Auto-sync: 2026-06-29 13:04:05 (part 2)
Browse files- results/paper_analysis.json +111 -19
- results/paper_analysis.md +9 -9
- results/paper_table_status.json +15 -15
- results/paper_table_status.md +2 -2
- scripts/build_paper_analysis.py +8 -0
- scripts/build_paper_table_status.py +9 -0
results/paper_analysis.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
-
"best_clean_key": "
|
| 3 |
-
"generated_utc": "2026-06-
|
| 4 |
"mechanism_gap": {
|
| 5 |
-
"best_clean_vs_direct_same_ckpt": 0.
|
| 6 |
-
"best_clean_vs_h16": 0.
|
| 7 |
"same_state_full_vs_no_expert": 0.12347826086956515,
|
| 8 |
-
"same_state_no_expert_vs_best_clean": 0.
|
| 9 |
"same_state_no_expert_vs_h16": 0.27246376811594214
|
| 10 |
},
|
| 11 |
"methods": {
|
|
@@ -948,9 +948,101 @@
|
|
| 948 |
"std_success": 0.011579701075616346
|
| 949 |
},
|
| 950 |
"residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003": {
|
|
|
|
| 951 |
"label": "K4 composed type-consensus tangents, masked, drop near-miss+no-op composite",
|
| 952 |
-
"
|
| 953 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 954 |
},
|
| 955 |
"residual_k4_composemasked_grid035040045": {
|
| 956 |
"ci95_success": 0.03024411731871433,
|
|
@@ -6321,12 +6413,12 @@
|
|
| 6321 |
},
|
| 6322 |
"paired_deltas": {
|
| 6323 |
"best_clean - canonical_h16": {
|
| 6324 |
-
"ci95_delta": 0.
|
| 6325 |
-
"left": "
|
| 6326 |
-
"mean_delta": 0.
|
| 6327 |
"right": "h16_policy_canonical",
|
| 6328 |
"seed_deltas": {
|
| 6329 |
-
"0": 0.
|
| 6330 |
"1": 0.038260869565217404,
|
| 6331 |
"2": 0.06782608695652176
|
| 6332 |
},
|
|
@@ -6335,15 +6427,15 @@
|
|
| 6335 |
1,
|
| 6336 |
2
|
| 6337 |
],
|
| 6338 |
-
"std_delta": 0.
|
| 6339 |
},
|
| 6340 |
"best_clean - direct_same_ckpt": {
|
| 6341 |
-
"ci95_delta": 0.
|
| 6342 |
-
"left": "
|
| 6343 |
-
"mean_delta": 0.
|
| 6344 |
"right": "near_miss_policy_bc5",
|
| 6345 |
"seed_deltas": {
|
| 6346 |
-
"0": 0.
|
| 6347 |
"1": 0.055652173913043446,
|
| 6348 |
"2": 0.08869565217391306
|
| 6349 |
},
|
|
@@ -6352,7 +6444,7 @@
|
|
| 6352 |
1,
|
| 6353 |
2
|
| 6354 |
],
|
| 6355 |
-
"std_delta": 0.
|
| 6356 |
},
|
| 6357 |
"full_lattice - no_expert_lattice": {
|
| 6358 |
"ci95_delta": 0.02628111194117426,
|
|
@@ -6408,14 +6500,14 @@
|
|
| 6408 |
},
|
| 6409 |
"per_task_deltas": {
|
| 6410 |
"best_clean_vs_h16": {
|
| 6411 |
-
"LiftPegUpright-v1": 0.
|
| 6412 |
"PickCube-v1": 0.1219595621769535,
|
| 6413 |
"PullCube-v1": 0.006887167413483192,
|
| 6414 |
"PushCube-v1": 0.006846429323783432,
|
| 6415 |
"StackCube-v1": 0.0594738594738595
|
| 6416 |
},
|
| 6417 |
"no_expert_vs_best_clean": {
|
| 6418 |
-
"LiftPegUpright-v1": 0.
|
| 6419 |
"PickCube-v1": 0.29366195398804096,
|
| 6420 |
"PullCube-v1": -0.006261866788182552,
|
| 6421 |
"PushCube-v1": 0.043639384064915965,
|
|
|
|
| 1 |
{
|
| 2 |
+
"best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003",
|
| 3 |
+
"generated_utc": "2026-06-29T17:02:31+00:00",
|
| 4 |
"mechanism_gap": {
|
| 5 |
+
"best_clean_vs_direct_same_ckpt": 0.07304347826086954,
|
| 6 |
+
"best_clean_vs_h16": 0.058550724637681184,
|
| 7 |
"same_state_full_vs_no_expert": 0.12347826086956515,
|
| 8 |
+
"same_state_no_expert_vs_best_clean": 0.21391304347826096,
|
| 9 |
"same_state_no_expert_vs_h16": 0.27246376811594214
|
| 10 |
},
|
| 11 |
"methods": {
|
|
|
|
| 948 |
"std_success": 0.011579701075616346
|
| 949 |
},
|
| 950 |
"residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003": {
|
| 951 |
+
"ci95_success": 0.024567904418045634,
|
| 952 |
"label": "K4 composed type-consensus tangents, masked, drop near-miss+no-op composite",
|
| 953 |
+
"mean_action_mse_to_best": 0.40635193244135687,
|
| 954 |
+
"mean_progress": 0.5706848979810625,
|
| 955 |
+
"mean_success": 0.35594202898550725,
|
| 956 |
+
"num_completed": 3,
|
| 957 |
+
"per_task_success": {
|
| 958 |
+
"LiftPegUpright-v1": {
|
| 959 |
+
"mean_num_groups": 102.0,
|
| 960 |
+
"mean_success": 0.26340865087329823,
|
| 961 |
+
"std_success": 0.06057881385012102
|
| 962 |
+
},
|
| 963 |
+
"PickCube-v1": {
|
| 964 |
+
"mean_num_groups": 196.66666666666666,
|
| 965 |
+
"mean_success": 0.3278169093386485,
|
| 966 |
+
"std_success": 0.02533117351097172
|
| 967 |
+
},
|
| 968 |
+
"PullCube-v1": {
|
| 969 |
+
"mean_num_groups": 81.0,
|
| 970 |
+
"mean_success": 0.20474165211007314,
|
| 971 |
+
"std_success": 0.019188581862611215
|
| 972 |
+
},
|
| 973 |
+
"PushCube-v1": {
|
| 974 |
+
"mean_num_groups": 102.0,
|
| 975 |
+
"mean_success": 0.757403601199683,
|
| 976 |
+
"std_success": 0.07781968814460603
|
| 977 |
+
},
|
| 978 |
+
"StackCube-v1": {
|
| 979 |
+
"mean_num_groups": 93.33333333333333,
|
| 980 |
+
"mean_success": 0.20044030044030045,
|
| 981 |
+
"std_success": 0.08896982358351117
|
| 982 |
+
}
|
| 983 |
+
},
|
| 984 |
+
"seed_action_mse_to_best": {
|
| 985 |
+
"0": 0.39301921862784933,
|
| 986 |
+
"1": 0.39824521353549286,
|
| 987 |
+
"2": 0.4277913651607283
|
| 988 |
+
},
|
| 989 |
+
"seed_progress": {
|
| 990 |
+
"0": 0.5583201415569562,
|
| 991 |
+
"1": 0.5714951352645522,
|
| 992 |
+
"2": 0.5822394171216793
|
| 993 |
+
},
|
| 994 |
+
"seed_success": {
|
| 995 |
+
"0": 0.35304347826086957,
|
| 996 |
+
"1": 0.34782608695652173,
|
| 997 |
+
"2": 0.36695652173913046
|
| 998 |
+
},
|
| 999 |
+
"selected_candidate_type_counts": {
|
| 1000 |
+
"retrieval_residual_policy_residual": 1613,
|
| 1001 |
+
"retrieval_residual_residual_near_miss": 8,
|
| 1002 |
+
"retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
|
| 1003 |
+
"retrieval_residual_residual_no_op": 56,
|
| 1004 |
+
"retrieval_residual_residual_no_op+residual_wrong_gripper": 24,
|
| 1005 |
+
"retrieval_residual_residual_wrong_gripper": 19
|
| 1006 |
+
},
|
| 1007 |
+
"selected_residual_scale_counts": {
|
| 1008 |
+
"0.35": 1631,
|
| 1009 |
+
"0.4": 11,
|
| 1010 |
+
"0.45": 83
|
| 1011 |
+
},
|
| 1012 |
+
"selected_type_outcomes": {
|
| 1013 |
+
"retrieval_residual_policy_residual": {
|
| 1014 |
+
"count": 1613.0,
|
| 1015 |
+
"mean_progress": 0.560971513695882,
|
| 1016 |
+
"success_rate": 0.3453192808431494
|
| 1017 |
+
},
|
| 1018 |
+
"retrieval_residual_residual_near_miss": {
|
| 1019 |
+
"count": 8.0,
|
| 1020 |
+
"mean_progress": 0.8760369084775448,
|
| 1021 |
+
"success_rate": 0.75
|
| 1022 |
+
},
|
| 1023 |
+
"retrieval_residual_residual_near_miss+residual_wrong_gripper": {
|
| 1024 |
+
"count": 5.0,
|
| 1025 |
+
"mean_progress": 0.7557853102684021,
|
| 1026 |
+
"success_rate": 0.6
|
| 1027 |
+
},
|
| 1028 |
+
"retrieval_residual_residual_no_op": {
|
| 1029 |
+
"count": 56.0,
|
| 1030 |
+
"mean_progress": 0.7249722117225507,
|
| 1031 |
+
"success_rate": 0.5
|
| 1032 |
+
},
|
| 1033 |
+
"retrieval_residual_residual_no_op+residual_wrong_gripper": {
|
| 1034 |
+
"count": 24.0,
|
| 1035 |
+
"mean_progress": 0.671148752172788,
|
| 1036 |
+
"success_rate": 0.4583333333333333
|
| 1037 |
+
},
|
| 1038 |
+
"retrieval_residual_residual_wrong_gripper": {
|
| 1039 |
+
"count": 19.0,
|
| 1040 |
+
"mean_progress": 0.6363769314791027,
|
| 1041 |
+
"success_rate": 0.47368421052631576
|
| 1042 |
+
}
|
| 1043 |
+
},
|
| 1044 |
+
"source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
|
| 1045 |
+
"std_success": 0.009889114266221456
|
| 1046 |
},
|
| 1047 |
"residual_k4_composemasked_grid035040045": {
|
| 1048 |
"ci95_success": 0.03024411731871433,
|
|
|
|
| 6413 |
},
|
| 6414 |
"paired_deltas": {
|
| 6415 |
"best_clean - canonical_h16": {
|
| 6416 |
+
"ci95_delta": 0.04370704391798567,
|
| 6417 |
+
"left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003",
|
| 6418 |
+
"mean_delta": 0.05855072463768116,
|
| 6419 |
"right": "h16_policy_canonical",
|
| 6420 |
"seed_deltas": {
|
| 6421 |
+
"0": 0.06956521739130433,
|
| 6422 |
"1": 0.038260869565217404,
|
| 6423 |
"2": 0.06782608695652176
|
| 6424 |
},
|
|
|
|
| 6427 |
1,
|
| 6428 |
2
|
| 6429 |
],
|
| 6430 |
+
"std_delta": 0.01759303293390552
|
| 6431 |
},
|
| 6432 |
"best_clean - direct_same_ckpt": {
|
| 6433 |
+
"ci95_delta": 0.04121578446082111,
|
| 6434 |
+
"left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003",
|
| 6435 |
+
"mean_delta": 0.07304347826086956,
|
| 6436 |
"right": "near_miss_policy_bc5",
|
| 6437 |
"seed_deltas": {
|
| 6438 |
+
"0": 0.07478260869565218,
|
| 6439 |
"1": 0.055652173913043446,
|
| 6440 |
"2": 0.08869565217391306
|
| 6441 |
},
|
|
|
|
| 6444 |
1,
|
| 6445 |
2
|
| 6446 |
],
|
| 6447 |
+
"std_delta": 0.0165902469811643
|
| 6448 |
},
|
| 6449 |
"full_lattice - no_expert_lattice": {
|
| 6450 |
"ci95_delta": 0.02628111194117426,
|
|
|
|
| 6500 |
},
|
| 6501 |
"per_task_deltas": {
|
| 6502 |
"best_clean_vs_h16": {
|
| 6503 |
+
"LiftPegUpright-v1": 0.03278131303915899,
|
| 6504 |
"PickCube-v1": 0.1219595621769535,
|
| 6505 |
"PullCube-v1": 0.006887167413483192,
|
| 6506 |
"PushCube-v1": 0.006846429323783432,
|
| 6507 |
"StackCube-v1": 0.0594738594738595
|
| 6508 |
},
|
| 6509 |
"no_expert_vs_best_clean": {
|
| 6510 |
+
"LiftPegUpright-v1": 0.35374851747103375,
|
| 6511 |
"PickCube-v1": 0.29366195398804096,
|
| 6512 |
"PullCube-v1": -0.006261866788182552,
|
| 6513 |
"PushCube-v1": 0.043639384064915965,
|
results/paper_analysis.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# Paper Analysis
|
| 2 |
|
| 3 |
-
Generated: `2026-06-
|
| 4 |
|
| 5 |
## Main Seed Statistics
|
| 6 |
|
|
@@ -47,7 +47,7 @@ Generated: `2026-06-29T16:57:54+00:00`
|
|
| 47 |
| residual_k4_compose_grid035040045_noopbonus003 | K4 composed type-consensus tangents, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 34.14% +/- 1.58 | +/- 3.92 | 56.00% | 0.482 | +4.41 pp |
|
| 48 |
| residual_k4_composemasked_grid035040045 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45 | 3 | 35.30% +/- 1.22 | +/- 3.02 | 56.91% | 0.410 | +5.57 pp |
|
| 49 |
| residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 35.54% +/- 1.02 | +/- 2.53 | 57.02% | 0.411 | +5.80 pp |
|
| 50 |
-
| residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, drop near-miss+no-op composite |
|
| 51 |
| residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, component no-op bonus 0.03 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.98% | 0.413 | +5.62 pp |
|
| 52 |
| residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.02 | 3 | 35.54% +/- 1.06 | +/- 2.64 | 57.03% | 0.408 | +5.80 pp |
|
| 53 |
| residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.05 | 3 | 35.48% +/- 0.97 | +/- 2.41 | 56.94% | 0.405 | +5.74 pp |
|
|
@@ -90,8 +90,8 @@ Generated: `2026-06-29T16:57:54+00:00`
|
|
| 90 |
|
| 91 |
| comparison | seeds | mean delta | 95% CI | seed deltas |
|
| 92 |
|---|---:|---:|---:|---|
|
| 93 |
-
| best_clean - canonical_h16 | 3 | +5.
|
| 94 |
-
| best_clean - direct_same_ckpt | 3 | +7.
|
| 95 |
| no_expert_lattice - canonical_h16 | 3 | +27.25 pp | +/- 8.58 | 0:+23.30, 1:+28.70, 2:+29.74 |
|
| 96 |
| full_lattice - no_expert_lattice | 3 | +12.35 pp | +/- 2.63 | 0:+13.57, 1:+11.83, 2:+11.65 |
|
| 97 |
| policy_candidate_lattice - no_expert_lattice | 3 | -16.29 pp | +/- 7.55 | 0:-15.48, 1:-13.74, 2:-19.65 |
|
|
@@ -100,7 +100,7 @@ Generated: `2026-06-29T16:57:54+00:00`
|
|
| 100 |
|
| 101 |
| task | h16 policy | best clean | near-miss lattice | no-expert lattice | full lattice | clean-h16 delta | noexpert-clean gap |
|
| 102 |
|---|---:|---:|---:|---:|---:|---:|---:|
|
| 103 |
-
| LiftPegUpright-v1 | 23.06% | 26.
|
| 104 |
| PickCube-v1 | 20.59% | 32.78% | 58.13% | 62.15% | 84.19% | +12.20 pp | +29.37 pp |
|
| 105 |
| PullCube-v1 | 19.79% | 20.47% | 15.02% | 19.85% | 22.41% | +0.69 pp | -0.63 pp |
|
| 106 |
| PushCube-v1 | 75.06% | 75.74% | 82.54% | 80.10% | 81.92% | +0.68 pp | +4.36 pp |
|
|
@@ -108,9 +108,9 @@ Generated: `2026-06-29T16:57:54+00:00`
|
|
| 108 |
|
| 109 |
## Mechanism Gap
|
| 110 |
|
| 111 |
-
- Best clean residual transport improves over canonical h16 by +5.
|
| 112 |
- Same-state no-expert lattice improves over canonical h16 by +27.25 pp.
|
| 113 |
-
- Remaining clean-to-same-state proposal gap is +21.
|
| 114 |
- Full lattice adds expert proposals and reaches 69.33%, a +12.35 pp gain over no-expert.
|
| 115 |
|
| 116 |
## Selection Histograms
|
|
@@ -119,8 +119,8 @@ Generated: `2026-06-29T16:57:54+00:00`
|
|
| 119 |
- `same_state_no_expert`: lattice_near_miss=1263 (73.2%), lattice_no_op=222 (12.9%), lattice_random_negative=144 (8.3%), lattice_wrong_gripper=62 (3.6%), lattice_wrong_direction=34 (2.0%)
|
| 120 |
- `same_state_policy_baseline`: policy_continuous=1022 (59.2%), lattice_near_miss=448 (26.0%), lattice_no_op=119 (6.9%), lattice_random_negative=75 (4.3%), lattice_wrong_gripper=45 (2.6%), lattice_wrong_direction=16 (0.9%)
|
| 121 |
- `same_state_full`: lattice_expert=977 (56.6%), lattice_near_miss=348 (20.2%), lattice_no_op=177 (10.3%), lattice_random_negative=138 (8.0%), lattice_wrong_gripper=55 (3.2%), lattice_wrong_direction=30 (1.7%)
|
| 122 |
-
- `
|
| 123 |
-
- `
|
| 124 |
|
| 125 |
## Selected-Type Outcomes
|
| 126 |
|
|
|
|
| 1 |
# Paper Analysis
|
| 2 |
|
| 3 |
+
Generated: `2026-06-29T17:02:31+00:00`
|
| 4 |
|
| 5 |
## Main Seed Statistics
|
| 6 |
|
|
|
|
| 47 |
| residual_k4_compose_grid035040045_noopbonus003 | K4 composed type-consensus tangents, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 34.14% +/- 1.58 | +/- 3.92 | 56.00% | 0.482 | +4.41 pp |
|
| 48 |
| residual_k4_composemasked_grid035040045 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45 | 3 | 35.30% +/- 1.22 | +/- 3.02 | 56.91% | 0.410 | +5.57 pp |
|
| 49 |
| residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 35.54% +/- 1.02 | +/- 2.53 | 57.02% | 0.411 | +5.80 pp |
|
| 50 |
+
| residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, drop near-miss+no-op composite | 3 | 35.59% +/- 0.99 | +/- 2.46 | 57.07% | 0.406 | +5.86 pp |
|
| 51 |
| residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, component no-op bonus 0.03 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.98% | 0.413 | +5.62 pp |
|
| 52 |
| residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.02 | 3 | 35.54% +/- 1.06 | +/- 2.64 | 57.03% | 0.408 | +5.80 pp |
|
| 53 |
| residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.05 | 3 | 35.48% +/- 0.97 | +/- 2.41 | 56.94% | 0.405 | +5.74 pp |
|
|
|
|
| 90 |
|
| 91 |
| comparison | seeds | mean delta | 95% CI | seed deltas |
|
| 92 |
|---|---:|---:|---:|---|
|
| 93 |
+
| best_clean - canonical_h16 | 3 | +5.86 pp | +/- 4.37 | 0:+6.96, 1:+3.83, 2:+6.78 |
|
| 94 |
+
| best_clean - direct_same_ckpt | 3 | +7.30 pp | +/- 4.12 | 0:+7.48, 1:+5.57, 2:+8.87 |
|
| 95 |
| no_expert_lattice - canonical_h16 | 3 | +27.25 pp | +/- 8.58 | 0:+23.30, 1:+28.70, 2:+29.74 |
|
| 96 |
| full_lattice - no_expert_lattice | 3 | +12.35 pp | +/- 2.63 | 0:+13.57, 1:+11.83, 2:+11.65 |
|
| 97 |
| policy_candidate_lattice - no_expert_lattice | 3 | -16.29 pp | +/- 7.55 | 0:-15.48, 1:-13.74, 2:-19.65 |
|
|
|
|
| 100 |
|
| 101 |
| task | h16 policy | best clean | near-miss lattice | no-expert lattice | full lattice | clean-h16 delta | noexpert-clean gap |
|
| 102 |
|---|---:|---:|---:|---:|---:|---:|---:|
|
| 103 |
+
| LiftPegUpright-v1 | 23.06% | 26.34% | 62.90% | 61.72% | 73.19% | +3.28 pp | +35.37 pp |
|
| 104 |
| PickCube-v1 | 20.59% | 32.78% | 58.13% | 62.15% | 84.19% | +12.20 pp | +29.37 pp |
|
| 105 |
| PullCube-v1 | 19.79% | 20.47% | 15.02% | 19.85% | 22.41% | +0.69 pp | -0.63 pp |
|
| 106 |
| PushCube-v1 | 75.06% | 75.74% | 82.54% | 80.10% | 81.92% | +0.68 pp | +4.36 pp |
|
|
|
|
| 108 |
|
| 109 |
## Mechanism Gap
|
| 110 |
|
| 111 |
+
- Best clean residual transport improves over canonical h16 by +5.86 pp.
|
| 112 |
- Same-state no-expert lattice improves over canonical h16 by +27.25 pp.
|
| 113 |
+
- Remaining clean-to-same-state proposal gap is +21.39 pp.
|
| 114 |
- Full lattice adds expert proposals and reaches 69.33%, a +12.35 pp gain over no-expert.
|
| 115 |
|
| 116 |
## Selection Histograms
|
|
|
|
| 119 |
- `same_state_no_expert`: lattice_near_miss=1263 (73.2%), lattice_no_op=222 (12.9%), lattice_random_negative=144 (8.3%), lattice_wrong_gripper=62 (3.6%), lattice_wrong_direction=34 (2.0%)
|
| 120 |
- `same_state_policy_baseline`: policy_continuous=1022 (59.2%), lattice_near_miss=448 (26.0%), lattice_no_op=119 (6.9%), lattice_random_negative=75 (4.3%), lattice_wrong_gripper=45 (2.6%), lattice_wrong_direction=16 (0.9%)
|
| 121 |
- `same_state_full`: lattice_expert=977 (56.6%), lattice_near_miss=348 (20.2%), lattice_no_op=177 (10.3%), lattice_random_negative=138 (8.0%), lattice_wrong_gripper=55 (3.2%), lattice_wrong_direction=30 (1.7%)
|
| 122 |
+
- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003`: retrieval_residual_policy_residual=1613 (93.5%), retrieval_residual_residual_no_op=56 (3.2%), retrieval_residual_residual_no_op+residual_wrong_gripper=24 (1.4%), retrieval_residual_residual_wrong_gripper=19 (1.1%), retrieval_residual_residual_near_miss=8 (0.5%), retrieval_residual_residual_near_miss+residual_wrong_gripper=5 (0.3%)
|
| 123 |
+
- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003` residual scale counts: {'0.35': 1631, '0.4': 11, '0.45': 83}
|
| 124 |
|
| 125 |
## Selected-Type Outcomes
|
| 126 |
|
results/paper_table_status.json
CHANGED
|
@@ -1306,14 +1306,14 @@
|
|
| 1306 |
"story_role": "exact incompatibility mask for one weak composed tangent pair",
|
| 1307 |
"fallback_success": null,
|
| 1308 |
"pending_job": "14915009/14915013",
|
| 1309 |
-
"path_exists":
|
| 1310 |
-
"status": "
|
| 1311 |
-
"success":
|
| 1312 |
-
"std_success":
|
| 1313 |
"completed_seeds": null,
|
| 1314 |
-
"num_completed":
|
| 1315 |
"best_config": null,
|
| 1316 |
-
"gain_vs_h16_policy":
|
| 1317 |
},
|
| 1318 |
{
|
| 1319 |
"key": "retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
|
|
@@ -2419,23 +2419,23 @@
|
|
| 2419 |
}
|
| 2420 |
],
|
| 2421 |
"best_clean": {
|
| 2422 |
-
"key": "
|
| 2423 |
-
"label": "K4 composed type-consensus residual retrieval, masked,
|
| 2424 |
-
"path": "
|
| 2425 |
"clean_deployment": "yes",
|
| 2426 |
"same_state_proposals": "no",
|
| 2427 |
"expert_proposal": "no",
|
| 2428 |
-
"story_role": "
|
| 2429 |
"fallback_success": null,
|
| 2430 |
-
"pending_job": "
|
| 2431 |
"path_exists": true,
|
| 2432 |
"status": "complete",
|
| 2433 |
-
"success": 0.
|
| 2434 |
-
"std_success": 0.
|
| 2435 |
"completed_seeds": null,
|
| 2436 |
"num_completed": 3,
|
| 2437 |
"best_config": null,
|
| 2438 |
-
"gain_vs_h16_policy": 0.
|
| 2439 |
},
|
| 2440 |
"best_mechanism_no_expert": {
|
| 2441 |
"key": "no_expert_lattice",
|
|
@@ -2460,7 +2460,7 @@
|
|
| 2460 |
"Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.",
|
| 2461 |
"Use full lattice only as an upper result because it includes expert proposals.",
|
| 2462 |
"Do not claim external SOTA from this table alone; add current external baselines separately.",
|
| 2463 |
-
"Current best clean deployment row is K4 composed type-consensus residual retrieval, masked,
|
| 2464 |
"Trust-region field optimization should be framed as a negative/diagnostic ablation.",
|
| 2465 |
"Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
|
| 2466 |
"KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
|
|
|
|
| 1306 |
"story_role": "exact incompatibility mask for one weak composed tangent pair",
|
| 1307 |
"fallback_success": null,
|
| 1308 |
"pending_job": "14915009/14915013",
|
| 1309 |
+
"path_exists": true,
|
| 1310 |
+
"status": "complete",
|
| 1311 |
+
"success": 0.35594202898550725,
|
| 1312 |
+
"std_success": 0.009889114266221456,
|
| 1313 |
"completed_seeds": null,
|
| 1314 |
+
"num_completed": 3,
|
| 1315 |
"best_config": null,
|
| 1316 |
+
"gain_vs_h16_policy": 0.058550724637681184
|
| 1317 |
},
|
| 1318 |
{
|
| 1319 |
"key": "retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
|
|
|
|
| 2419 |
}
|
| 2420 |
],
|
| 2421 |
"best_clean": {
|
| 2422 |
+
"key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003",
|
| 2423 |
+
"label": "K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite",
|
| 2424 |
+
"path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
|
| 2425 |
"clean_deployment": "yes",
|
| 2426 |
"same_state_proposals": "no",
|
| 2427 |
"expert_proposal": "no",
|
| 2428 |
+
"story_role": "exact incompatibility mask for one weak composed tangent pair",
|
| 2429 |
"fallback_success": null,
|
| 2430 |
+
"pending_job": "14915009/14915013",
|
| 2431 |
"path_exists": true,
|
| 2432 |
"status": "complete",
|
| 2433 |
+
"success": 0.35594202898550725,
|
| 2434 |
+
"std_success": 0.009889114266221456,
|
| 2435 |
"completed_seeds": null,
|
| 2436 |
"num_completed": 3,
|
| 2437 |
"best_config": null,
|
| 2438 |
+
"gain_vs_h16_policy": 0.058550724637681184
|
| 2439 |
},
|
| 2440 |
"best_mechanism_no_expert": {
|
| 2441 |
"key": "no_expert_lattice",
|
|
|
|
| 2460 |
"Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.",
|
| 2461 |
"Use full lattice only as an upper result because it includes expert proposals.",
|
| 2462 |
"Do not claim external SOTA from this table alone; add current external baselines separately.",
|
| 2463 |
+
"Current best clean deployment row is K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite at 35.59%.",
|
| 2464 |
"Trust-region field optimization should be framed as a negative/diagnostic ablation.",
|
| 2465 |
"Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
|
| 2466 |
"KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
|
results/paper_table_status.md
CHANGED
|
@@ -71,7 +71,7 @@ Baseline h=16 policy: 29.74%
|
|
| 71 |
| retrieval_residual_k4_compose_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 34.14% | +4.41 pp | yes | no | no | local tangent composition on the current best typed prior |
|
| 72 |
| retrieval_residual_k4_composemasked_grid035040045 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20 | complete | 35.30% | +5.57 pp | yes | no | no | local tangent composition with anti-goal composite masks |
|
| 73 |
| retrieval_residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 35.54% | +5.80 pp | yes | no | no | local tangent composition with anti-goal composite masks on the current best typed prior |
|
| 74 |
-
| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite |
|
| 75 |
| retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03 | complete | 35.36% | +5.62 pp | yes | no | no | component-wise sparse prior on the masked local tangent composition chart |
|
| 76 |
| retrieval_residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.02 | complete | 35.54% | +5.80 pp | yes | no | no | trust-radius penalty on composed local tangent candidates |
|
| 77 |
| retrieval_residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.05 | complete | 35.48% | +5.74 pp | yes | no | no | stronger trust-radius penalty on composed local tangent candidates |
|
|
@@ -136,7 +136,7 @@ Baseline h=16 policy: 29.74%
|
|
| 136 |
- Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.
|
| 137 |
- Use full lattice only as an upper result because it includes expert proposals.
|
| 138 |
- Do not claim external SOTA from this table alone; add current external baselines separately.
|
| 139 |
-
- Current best clean deployment row is K4 composed type-consensus residual retrieval, masked,
|
| 140 |
- Trust-region field optimization should be framed as a negative/diagnostic ablation.
|
| 141 |
- Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.
|
| 142 |
- KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.
|
|
|
|
| 71 |
| retrieval_residual_k4_compose_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 34.14% | +4.41 pp | yes | no | no | local tangent composition on the current best typed prior |
|
| 72 |
| retrieval_residual_k4_composemasked_grid035040045 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20 | complete | 35.30% | +5.57 pp | yes | no | no | local tangent composition with anti-goal composite masks |
|
| 73 |
| retrieval_residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 35.54% | +5.80 pp | yes | no | no | local tangent composition with anti-goal composite masks on the current best typed prior |
|
| 74 |
+
| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite | complete | 35.59% | +5.86 pp | yes | no | no | exact incompatibility mask for one weak composed tangent pair |
|
| 75 |
| retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03 | complete | 35.36% | +5.62 pp | yes | no | no | component-wise sparse prior on the masked local tangent composition chart |
|
| 76 |
| retrieval_residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.02 | complete | 35.54% | +5.80 pp | yes | no | no | trust-radius penalty on composed local tangent candidates |
|
| 77 |
| retrieval_residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.05 | complete | 35.48% | +5.74 pp | yes | no | no | stronger trust-radius penalty on composed local tangent candidates |
|
|
|
|
| 136 |
- Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.
|
| 137 |
- Use full lattice only as an upper result because it includes expert proposals.
|
| 138 |
- Do not claim external SOTA from this table alone; add current external baselines separately.
|
| 139 |
+
- Current best clean deployment row is K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite at 35.59%.
|
| 140 |
- Trust-region field optimization should be framed as a negative/diagnostic ablation.
|
| 141 |
- Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.
|
| 142 |
- KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.
|
scripts/build_paper_analysis.py
CHANGED
|
@@ -361,6 +361,14 @@ METHODS = [
|
|
| 361 |
"k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
|
| 362 |
),
|
| 363 |
),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 364 |
MethodSpec(
|
| 365 |
key="residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
|
| 366 |
label="K4 composed type-consensus tangents, masked, component no-op bonus 0.03",
|
|
|
|
| 361 |
"k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
|
| 362 |
),
|
| 363 |
),
|
| 364 |
+
MethodSpec(
|
| 365 |
+
key="residual_k4_composemasked_dropnmnoop_l2comp002_grid035040045_noopbonus003",
|
| 366 |
+
label="K4 composed type-consensus tangents, masked, drop near-miss+no-op composite, composite L2 penalty 0.02",
|
| 367 |
+
summary_path=(
|
| 368 |
+
"h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
|
| 369 |
+
"k4_composemasked_dropnmnoop_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
|
| 370 |
+
),
|
| 371 |
+
),
|
| 372 |
MethodSpec(
|
| 373 |
key="residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
|
| 374 |
label="K4 composed type-consensus tangents, masked, component no-op bonus 0.03",
|
scripts/build_paper_table_status.py
CHANGED
|
@@ -705,6 +705,15 @@ SPECS = [
|
|
| 705 |
story_role="exact incompatibility mask for one weak composed tangent pair",
|
| 706 |
pending_job="14915009/14915013",
|
| 707 |
),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 708 |
ResultSpec(
|
| 709 |
key="retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
|
| 710 |
label="K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03",
|
|
|
|
| 705 |
story_role="exact incompatibility mask for one weak composed tangent pair",
|
| 706 |
pending_job="14915009/14915013",
|
| 707 |
),
|
| 708 |
+
ResultSpec(
|
| 709 |
+
key="retrieval_residual_k4_composemasked_dropnmnoop_l2comp002_grid035040045_noopbonus003",
|
| 710 |
+
label="K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite, composite L2 penalty 0.02",
|
| 711 |
+
path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
|
| 712 |
+
clean_deployment="yes",
|
| 713 |
+
same_state_proposals="no",
|
| 714 |
+
expert_proposal="no",
|
| 715 |
+
story_role="compatible local tangent composition with trust-radius regularization",
|
| 716 |
+
),
|
| 717 |
ResultSpec(
|
| 718 |
key="retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
|
| 719 |
label="K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03",
|