thanhkt commited on
Commit
3ca327b
·
verified ·
1 Parent(s): d115fd8

Upload results.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. results.md +105 -0
results.md ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Plan-Bind D1b results: stack bowls, several training seeds
2
+
3
+ Source: `experiments/261029-plan-bind-d1b-seeds/outputs/full/eval`. Training seeds 1000, 1001, 1002. Success in percent: the mean over the seeds, then each seed's success with its [95% Wilson interval].
4
+
5
+ ## Main comparison (last checkpoint)
6
+
7
+ | Policy | Seeds | 0 ms, latency | 0 ms, naive | 200 ms, latency | 300 ms, latency |
8
+ |---|---|---|---|---|---|
9
+ | bsp10 | 3 | 100 (100 [96-100], 100 [96-100], 100 [96-100]) | - | - | - |
10
+ | bsp10rt | 3 | 89 (90 [83-94], 85 [77-91], 91 [84-95]) | - | 58 (61 [51-70], 49 [39-59], 65 [55-74]) | 45 (43 [34-53], 48 [38-58], 45 [36-55]) |
11
+ | onestep | 3 | 71 (59 [49-68], 84 [76-90], 69 [59-77]) | 33 (34 [25-44], 32 [24-42], 34 [25-44]) | 68 (64 [54-73], 76 [67-83], 63 [53-72]) | 80 (82 [73-88], 82 [73-88], 76 [67-83]) |
12
+ | onestep_k2 | 1 | 43 [34-53] | - | - | - |
13
+ | onestep_long | 1 | 72 [63-80] | - | - | - |
14
+ | planbind | 3 | 96 (98 [93-99], 96 [90-98], 94 [88-97]) | 92 (93 [86-97], 92 [85-96], 92 [85-96]) | 98 (97 [92-99], 100 [96-100], 98 [93-99]) | 98 (99 [95-100], 97 [92-99], 98 [93-99]) |
15
+
16
+ ## Plan-Bind minus a control (same training seed, same eval episodes)
17
+
18
+ Per seed: the difference in percentage points, the episodes solved only by Plan-Bind / only by the control, and the exact McNemar p on them. The `mean` row averages the per-seed differences and counts the seeds where Plan-Bind is ahead. `(unpaired)` marks cells without per-episode outcomes.
19
+
20
+ | Control | Setting | Seed | Difference | Only Plan-Bind / only control | McNemar p |
21
+ |---|---|---|---|---|---|
22
+ | onestep | 0 ms, latency | 1000 | +39 | 41 / 2 | <0.0001 |
23
+ | onestep | 0 ms, latency | 1001 | +12 | 14 / 2 | 0.0042 |
24
+ | onestep | 0 ms, latency | 1002 | +25 | 28 / 3 | <0.0001 |
25
+ | onestep | 0 ms, latency | mean | +25.3 | ahead on 3 of 3 seeds | - |
26
+ | onestep | 0 ms, naive | 1000 | +59 | 61 / 2 | <0.0001 |
27
+ | onestep | 0 ms, naive | 1001 | +60 | 62 / 2 | <0.0001 |
28
+ | onestep | 0 ms, naive | 1002 | +58 | 60 / 2 | <0.0001 |
29
+ | onestep | 0 ms, naive | mean | +59.0 | ahead on 3 of 3 seeds | - |
30
+ | onestep | 200 ms, latency | 1000 | +33 | 35 / 2 | <0.0001 |
31
+ | onestep | 200 ms, latency | 1001 | +24 | 24 / 0 | <0.0001 |
32
+ | onestep | 200 ms, latency | 1002 | +35 | 37 / 2 | <0.0001 |
33
+ | onestep | 200 ms, latency | mean | +30.7 | ahead on 3 of 3 seeds | - |
34
+ | onestep | 300 ms, latency | 1000 | +17 | 18 / 1 | <0.0001 |
35
+ | onestep | 300 ms, latency | 1001 | +15 | 18 / 3 | 0.0015 |
36
+ | onestep | 300 ms, latency | 1002 | +22 | 24 / 2 | <0.0001 |
37
+ | onestep | 300 ms, latency | mean | +18.0 | ahead on 3 of 3 seeds | - |
38
+ | bsp10rt | 0 ms, latency | 1000 | +8 | 10 / 2 | 0.04 |
39
+ | bsp10rt | 0 ms, latency | 1001 | +11 | 15 / 4 | 0.02 |
40
+ | bsp10rt | 0 ms, latency | 1002 | +3 | 8 / 5 | 0.58 |
41
+ | bsp10rt | 0 ms, latency | mean | +7.3 | ahead on 3 of 3 seeds | - |
42
+ | bsp10rt | 200 ms, latency | 1000 | +36 | 38 / 2 | <0.0001 |
43
+ | bsp10rt | 200 ms, latency | 1001 | +51 | 51 / 0 | <0.0001 |
44
+ | bsp10rt | 200 ms, latency | 1002 | +33 | 35 / 2 | <0.0001 |
45
+ | bsp10rt | 200 ms, latency | mean | +40.0 | ahead on 3 of 3 seeds | - |
46
+ | bsp10rt | 300 ms, latency | 1000 | +56 | 56 / 0 | <0.0001 |
47
+ | bsp10rt | 300 ms, latency | 1001 | +49 | 51 / 2 | <0.0001 |
48
+ | bsp10rt | 300 ms, latency | 1002 | +53 | 53 / 0 | <0.0001 |
49
+ | bsp10rt | 300 ms, latency | mean | +52.7 | ahead on 3 of 3 seeds | - |
50
+
51
+ ## Is the one-step control strong enough? (commit latency)
52
+
53
+ `onestep_pos` feeds the UNet a trained ramp instead of zeros, so its input carries sequence position. `onestep_k2` calls twice as often; the `every 4` column runs it at D1's call rate. `onestep_long` trains twice as long.
54
+
55
+ | Policy | 0 ms | 0 ms, every 4 frames | 200 ms |
56
+ |---|---|---|---|
57
+ | onestep | 71 (59 [49-68], 84 [76-90], 69 [59-77]) | - | 68 (64 [54-73], 76 [67-83], 63 [53-72]) |
58
+ | onestep_k2 | 43 [34-53] | 44 [35-54] | - |
59
+ | onestep_long | 72 [63-80] | - | - |
60
+ | planbind | 96 (98 [93-99], 96 [90-98], 94 [88-97]) | - | 98 (97 [92-99], 100 [96-100], 98 [93-99]) |
61
+
62
+ ## All cells
63
+
64
+ | Policy | Latency (frames) | Commit | Every | Seed | Checkpoint | Episodes | Success |
65
+ |---|---|---|---|---|---|---|---|
66
+ | bsp10 | 0 | - | - | 1000 | last | 100 | 100 |
67
+ | bsp10 | 0 | - | - | 1001 | last | 100 | 100 |
68
+ | bsp10 | 0 | - | - | 1002 | last | 100 | 100 |
69
+ | bsp10rt | 0 | latency | - | 1000 | last | 100 | 90 |
70
+ | bsp10rt | 0 | latency | - | 1001 | last | 100 | 85 |
71
+ | bsp10rt | 0 | latency | - | 1002 | last | 100 | 91 |
72
+ | bsp10rt | 6 | latency | - | 1000 | last | 100 | 61 |
73
+ | bsp10rt | 6 | latency | - | 1001 | last | 100 | 49 |
74
+ | bsp10rt | 6 | latency | - | 1002 | last | 100 | 65 |
75
+ | bsp10rt | 9 | latency | - | 1000 | last | 100 | 43 |
76
+ | bsp10rt | 9 | latency | - | 1001 | last | 100 | 48 |
77
+ | bsp10rt | 9 | latency | - | 1002 | last | 100 | 45 |
78
+ | onestep | 0 | latency | - | 1000 | last | 100 | 59 |
79
+ | onestep | 0 | latency | - | 1001 | last | 100 | 84 |
80
+ | onestep | 0 | latency | - | 1002 | last | 100 | 69 |
81
+ | onestep | 0 | none | - | 1000 | last | 100 | 34 |
82
+ | onestep | 0 | none | - | 1001 | last | 100 | 32 |
83
+ | onestep | 0 | none | - | 1002 | last | 100 | 34 |
84
+ | onestep | 6 | latency | - | 1000 | last | 100 | 64 |
85
+ | onestep | 6 | latency | - | 1001 | last | 100 | 76 |
86
+ | onestep | 6 | latency | - | 1002 | last | 100 | 63 |
87
+ | onestep | 9 | latency | - | 1000 | last | 100 | 82 |
88
+ | onestep | 9 | latency | - | 1001 | last | 100 | 82 |
89
+ | onestep | 9 | latency | - | 1002 | last | 100 | 76 |
90
+ | onestep_k2 | 0 | latency | - | 1000 | last | 100 | 43 |
91
+ | onestep_k2 | 0 | latency | 4 | 1000 | last | 100 | 44 |
92
+ | onestep_long | 0 | latency | - | 1000 | last | 100 | 72 |
93
+ | planbind | 0 | latency | - | 1000 | last | 100 | 98 |
94
+ | planbind | 0 | latency | - | 1001 | last | 100 | 96 |
95
+ | planbind | 0 | latency | - | 1002 | last | 100 | 94 |
96
+ | planbind | 0 | none | - | 1000 | last | 100 | 93 |
97
+ | planbind | 0 | none | - | 1001 | last | 100 | 92 |
98
+ | planbind | 0 | none | - | 1002 | last | 100 | 92 |
99
+ | planbind | 6 | latency | - | 1000 | last | 100 | 97 |
100
+ | planbind | 6 | latency | - | 1001 | last | 100 | 100 |
101
+ | planbind | 6 | latency | - | 1002 | last | 100 | 98 |
102
+ | planbind | 9 | latency | - | 1000 | last | 100 | 99 |
103
+ | planbind | 9 | latency | - | 1001 | last | 100 | 97 |
104
+ | planbind | 9 | latency | - | 1002 | last | 100 | 98 |
105
+