reads candidate rewards; no-expert and near-miss-only isolate mechanism; field_optim, broad proposal BC, and seed-0 field-teacher distillation fail | add allmap field-teacher rollout evidence |
68
  | Method is just a bundle of tricks | use mechanism ablations to show one central idea: local counterfactual field | avoid presenting unrelated variants as core |
69
  | Not SOTA enough | current clean deploy result is modest | need external baselines and stronger proposal generator before claiming SOTA |
70
 
71
  ## Active Jobs
72
 
73
- Last checked: `2026-06-28 05:26 UTC`.
74
 
75
  - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
76
  direct rollout is 26.84%, field-guided best is 27.65%.
77
- - `14858449`: completed export of an all-split target map for aligned validation
78
- checkpoint selection and seed-invariant student train coverage.
79
- - `14858450`: completed 3-seed `field_selected_noexpert_bc5_allmap` training
80
- with 100% train/val target-map coverage.
81
- - `14858451`/`14858452`: pending direct rollout evaluation and summary for allmap.
82
- - `14858453`/`14858454`: pending field-guided rollout sweep and summary for allmap.
83
- - `14858455`: rebuild `paper_table_status.*` after allmap summaries.
84
  - `14858978`: completed CPU Apptainer unit smoke for residual-scale selection.
85
  Earlier smoke jobs `14858889`/`14858894` caught and fixed two scale wiring bugs
86
  before rollout jobs started.
87
- - `14858875`/`14858876`: pending nearest residual scale `0.25` eval/summary.
88
- - `14858877`/`14858878`: pending nearest residual scale `0.50` eval/summary.
89
- - `14858879`/`14858880`: pending nearest residual scale `0.75` eval/summary.
90
- - `14858881`/`14858882`: pending nearest residual scale `1.25` eval/summary.
91
- - `14858883`: rebuild `paper_table_status.*` after residual-scale summaries.
92
  - `14859041`: completed CPU Apptainer unit smoke for hybrid residual+Gaussian selection.
93
- - `14859042`/`14859043`: pending hybrid residual+Gaussian K32/sigma0.35 eval/summary.
94
- - `14859044`/`14859045`: pending hybrid residual+Gaussian K64/sigma0.50 eval/summary.
95
- - `14859046`: rebuild `paper_table_status.*` after hybrid summaries.
96
-
97
- ## Decision Rule For Field-Teacher Jobs
98
-
99
- - If allmap field-teacher distillation beats 32.93%, promote it as the best
100
- deployment-clean bridge and keep same-state lattice as the mechanism result.
101
- - If it lands near residual retrieval, present residual retrieval and
102
- field-teacher distillation as complementary evidence for transferable local
103
- counterfactual geometry.
104
- - If it fails, keep the central paper story focused on the same-state mechanism
105
- and the clean-proposal bottleneck, with residual retrieval as the strongest
106
- deployment-clean bridge.
107
-
108
- ## Decision Rule For Residual-Scale Jobs
109
-
110
- - If any residual scale beats 32.93%, promote tangent-transport residuals as the
111
- best deployment-clean bridge.
112
- - If a smaller scale beats 32.12% but not 32.93%, present it as evidence that
113
- counterfactual residuals transfer as local tangent directions with a calibrated
114
- step length.
115
- - If all scales fail, keep scale `1.0` nearest residual retrieval as the clean
116
- positive bridge and treat magnitude calibration as a negative ablation.
117
- - If a hybrid residual+Gaussian row beats both residual-only and Gaussian-only
118
- rows, frame it as complementarity between transported local tangent directions
119
- and policy-centered local exploration under one learned field.
 
16
  | Same-state local counterfactual proposals are the mechanism | near-miss-only lattice is 55.94%; removing expert+near_miss drops to 25.57% | Strongly supported |
17
  | Conservative same-state result is large | no-expert lattice is 56.99% vs 29.74% policy | Main result |
18
  | Full lattice gives upper result | full lattice is 69.33%, oracle is 86.78% | Strong but label expert proposal clearly |
19
+ | Deployment-clean proposal is currently a bottleneck | best clean residual transport is 33.33%, far below 56.99% | Supported |
20
  | Gradient-based field optimization does not solve the clean proposal gap | `field_optim` best observed result is 25.39% | Negative diagnostic |
21
  | A broader non-expert proposal target does not reduce the proposal gap | direct broad non-expert policy is 27.88%; with field scoring it is 26.49% | Negative diagnostic |
22
  | Counterfactual residuals transfer better than absolute retrieved actions | nearest residual retrieval is 32.12% vs absolute retrieval 28.93%; KNN4 residual drops to 29.91% | Supported as a clean bridge |
23
+ | Residual magnitude is a real clean-deployment knob | scale 0.50 reaches 33.33%; scale 0.25 ties 32.93%; larger scales fall back | Supported |
24
+ | Residual transport and Gaussian local proposals are not complementary here | hybrid K32/K64 reach 31.30%/30.90%, below residual-only | Negative diagnostic |
25
  | Seed-0 train-split field-teacher distillation does not solve the proposal gap | direct student is 26.84%; with field scoring it is 27.65% | Negative diagnostic |
26
+ | All-split field-teacher distillation does not fix checkpointing/coverage | allmap direct is 28.00%; field-guided best is 26.49% despite 100% target coverage | Negative diagnostic |
27
+ | Residual family consistency is the next hypothesis | field-selected random/wrong-direction residuals have low rollout success; masked residual jobs are active | Active |
28
 
29
  ## Main Table Candidate
30
 
 
39
  5. Trust-region field optimization: 25.39%
40
  6. Broad non-expert proposal + field: 26.49%
41
  7. Field-selected no-expert proposal + field, seed-0 train map: 27.65%
42
+ 8. Field-selected no-expert proposal + field, aligned allmap: 26.49%
43
+ 9. Train-state residual retrieval: 32.12%
44
+ 10. Train-state residual retrieval, scale 0.50: 33.33%
45
+ 11. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
46
+ 12. Lattice, near-miss only: 55.94%
47
+ 13. Lattice, no expert: 56.99%
48
+ 14. Lattice, full: 69.33%
49
+ 15. Oracle ceiling: 86.78%
50
 
51
  ## Novelty Framing
52
 
 
68
  |---|---|---|
69
  | Same-state lattice is not deployment-clean | show no-expert lattice and near-miss-only mechanism; show retrieval/Gaussian failures | improve clean proposal route |
70
  | Full lattice includes expert proposal | label as upper deployment/ceiling, not main conservative result | keep no-expert row as main |
71
+ | Gains are from candidate leakage, not learning | selection never reads candidate rewards; no-expert and near-miss-only isolate mechanism; field_optim, broad proposal BC, and field-teacher distillation fail | keep same-state rows labeled as mechanism, not clean deployment |
72
  | Method is just a bundle of tricks | use mechanism ablations to show one central idea: local counterfactual field | avoid presenting unrelated variants as core |
73
  | Not SOTA enough | current clean deploy result is modest | need external baselines and stronger proposal generator before claiming SOTA |
74
 
75
  ## Active Jobs
76
 
77
+ Last checked: `2026-06-28 05:47 UTC`.
78
 
79
  - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
80
  direct rollout is 26.84%, field-guided best is 27.65%.
81
+ - `14858449`-`14858455`: completed all-split `field_selected_noexpert_bc5_allmap`;
82
+ direct rollout is 28.00%, field-guided best is 26.49%, so aligned coverage did
83
+ not fix the proposal bottleneck.
 
 
 
 
84
  - `14858978`: completed CPU Apptainer unit smoke for residual-scale selection.
85
  Earlier smoke jobs `14858889`/`14858894` caught and fixed two scale wiring bugs
86
  before rollout jobs started.
87
+ - `14858875`-`14858883`: completed nearest residual scale sweep. Scale `0.50`
88
+ is the current best clean deployment bridge at 33.33%; scale `0.25` ties the
89
+ previous 32.93% clean best; larger scales are weaker.
 
 
90
  - `14859041`: completed CPU Apptainer unit smoke for hybrid residual+Gaussian selection.
91
+ - `14859042`-`14859046`: completed hybrid residual+Gaussian jobs; K32 reaches
92
+ 31.30% and K64 reaches 30.90%, both below residual-only transport.
93
+ - `14859141`/`14859142`: active masked residual eval/summary, scale `0.50`,
94
+ excluding `residual_random_negative`.
95
+ - `14859143`/`14859144`: active masked residual eval/summary, scale `0.50`,
96
+ excluding `residual_random_negative` and `residual_wrong_direction`.
97
+ - `14859145`/`14859146`: active masked residual eval/summary, scale `0.25`,
98
+ excluding `residual_random_negative` and `residual_wrong_direction`.
99
+ - `14859147`/`14859148`: active typed residual eval/summary, scale `0.50`,
100
+ keeping policy/no-op/wrong-gripper residual families.
101
+ - `14859149`: rebuild `paper_table_status.*` after masked residual summaries.
102
+
103
+ ## Decision Rule For Masked Residual Jobs
104
+
105
+ - If a masked row beats 33.33%, promote it as evidence that transferable
106
+ counterfactual residuals need family-consistent local tangent proposals, not
107
+ anti-goal residuals.
108
+ - If masks land near 33.33% but do not beat it, keep scale `0.50` as the clean
109
+ residual result and present masking as a diagnostic of field over-selection.
110
+ - If masks fail, keep the story focused on residual scale calibration and the
111
+ larger same-state counterfactual mechanism.
 
 
 
 
 
 
results/paper_table_status.json CHANGED
@@ -69,7 +69,7 @@
69
  "clean_deployment": "yes",
70
  "same_state_proposals": "no",
71
  "expert_proposal": "no",
72
- "story_role": "current best clean deployment bridge",
73
  "fallback_success": 0.3293,
74
  "pending_job": "",
75
  "path_exists": true,
@@ -289,6 +289,139 @@
289
  "best_config": null,
290
  "gain_vs_h16_policy": 0.035942028985507246
291
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
292
  {
293
  "key": "retrieval_residual_scale075",
294
  "label": "Train-state residual retrieval, scale 0.75",
 
69
  "clean_deployment": "yes",
70
  "same_state_proposals": "no",
71
  "expert_proposal": "no",
72
+ "story_role": "strong clean proposal-field bridge",
73
  "fallback_success": 0.3293,
74
  "pending_job": "",
75
  "path_exists": true,
 
289
  "best_config": null,
290
  "gain_vs_h16_policy": 0.035942028985507246
291
  },
292
+ {
293
+ "key": "retrieval_residual_scale050_zscore",
294
+ "label": "Train-state residual retrieval, scale 0.50, z-score retrieval",
295
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_zscore_summary.json",
296
+ "clean_deployment": "yes",
297
+ "same_state_proposals": "no",
298
+ "expert_proposal": "no",
299
+ "story_role": "state-normalized tangent retrieval ablation",
300
+ "fallback_success": null,
301
+ "pending_job": "",
302
+ "path_exists": false,
303
+ "status": "pending",
304
+ "success": null,
305
+ "std_success": null,
306
+ "completed_seeds": null,
307
+ "num_completed": null,
308
+ "best_config": null,
309
+ "gain_vs_h16_policy": null
310
+ },
311
+ {
312
+ "key": "retrieval_residual_scale050_zscore_no_random_wrongdir",
313
+ "label": "Train-state residual retrieval, scale 0.50, z-score retrieval, no random/wrong-direction residuals",
314
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_zscore_no_random_wrongdir_summary.json",
315
+ "clean_deployment": "yes",
316
+ "same_state_proposals": "no",
317
+ "expert_proposal": "no",
318
+ "story_role": "state-normalized typed tangent retrieval ablation",
319
+ "fallback_success": null,
320
+ "pending_job": "",
321
+ "path_exists": false,
322
+ "status": "pending",
323
+ "success": null,
324
+ "std_success": null,
325
+ "completed_seeds": null,
326
+ "num_completed": null,
327
+ "best_config": null,
328
+ "gain_vs_h16_policy": null
329
+ },
330
+ {
331
+ "key": "retrieval_residual_scale025_zscore_no_random_wrongdir",
332
+ "label": "Train-state residual retrieval, scale 0.25, z-score retrieval, no random/wrong-direction residuals",
333
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_zscore_no_random_wrongdir_summary.json",
334
+ "clean_deployment": "yes",
335
+ "same_state_proposals": "no",
336
+ "expert_proposal": "no",
337
+ "story_role": "state-normalized typed tangent retrieval ablation",
338
+ "fallback_success": null,
339
+ "pending_job": "",
340
+ "path_exists": false,
341
+ "status": "pending",
342
+ "success": null,
343
+ "std_success": null,
344
+ "completed_seeds": null,
345
+ "num_completed": null,
346
+ "best_config": null,
347
+ "gain_vs_h16_policy": null
348
+ },
349
+ {
350
+ "key": "retrieval_residual_scale050_no_random",
351
+ "label": "Train-state residual retrieval, scale 0.50, no random residuals",
352
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_no_random_summary.json",
353
+ "clean_deployment": "yes",
354
+ "same_state_proposals": "no",
355
+ "expert_proposal": "no",
356
+ "story_role": "anti-goal residual family mask ablation",
357
+ "fallback_success": null,
358
+ "pending_job": "14859141/14859142",
359
+ "path_exists": false,
360
+ "status": "pending",
361
+ "success": null,
362
+ "std_success": null,
363
+ "completed_seeds": null,
364
+ "num_completed": null,
365
+ "best_config": null,
366
+ "gain_vs_h16_policy": null
367
+ },
368
+ {
369
+ "key": "retrieval_residual_scale050_no_random_wrongdir",
370
+ "label": "Train-state residual retrieval, scale 0.50, no random/wrong-direction residuals",
371
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_no_random_wrongdir_summary.json",
372
+ "clean_deployment": "yes",
373
+ "same_state_proposals": "no",
374
+ "expert_proposal": "no",
375
+ "story_role": "anti-goal residual family mask ablation",
376
+ "fallback_success": null,
377
+ "pending_job": "14859143/14859144",
378
+ "path_exists": false,
379
+ "status": "pending",
380
+ "success": null,
381
+ "std_success": null,
382
+ "completed_seeds": null,
383
+ "num_completed": null,
384
+ "best_config": null,
385
+ "gain_vs_h16_policy": null
386
+ },
387
+ {
388
+ "key": "retrieval_residual_scale025_no_random_wrongdir",
389
+ "label": "Train-state residual retrieval, scale 0.25, no random/wrong-direction residuals",
390
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_no_random_wrongdir_summary.json",
391
+ "clean_deployment": "yes",
392
+ "same_state_proposals": "no",
393
+ "expert_proposal": "no",
394
+ "story_role": "anti-goal residual family mask ablation",
395
+ "fallback_success": null,
396
+ "pending_job": "14859145/14859146",
397
+ "path_exists": false,
398
+ "status": "pending",
399
+ "success": null,
400
+ "std_success": null,
401
+ "completed_seeds": null,
402
+ "num_completed": null,
403
+ "best_config": null,
404
+ "gain_vs_h16_policy": null
405
+ },
406
+ {
407
+ "key": "retrieval_residual_scale050_safe_types",
408
+ "label": "Train-state residual retrieval, scale 0.50, policy/no-op/wrong-gripper residuals",
409
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_safe_types_summary.json",
410
+ "clean_deployment": "yes",
411
+ "same_state_proposals": "no",
412
+ "expert_proposal": "no",
413
+ "story_role": "typed tangent-family mask ablation",
414
+ "fallback_success": null,
415
+ "pending_job": "14859147/14859148",
416
+ "path_exists": false,
417
+ "status": "pending",
418
+ "success": null,
419
+ "std_success": null,
420
+ "completed_seeds": null,
421
+ "num_completed": null,
422
+ "best_config": null,
423
+ "gain_vs_h16_policy": null
424
+ },
425
  {
426
  "key": "retrieval_residual_scale075",
427
  "label": "Train-state residual retrieval, scale 0.75",
results/paper_table_status.md CHANGED
@@ -7,7 +7,7 @@ Baseline h=16 policy: 29.74%
7
  | h16_policy | Direct h=16 policy | fallback canonical | 29.74% | +0.00 pp | yes | no | no | behavior-cloning baseline |
8
  | gaussian_field | Gaussian field search | complete k32_sigma0.35 | 29.10% | -0.64 pp | yes | no | no | negative off-manifold field ablation |
9
  | retrieval_lattice_no_expert | Nearest train-state lattice, no expert | complete | 27.13% | -2.61 pp | yes | no | no | negative generic action-library ablation |
10
- | near_miss_policy_bc5_field | Near-miss proposal policy + field | complete k64_sigma0.50 | 32.93% | +3.19 pp | yes | no | no | current best clean deployment bridge |
11
  | field_optim | Trust-region field optimization | complete k32_sigma0.50 | 25.39% | -4.35 pp | yes | no | no | differentiable field-ascent diagnostic |
12
  | nonexpert_policy_bc5 | Best non-expert proposal policy | complete | 27.88% | -1.86 pp | yes | no | no | broader non-expert proposal-model ablation |
13
  | nonexpert_policy_bc5_field | Best non-expert proposal policy + field | complete k64_sigma0.50 | 26.49% | -3.25 pp | yes | no | no | broader proposal-field ablation |
@@ -18,6 +18,13 @@ Baseline h=16 policy: 29.74%
18
  | retrieval_residual | Train-state counterfactual residual retrieval | complete | 32.12% | +2.38 pp | yes | no | no | transferable local tangent proposal |
19
  | retrieval_residual_scale025 | Train-state residual retrieval, scale 0.25 | complete | 32.93% | +3.19 pp | yes | no | no | tangent transport scale ablation |
20
  | retrieval_residual_scale050 | Train-state residual retrieval, scale 0.50 | complete | 33.33% | +3.59 pp | yes | no | no | tangent transport scale ablation |
 
 
 
 
 
 
 
21
  | retrieval_residual_scale075 | Train-state residual retrieval, scale 0.75 | complete | 32.70% | +2.96 pp | yes | no | no | tangent transport scale ablation |
22
  | retrieval_residual_scale125 | Train-state residual retrieval, scale 1.25 | complete | 32.52% | +2.78 pp | yes | no | no | tangent transport scale ablation |
23
  | retrieval_residual_hybrid_k32 | Train-state residual + Gaussian proposals, K32 sigma0.35 | complete | 31.30% | +1.57 pp | yes | no | no | hybrid tangent/local proposal bridge |
 
7
  | h16_policy | Direct h=16 policy | fallback canonical | 29.74% | +0.00 pp | yes | no | no | behavior-cloning baseline |
8
  | gaussian_field | Gaussian field search | complete k32_sigma0.35 | 29.10% | -0.64 pp | yes | no | no | negative off-manifold field ablation |
9
  | retrieval_lattice_no_expert | Nearest train-state lattice, no expert | complete | 27.13% | -2.61 pp | yes | no | no | negative generic action-library ablation |
10
+ | near_miss_policy_bc5_field | Near-miss proposal policy + field | complete k64_sigma0.50 | 32.93% | +3.19 pp | yes | no | no | strong clean proposal-field bridge |
11
  | field_optim | Trust-region field optimization | complete k32_sigma0.50 | 25.39% | -4.35 pp | yes | no | no | differentiable field-ascent diagnostic |
12
  | nonexpert_policy_bc5 | Best non-expert proposal policy | complete | 27.88% | -1.86 pp | yes | no | no | broader non-expert proposal-model ablation |
13
  | nonexpert_policy_bc5_field | Best non-expert proposal policy + field | complete k64_sigma0.50 | 26.49% | -3.25 pp | yes | no | no | broader proposal-field ablation |
 
18
  | retrieval_residual | Train-state counterfactual residual retrieval | complete | 32.12% | +2.38 pp | yes | no | no | transferable local tangent proposal |
19
  | retrieval_residual_scale025 | Train-state residual retrieval, scale 0.25 | complete | 32.93% | +3.19 pp | yes | no | no | tangent transport scale ablation |
20
  | retrieval_residual_scale050 | Train-state residual retrieval, scale 0.50 | complete | 33.33% | +3.59 pp | yes | no | no | tangent transport scale ablation |
21
+ | retrieval_residual_scale050_zscore | Train-state residual retrieval, scale 0.50, z-score retrieval | pending | pending | pending | yes | no | no | state-normalized tangent retrieval ablation |
22
+ | retrieval_residual_scale050_zscore_no_random_wrongdir | Train-state residual retrieval, scale 0.50, z-score retrieval, no random/wrong-direction residuals | pending | pending | pending | yes | no | no | state-normalized typed tangent retrieval ablation |
23
+ | retrieval_residual_scale025_zscore_no_random_wrongdir | Train-state residual retrieval, scale 0.25, z-score retrieval, no random/wrong-direction residuals | pending | pending | pending | yes | no | no | state-normalized typed tangent retrieval ablation |
24
+ | retrieval_residual_scale050_no_random | Train-state residual retrieval, scale 0.50, no random residuals | pending 14859141/14859142 | pending | pending | yes | no | no | anti-goal residual family mask ablation |
25
+ | retrieval_residual_scale050_no_random_wrongdir | Train-state residual retrieval, scale 0.50, no random/wrong-direction residuals | pending 14859143/14859144 | pending | pending | yes | no | no | anti-goal residual family mask ablation |
26
+ | retrieval_residual_scale025_no_random_wrongdir | Train-state residual retrieval, scale 0.25, no random/wrong-direction residuals | pending 14859145/14859146 | pending | pending | yes | no | no | anti-goal residual family mask ablation |
27
+ | retrieval_residual_scale050_safe_types | Train-state residual retrieval, scale 0.50, policy/no-op/wrong-gripper residuals | pending 14859147/14859148 | pending | pending | yes | no | no | typed tangent-family mask ablation |
28
  | retrieval_residual_scale075 | Train-state residual retrieval, scale 0.75 | complete | 32.70% | +2.96 pp | yes | no | no | tangent transport scale ablation |
29
  | retrieval_residual_scale125 | Train-state residual retrieval, scale 1.25 | complete | 32.52% | +2.78 pp | yes | no | no | tangent transport scale ablation |
30
  | retrieval_residual_hybrid_k32 | Train-state residual + Gaussian proposals, K32 sigma0.35 | complete | 31.30% | +1.57 pp | yes | no | no | hybrid tangent/local proposal bridge |
scripts/build_paper_table_status.py CHANGED
@@ -62,7 +62,7 @@ SPECS = [
62
  clean_deployment="yes",
63
  same_state_proposals="no",
64
  expert_proposal="no",
65
- story_role="current best clean deployment bridge",
66
  fallback_success=0.3293,
67
  ),
68
  ResultSpec(
@@ -165,6 +165,73 @@ SPECS = [
165
  story_role="tangent transport scale ablation",
166
  pending_job="14858877/14858878",
167
  ),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
168
  ResultSpec(
169
  key="retrieval_residual_scale075",
170
  label="Train-state residual retrieval, scale 0.75",
 
62
  clean_deployment="yes",
63
  same_state_proposals="no",
64
  expert_proposal="no",
65
+ story_role="strong clean proposal-field bridge",
66
  fallback_success=0.3293,
67
  ),
68
  ResultSpec(
 
165
  story_role="tangent transport scale ablation",
166
  pending_job="14858877/14858878",
167
  ),
168
+ ResultSpec(
169
+ key="retrieval_residual_scale050_zscore",
170
+ label="Train-state residual retrieval, scale 0.50, z-score retrieval",
171
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_zscore_summary.json",
172
+ clean_deployment="yes",
173
+ same_state_proposals="no",
174
+ expert_proposal="no",
175
+ story_role="state-normalized tangent retrieval ablation",
176
+ ),
177
+ ResultSpec(
178
+ key="retrieval_residual_scale050_zscore_no_random_wrongdir",
179
+ label="Train-state residual retrieval, scale 0.50, z-score retrieval, no random/wrong-direction residuals",
180
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_zscore_no_random_wrongdir_summary.json",
181
+ clean_deployment="yes",
182
+ same_state_proposals="no",
183
+ expert_proposal="no",
184
+ story_role="state-normalized typed tangent retrieval ablation",
185
+ ),
186
+ ResultSpec(
187
+ key="retrieval_residual_scale025_zscore_no_random_wrongdir",
188
+ label="Train-state residual retrieval, scale 0.25, z-score retrieval, no random/wrong-direction residuals",
189
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_zscore_no_random_wrongdir_summary.json",
190
+ clean_deployment="yes",
191
+ same_state_proposals="no",
192
+ expert_proposal="no",
193
+ story_role="state-normalized typed tangent retrieval ablation",
194
+ ),
195
+ ResultSpec(
196
+ key="retrieval_residual_scale050_no_random",
197
+ label="Train-state residual retrieval, scale 0.50, no random residuals",
198
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_no_random_summary.json",
199
+ clean_deployment="yes",
200
+ same_state_proposals="no",
201
+ expert_proposal="no",
202
+ story_role="anti-goal residual family mask ablation",
203
+ pending_job="14859141/14859142",
204
+ ),
205
+ ResultSpec(
206
+ key="retrieval_residual_scale050_no_random_wrongdir",
207
+ label="Train-state residual retrieval, scale 0.50, no random/wrong-direction residuals",
208
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_no_random_wrongdir_summary.json",
209
+ clean_deployment="yes",
210
+ same_state_proposals="no",
211
+ expert_proposal="no",
212
+ story_role="anti-goal residual family mask ablation",
213
+ pending_job="14859143/14859144",
214
+ ),
215
+ ResultSpec(
216
+ key="retrieval_residual_scale025_no_random_wrongdir",
217
+ label="Train-state residual retrieval, scale 0.25, no random/wrong-direction residuals",
218
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_no_random_wrongdir_summary.json",
219
+ clean_deployment="yes",
220
+ same_state_proposals="no",
221
+ expert_proposal="no",
222
+ story_role="anti-goal residual family mask ablation",
223
+ pending_job="14859145/14859146",
224
+ ),
225
+ ResultSpec(
226
+ key="retrieval_residual_scale050_safe_types",
227
+ label="Train-state residual retrieval, scale 0.50, policy/no-op/wrong-gripper residuals",
228
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_safe_types_summary.json",
229
+ clean_deployment="yes",
230
+ same_state_proposals="no",
231
+ expert_proposal="no",
232
+ story_role="typed tangent-family mask ablation",
233
+ pending_job="14859147/14859148",
234
+ ),
235
  ResultSpec(
236
  key="retrieval_residual_scale075",
237
  label="Train-state residual retrieval, scale 0.75",
scripts/eval_maniskill_policy_rollout.py CHANGED
@@ -103,6 +103,13 @@ def main(argv: list[str] | None = None) -> int:
103
  default=1,
104
  help="Nearest train states to use for retrieval_lattice/retrieval_residual proposals.",
105
  )
 
 
 
 
 
 
 
106
  parser.add_argument(
107
  "--retrieval-residual-scale",
108
  type=float,
@@ -138,6 +145,7 @@ def main(argv: list[str] | None = None) -> int:
138
  field_optim_trust_radius=args.field_optim_trust_radius,
139
  field_optim_l2_penalty=args.field_optim_l2_penalty,
140
  retrieval_neighbors=args.retrieval_neighbors,
 
141
  retrieval_residual_scale=args.retrieval_residual_scale,
142
  lattice_exclude_types=lattice_exclude_types,
143
  )
 
103
  default=1,
104
  help="Nearest train states to use for retrieval_lattice/retrieval_residual proposals.",
105
  )
106
+ parser.add_argument(
107
+ "--retrieval-metric",
108
+ choices=("raw", "zscore"),
109
+ default="raw",
110
+ help="State-space metric for retrieval proposals. 'raw' preserves earlier results; "
111
+ "'zscore' standardizes each task's train-bank features before nearest-neighbor lookup.",
112
+ )
113
  parser.add_argument(
114
  "--retrieval-residual-scale",
115
  type=float,
 
145
  field_optim_trust_radius=args.field_optim_trust_radius,
146
  field_optim_l2_penalty=args.field_optim_l2_penalty,
147
  retrieval_neighbors=args.retrieval_neighbors,
148
+ retrieval_metric=args.retrieval_metric,
149
  retrieval_residual_scale=args.retrieval_residual_scale,
150
  lattice_exclude_types=lattice_exclude_types,
151
  )
scripts/slurm/eval_maniskill_policy_rollout.sbatch CHANGED
@@ -49,6 +49,7 @@ FIELD_OPTIM_STEP_SIZE="${FIELD_OPTIM_STEP_SIZE:-0.05}"
49
  FIELD_OPTIM_TRUST_RADIUS="${FIELD_OPTIM_TRUST_RADIUS:-0.5}"
50
  FIELD_OPTIM_L2_PENALTY="${FIELD_OPTIM_L2_PENALTY:-0.0}"
51
  RETRIEVAL_NEIGHBORS="${RETRIEVAL_NEIGHBORS:-1}"
 
52
  RETRIEVAL_RESIDUAL_SCALE="${RETRIEVAL_RESIDUAL_SCALE:-1.0}"
53
  LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
54
  if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
@@ -98,6 +99,7 @@ apptainer exec --nv \
98
  --field-optim-trust-radius "$FIELD_OPTIM_TRUST_RADIUS" \
99
  --field-optim-l2-penalty "$FIELD_OPTIM_L2_PENALTY" \
100
  --retrieval-neighbors "$RETRIEVAL_NEIGHBORS" \
 
101
  --retrieval-residual-scale "$RETRIEVAL_RESIDUAL_SCALE" \
102
  --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
103
  "${EXTRA_ARGS[@]}"
 
49
  FIELD_OPTIM_TRUST_RADIUS="${FIELD_OPTIM_TRUST_RADIUS:-0.5}"
50
  FIELD_OPTIM_L2_PENALTY="${FIELD_OPTIM_L2_PENALTY:-0.0}"
51
  RETRIEVAL_NEIGHBORS="${RETRIEVAL_NEIGHBORS:-1}"
52
+ RETRIEVAL_METRIC="${RETRIEVAL_METRIC:-raw}"
53
  RETRIEVAL_RESIDUAL_SCALE="${RETRIEVAL_RESIDUAL_SCALE:-1.0}"
54
  LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
55
  if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
 
99
  --field-optim-trust-radius "$FIELD_OPTIM_TRUST_RADIUS" \
100
  --field-optim-l2-penalty "$FIELD_OPTIM_L2_PENALTY" \
101
  --retrieval-neighbors "$RETRIEVAL_NEIGHBORS" \
102
+ --retrieval-metric "$RETRIEVAL_METRIC" \
103
  --retrieval-residual-scale "$RETRIEVAL_RESIDUAL_SCALE" \
104
  --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
105
  "${EXTRA_ARGS[@]}"
scripts/slurm/summarize_h16_policy_ckpt.sbatch CHANGED
@@ -60,6 +60,7 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
60
  "field_optim_trust_radius": data.get("field_optim_trust_radius", 0.0),
61
  "field_optim_l2_penalty": data.get("field_optim_l2_penalty", 0.0),
62
  "retrieval_neighbors": data.get("retrieval_neighbors", 0),
 
63
  "retrieval_residual_scale": data.get("retrieval_residual_scale", 0.0),
64
  "policy_rollout_success_rate": data.get("policy_rollout_success_rate", 0.0),
65
  "policy_rollout_progress": data.get("policy_rollout_progress", 0.0),
@@ -121,17 +122,18 @@ lines = [
121
  f"Mean progress: {summary['mean_progress']:.2%}",
122
  f"Mean action MSE to best: {summary['mean_action_mse_to_best']:.3f}",
123
  "",
124
- "| seed | mode | k | retrieval K | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |",
125
- "|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|",
126
  ]
127
  for row in rows:
128
  lines.append(
129
- "| {seed} | {mode} | {k} | {retrieval} | {scale:.2f} | {sigma:.2f} | {steps} | {trust:.2f} | "
130
  "{success:.2%} | {progress:.2%} | {oracle:.2%} | {mse:.3f} |".format(
131
  seed=row["seed"],
132
  mode=row.get("selection_mode") or "policy",
133
  k=row.get("num_candidates") or 1,
134
  retrieval=row.get("retrieval_neighbors") or 0,
 
135
  scale=row.get("retrieval_residual_scale") or 0.0,
136
  sigma=row.get("candidate_sigma") or 0.0,
137
  steps=row.get("field_optim_steps") or 0,
 
60
  "field_optim_trust_radius": data.get("field_optim_trust_radius", 0.0),
61
  "field_optim_l2_penalty": data.get("field_optim_l2_penalty", 0.0),
62
  "retrieval_neighbors": data.get("retrieval_neighbors", 0),
63
+ "retrieval_metric": data.get("retrieval_metric", "none"),
64
  "retrieval_residual_scale": data.get("retrieval_residual_scale", 0.0),
65
  "policy_rollout_success_rate": data.get("policy_rollout_success_rate", 0.0),
66
  "policy_rollout_progress": data.get("policy_rollout_progress", 0.0),
 
122
  f"Mean progress: {summary['mean_progress']:.2%}",
123
  f"Mean action MSE to best: {summary['mean_action_mse_to_best']:.3f}",
124
  "",
125
+ "| seed | mode | k | retrieval K | retrieval metric | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |",
126
+ "|---:|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|",
127
  ]
128
  for row in rows:
129
  lines.append(
130
+ "| {seed} | {mode} | {k} | {retrieval} | {metric} | {scale:.2f} | {sigma:.2f} | {steps} | {trust:.2f} | "
131
  "{success:.2%} | {progress:.2%} | {oracle:.2%} | {mse:.3f} |".format(
132
  seed=row["seed"],
133
  mode=row.get("selection_mode") or "policy",
134
  k=row.get("num_candidates") or 1,
135
  retrieval=row.get("retrieval_neighbors") or 0,
136
+ metric=row.get("retrieval_metric") or "none",
137
  scale=row.get("retrieval_residual_scale") or 0.0,
138
  sigma=row.get("candidate_sigma") or 0.0,
139
  steps=row.get("field_optim_steps") or 0,
tests/test_maniskill_policy_rollout.py CHANGED
@@ -497,3 +497,83 @@ def test_retrieval_residual_candidates_use_knn_train_residuals() -> None:
497
  np.asarray(attached.candidate_action_values, dtype=np.float32),
498
  np.asarray([[[0.0, 0.0]], [[0.2, 0.0]], [[0.0, 0.0]], [[-0.3, 0.0]]]),
499
  )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
497
  np.asarray(attached.candidate_action_values, dtype=np.float32),
498
  np.asarray([[[0.0, 0.0]], [[0.2, 0.0]], [[0.0, 0.0]], [[-0.3, 0.0]]]),
499
  )
500
+
501
+
502
+ def test_retrieval_residual_zscore_metric_standardizes_train_bank_features() -> None:
503
+ def record(group_id: str, candidate_type: str, action_value: float, feature: list[float]):
504
+ return SimpleNamespace(
505
+ group_id=group_id,
506
+ task_id="PickCube-v1",
507
+ candidate_type=candidate_type,
508
+ record_id=f"{group_id}-{candidate_type}-{action_value}",
509
+ observation_inline={"features": feature},
510
+ action_chunk=ActionChunk(
511
+ representation="continuous",
512
+ horizon=1,
513
+ values=[[action_value, 0.0]],
514
+ ),
515
+ )
516
+
517
+ groups = {
518
+ "train_a": [
519
+ record("train_a", "expert", 1.0, [0.0, 0.0]),
520
+ record("train_a", "near_miss", 1.1, [0.0, 0.0]),
521
+ ],
522
+ "train_b": [
523
+ record("train_b", "expert", 2.0, [10.0, 1.0]),
524
+ record("train_b", "near_miss", 2.2, [10.0, 1.0]),
525
+ ],
526
+ "train_c": [
527
+ record("train_c", "expert", 3.0, [11.0, 1.0]),
528
+ record("train_c", "near_miss", 3.3, [11.0, 1.0]),
529
+ ],
530
+ "heldout": [
531
+ record("heldout", "expert", 9.0, [0.0, 1.0]),
532
+ record("heldout", "near_miss", 9.9, [0.0, 1.0]),
533
+ ],
534
+ }
535
+ dataset = SimpleNamespace(
536
+ group_ids=list(groups),
537
+ get_group=lambda group_id: groups[group_id],
538
+ )
539
+ case = _RolloutCase(
540
+ group_id="heldout",
541
+ task_id="PickCube-v1",
542
+ source_dataset=Path("."),
543
+ state={},
544
+ observation={"features": [0.0, 1.0]},
545
+ instruction="pick",
546
+ oracle_score=1.0,
547
+ oracle_success=True,
548
+ expert_score=1.0,
549
+ expert_success=True,
550
+ best_action_values=[[9.9, 0.0]],
551
+ candidate_action_values=[],
552
+ candidate_types=[],
553
+ )
554
+
555
+ [raw_attached] = _attach_retrieved_residual_candidates(
556
+ dataset,
557
+ [case],
558
+ heldout_group_ids=["heldout"],
559
+ obs_dim=2,
560
+ observation_mode="state",
561
+ retrieval_neighbors=1,
562
+ retrieval_metric="raw",
563
+ )
564
+ [zscore_attached] = _attach_retrieved_residual_candidates(
565
+ dataset,
566
+ [case],
567
+ heldout_group_ids=["heldout"],
568
+ obs_dim=2,
569
+ observation_mode="state",
570
+ retrieval_neighbors=1,
571
+ retrieval_metric="zscore",
572
+ )
573
+
574
+ assert raw_attached.candidate_source_group_id == "train_a"
575
+ assert zscore_attached.candidate_source_group_id == "train_b"
576
+ assert np.allclose(
577
+ np.asarray(zscore_attached.candidate_action_values, dtype=np.float32),
578
+ np.asarray([[[0.0, 0.0]], [[0.2, 0.0]]]),
579
+ )