Linslab commited on
Commit
60a282b
·
verified ·
1 Parent(s): cfed8f0

Add small_flow_matching_policy/ (RADIO + Flow Matching checkpoints from Linslab/ManiLadder_Models)

Browse files

All 111 task folders of Linslab/ManiLadder_Models (revision e4f231e8e836f29720b6101ca05ae357964f54b0) now live under small_flow_matching_policy/<task>/. File contents are unchanged: 220 LFS files were re-uploaded and checked by SHA-256, 215 small files by Git blob hash.

This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +107 -0
  2. small_flow_matching_policy/BagDropIn-v1/README.md +55 -0
  3. small_flow_matching_policy/BagDropIn-v1/artifact_manifest.json +287 -0
  4. small_flow_matching_policy/BagDropIn-v1/ckpt_final.pt +3 -0
  5. small_flow_matching_policy/BagDropIn-v1/environment/bagdrop.py +1588 -0
  6. small_flow_matching_policy/BagDropIn-v1/eval_metrics.json +475 -0
  7. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200000_succ0.mp4 +3 -0
  8. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200001_succ0.mp4 +3 -0
  9. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200002_succ0.mp4 +3 -0
  10. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200003_succ0.mp4 +3 -0
  11. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200004_succ0.mp4 +3 -0
  12. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200005_succ0.mp4 +3 -0
  13. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200006_succ0.mp4 +3 -0
  14. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200007_succ0.mp4 +3 -0
  15. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200008_succ0.mp4 +3 -0
  16. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200009_succ0.mp4 +3 -0
  17. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200010_succ0.mp4 +3 -0
  18. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200011_succ0.mp4 +3 -0
  19. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200012_succ0.mp4 +3 -0
  20. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200013_succ1.mp4 +3 -0
  21. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200014_succ0.mp4 +3 -0
  22. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200015_succ0.mp4 +3 -0
  23. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200016_succ0.mp4 +3 -0
  24. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200017_succ0.mp4 +3 -0
  25. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200018_succ0.mp4 +3 -0
  26. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200019_succ0.mp4 +3 -0
  27. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200020_succ0.mp4 +3 -0
  28. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200021_succ0.mp4 +3 -0
  29. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200022_succ0.mp4 +3 -0
  30. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200023_succ0.mp4 +3 -0
  31. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200024_succ0.mp4 +3 -0
  32. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200025_succ0.mp4 +3 -0
  33. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200026_succ0.mp4 +3 -0
  34. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200027_succ0.mp4 +3 -0
  35. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200028_succ1.mp4 +3 -0
  36. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200029_succ0.mp4 +3 -0
  37. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200030_succ0.mp4 +3 -0
  38. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200031_succ0.mp4 +3 -0
  39. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200032_succ0.mp4 +3 -0
  40. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200033_succ1.mp4 +3 -0
  41. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200034_succ0.mp4 +3 -0
  42. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200035_succ0.mp4 +3 -0
  43. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200036_succ0.mp4 +3 -0
  44. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200037_succ1.mp4 +3 -0
  45. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200038_succ0.mp4 +3 -0
  46. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200039_succ0.mp4 +3 -0
  47. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200040_succ0.mp4 +3 -0
  48. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200041_succ0.mp4 +3 -0
  49. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200042_succ0.mp4 +3 -0
  50. small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200043_succ0.mp4 +3 -0
.gitattributes CHANGED
@@ -33,3 +33,110 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200000_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
37
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200001_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
38
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200002_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
39
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200003_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
40
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200004_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
41
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200005_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
42
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200006_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
43
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200007_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
44
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200008_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
45
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200009_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
46
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200010_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
47
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200011_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
48
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200012_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
49
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200013_succ1.mp4 filter=lfs diff=lfs merge=lfs -text
50
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200014_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
51
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200015_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
52
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200016_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
53
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200017_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
54
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200018_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
55
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200019_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
56
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200020_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
57
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200021_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
58
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200022_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
59
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200023_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
60
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200024_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
61
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200025_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
62
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200026_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
63
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200027_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
64
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200028_succ1.mp4 filter=lfs diff=lfs merge=lfs -text
65
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200029_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
66
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200030_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
67
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200031_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
68
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200032_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
69
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200033_succ1.mp4 filter=lfs diff=lfs merge=lfs -text
70
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200034_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
71
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200035_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
72
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200036_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
73
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200037_succ1.mp4 filter=lfs diff=lfs merge=lfs -text
74
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200038_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
75
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200039_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
76
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200040_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
77
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200041_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
78
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200042_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
79
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200043_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
80
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200044_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
81
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200045_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
82
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200046_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
83
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200047_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
84
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200048_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
85
+ small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200049_succ0.mp4 filter=lfs diff=lfs merge=lfs -text
86
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_000.mp4 filter=lfs diff=lfs merge=lfs -text
87
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_001.mp4 filter=lfs diff=lfs merge=lfs -text
88
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_002.mp4 filter=lfs diff=lfs merge=lfs -text
89
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_003.mp4 filter=lfs diff=lfs merge=lfs -text
90
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_004.mp4 filter=lfs diff=lfs merge=lfs -text
91
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_005.mp4 filter=lfs diff=lfs merge=lfs -text
92
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_006.mp4 filter=lfs diff=lfs merge=lfs -text
93
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_007.mp4 filter=lfs diff=lfs merge=lfs -text
94
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_008.mp4 filter=lfs diff=lfs merge=lfs -text
95
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_009.mp4 filter=lfs diff=lfs merge=lfs -text
96
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_010.mp4 filter=lfs diff=lfs merge=lfs -text
97
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_011.mp4 filter=lfs diff=lfs merge=lfs -text
98
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_012.mp4 filter=lfs diff=lfs merge=lfs -text
99
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_013.mp4 filter=lfs diff=lfs merge=lfs -text
100
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_014.mp4 filter=lfs diff=lfs merge=lfs -text
101
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_015.mp4 filter=lfs diff=lfs merge=lfs -text
102
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_016.mp4 filter=lfs diff=lfs merge=lfs -text
103
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_017.mp4 filter=lfs diff=lfs merge=lfs -text
104
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_018.mp4 filter=lfs diff=lfs merge=lfs -text
105
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_019.mp4 filter=lfs diff=lfs merge=lfs -text
106
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_020.mp4 filter=lfs diff=lfs merge=lfs -text
107
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_021.mp4 filter=lfs diff=lfs merge=lfs -text
108
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_022.mp4 filter=lfs diff=lfs merge=lfs -text
109
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_023.mp4 filter=lfs diff=lfs merge=lfs -text
110
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_024.mp4 filter=lfs diff=lfs merge=lfs -text
111
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_025.mp4 filter=lfs diff=lfs merge=lfs -text
112
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_026.mp4 filter=lfs diff=lfs merge=lfs -text
113
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_027.mp4 filter=lfs diff=lfs merge=lfs -text
114
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_028.mp4 filter=lfs diff=lfs merge=lfs -text
115
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_029.mp4 filter=lfs diff=lfs merge=lfs -text
116
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_030.mp4 filter=lfs diff=lfs merge=lfs -text
117
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_031.mp4 filter=lfs diff=lfs merge=lfs -text
118
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_032.mp4 filter=lfs diff=lfs merge=lfs -text
119
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_033.mp4 filter=lfs diff=lfs merge=lfs -text
120
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_034.mp4 filter=lfs diff=lfs merge=lfs -text
121
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_035.mp4 filter=lfs diff=lfs merge=lfs -text
122
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_036.mp4 filter=lfs diff=lfs merge=lfs -text
123
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_037.mp4 filter=lfs diff=lfs merge=lfs -text
124
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_038.mp4 filter=lfs diff=lfs merge=lfs -text
125
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_039.mp4 filter=lfs diff=lfs merge=lfs -text
126
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_040.mp4 filter=lfs diff=lfs merge=lfs -text
127
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_041.mp4 filter=lfs diff=lfs merge=lfs -text
128
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_042.mp4 filter=lfs diff=lfs merge=lfs -text
129
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_043.mp4 filter=lfs diff=lfs merge=lfs -text
130
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_044.mp4 filter=lfs diff=lfs merge=lfs -text
131
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_045.mp4 filter=lfs diff=lfs merge=lfs -text
132
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_046.mp4 filter=lfs diff=lfs merge=lfs -text
133
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_047.mp4 filter=lfs diff=lfs merge=lfs -text
134
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_048.mp4 filter=lfs diff=lfs merge=lfs -text
135
+ small_flow_matching_policy/BagHanging-v1/evaluation50/episode_049.mp4 filter=lfs diff=lfs merge=lfs -text
136
+ small_flow_matching_policy/BagHanging-v1/visualizations/ckpt20k_eval50_front_grid.mp4 filter=lfs diff=lfs merge=lfs -text
137
+ small_flow_matching_policy/BagHanging-v1/visualizations/expert20_front_grid.mp4 filter=lfs diff=lfs merge=lfs -text
138
+ small_flow_matching_policy/SwingRing-v1/evaluation/videos_standard_front/SwingRing-v1_ep000_20261028-114857_succ0_ret66.54.mp4 filter=lfs diff=lfs merge=lfs -text
139
+ small_flow_matching_policy/SwingRing-v1/evaluation/videos_standard_front/SwingRing-v1_ep001_20261028-114954_succ0_ret78.49.mp4 filter=lfs diff=lfs merge=lfs -text
140
+ small_flow_matching_policy/SwingRing-v1/evaluation/videos_standard_front/SwingRing-v1_ep002_20261028-115049_succ0_ret73.53.mp4 filter=lfs diff=lfs merge=lfs -text
141
+ small_flow_matching_policy/SwingRing-v1/evaluation/videos_standard_front/SwingRing-v1_ep003_20261028-115144_succ0_ret65.88.mp4 filter=lfs diff=lfs merge=lfs -text
142
+ small_flow_matching_policy/SwingRing-v1/evaluation/videos_standard_front/SwingRing-v1_ep004_20261028-115238_succ0_ret73.45.mp4 filter=lfs diff=lfs merge=lfs -text
small_flow_matching_policy/BagDropIn-v1/README.md ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # BagDropIn-v1 wall release (2026-11-09)
2
+
3
+ This release replaces the short support pole with a collision wall, preserving the hook pose, scripted policy and success rule. The wall is centered at `(0.30, -0.43, 0.20)` with dimensions `(0.55, 0.02, 0.50)` metres. The hook length remains 0.10 m. Success requires both handles hung and the payload inside for eight consecutive steps; hand clearance is diagnostic only.
4
+
5
+ ## Matched artifacts
6
+
7
+ - Raw: `Linslab/ManiLadder_Dataset`, `raw/BagDropIn-v1/` (20 scripted demonstrations, metadata and videos).
8
+ - Preprocessed: the same repository, `preprocessed/BagDropIn-v1/` (20 trajectories, RADIO features, videos and `norm_stats.json`). This wall release intentionally uses this path even though the dataset root README describes migration of other preprocessed datasets.
9
+ - Model: `Linslab/ManiLadder_Models`, `BagDropIn-v1/ckpt_final.pt`. **This compatibility filename contains the selected step-30000 checkpoint, not step 40000.** Inspect `checkpoint['step']` or `release.json`.
10
+ - Use the matching `BagDropIn-v1/norm_stats.json` from the model repository. Do not mix it with the previous short-pole release.
11
+ - Model task directory includes `eval_metrics.json`, `release.json`, `README.md`, `evaluation_videos/`, and a snapshot of the environment source.
12
+
13
+ Previous short-pole artifacts remain available through Hugging Face commit history. The previous dataset revision was `326d7b3e3001d8cae6c5906458aa320ffa557f1f`; the previous model revision was `02e2f9ecddeab80e6dc86bfe59e08605ccba7f73`.
14
+
15
+ ## Training and selection
16
+
17
+ The architecture is RADIO + transformer Flow Matching, **not pi0.5**. Training used 20 wall demonstrations, 40,000 steps, batch size 256, learning rate 1e-4, weight decay 1e-4, seed 0, action chunk length 32, observation/state horizon 2, and the repository training implementation. Checkpoints were saved every 2,000 steps. The run resumed from 6k and 10k while evaluation scheduling changed; native resume restores model/optimizer/step, not the complete RNG/data-loader state. Exact retraining equivalence is not claimed.
18
+
19
+ | Checkpoint | Development evaluation (20 episodes) | Independent evaluation (50 episodes) |
20
+ | --- | --- | --- |
21
+ | 10k | 0/20 (0%) | — |
22
+ | 20k | 1/20 (5%) | — |
23
+ | 30k (selected) | 4/20 (20%) | **4/50 (8%)** |
24
+ | 40k | 0/20 (0%) | 1/50 (2%) |
25
+
26
+ The selected model is the best among the evaluated 10k/20k/30k/40k checkpoints, not necessarily the best of all 2k saves. Selection used the development sweep; the reported 8% is the subsequent 50-episode result. Small-sample rates vary; this does not guarantee a future run falls in 5–15%.
27
+
28
+ Development seeds were 100000–100019. Both independent 50-episode runs used the same 50 distinct environment seeds 200000–200049. Each episode was a separate stock-evaluator process (default process seed 0); simulation initialization was not given a fixed `MANILADDER_SIM_SEED`. Parallel/sequential trajectory equivalence has not been established. Development results retained completed serial episodes when scheduling was switched to parallel.
29
+
30
+ Evaluation used `sample_steps=20`, `exec_steps=16`, unchanged task success, four-camera grid video, 30 fps, grid padding 4, resize mode `min`. Two PRO6000 GPUs each ran 12 episodes then 13 episodes, independently, through per-job CUDA MPS. Failed launches were excluded, never counted as task failures.
31
+
32
+ ## Evaluate the published checkpoint
33
+
34
+ Download the matched model/statistics and use this VR source revision. On an allocated compute node, a single stock-evaluator run can be launched as follows (its RNG stream differs from the per-episode-process evaluation above):
35
+
36
+ ```bash
37
+ python maniladder/algos/FlowMatching/eval_2d_radio_comb.py \
38
+ --ckpt /path/to/BagDropIn-v1/ckpt_final.pt \
39
+ --stats_json /path/to/BagDropIn-v1/norm_stats.json \
40
+ --env_name BagDropIn-v1 --episodes 50 --episode_seed_base 200000 \
41
+ --sample_steps 20 --exec_steps 16 --save_video \
42
+ --video_dir runs/BagDropIn-v1/wall_release_eval \
43
+ --metrics_dir runs/BagDropIn-v1/wall_release_eval \
44
+ --video_fps 30 --video_grid --video_grid_pad 4 --video_grid_resize min
45
+ ```
46
+
47
+ For the release's per-episode-process layout, run `scripts/eval_bagdrop_wall_release.py` once in each of two separate one-GPU allocations, using `--shard 0` and `--shard 1`, identical checkpoint/statistics paths and a shared new `--output` directory. It requires `nvidia-cuda-mps-control`; it respects Slurm GPU visibility and does not change compute mode. Set `PYOPENGL_PLATFORM=egl` and use an appropriate CPU/thread budget. Each shard saves 25 distinct episodes; the last finisher merges all 50 records. Existing completed batches are preserved, while partial directories require inspection before reuse.
48
+
49
+ ## Artifact checks
50
+
51
+ - Environment SHA256: `887e7ae2ea9188688335a8e2b39b5e3ea3532c47734387a2b82bfe18d500c041`.
52
+ - Normalization SHA256: `4f88d790058d13044db3dac1c7310e5b083e029d9d71fb800666a8ffc8e1e980`.
53
+ - Per-file sizes and hashes, selected checkpoint metadata, code revision, and upload verification are recorded in the release metadata.
54
+
55
+ Published source revision: https://github.com/HeegerGao/maniladder/tree/f736955e2bf698385e4e957d6a5ec60419acd4b7
small_flow_matching_policy/BagDropIn-v1/artifact_manifest.json ADDED
@@ -0,0 +1,287 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200000_succ0.mp4",
4
+ "size": 3264118,
5
+ "sha256": "190783f9b9e7529a0ec46569adf5c4794868f5154cb43b19b83fc61f1e9f7aa0"
6
+ },
7
+ {
8
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200001_succ0.mp4",
9
+ "size": 3783945,
10
+ "sha256": "3a90286ed1713ac357310234065a2696bc5fd6c511883eccd8391752f3338ae9"
11
+ },
12
+ {
13
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200002_succ0.mp4",
14
+ "size": 4493562,
15
+ "sha256": "bd1863802cc69ab556e69623865d192a2cf45b79c6d5f1b2a5c17013e14a5ab2"
16
+ },
17
+ {
18
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200003_succ0.mp4",
19
+ "size": 3600186,
20
+ "sha256": "ece794627193fc7bf1a2fa281b314b383aedd2bd93381d20a9ab191e91af5fde"
21
+ },
22
+ {
23
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200004_succ0.mp4",
24
+ "size": 3429978,
25
+ "sha256": "61baf86a76e4a54fb8695baa0d7e7773646c31814b06147fe665c4d0b7272451"
26
+ },
27
+ {
28
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200005_succ0.mp4",
29
+ "size": 3417354,
30
+ "sha256": "660f7e62ebc020113447d03eb493bab81953191a30f00c7d3899a442ed4373b8"
31
+ },
32
+ {
33
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200006_succ0.mp4",
34
+ "size": 3666759,
35
+ "sha256": "f2eb34f02fdce46b678a468523affa08b8dfb877362d8908e75351cd0c780c72"
36
+ },
37
+ {
38
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200007_succ0.mp4",
39
+ "size": 3238840,
40
+ "sha256": "206a29f5234deaadea078ca7d12d15b797e81fca64977998f31363a69263404e"
41
+ },
42
+ {
43
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200008_succ0.mp4",
44
+ "size": 4125735,
45
+ "sha256": "f0b7c38e53f05b33e89e0ce251774dbc8d41abe86c094e4b9394158c4f7227de"
46
+ },
47
+ {
48
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200009_succ0.mp4",
49
+ "size": 3501552,
50
+ "sha256": "663837c5c85af68467adec0e6466f1a641df85fe4eac48f70f3e588bc12ce9fc"
51
+ },
52
+ {
53
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200010_succ0.mp4",
54
+ "size": 4268729,
55
+ "sha256": "645a49b26ee5717eb284f18ca605fb3d2a2c72164574465a75945161fbc9f389"
56
+ },
57
+ {
58
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200011_succ0.mp4",
59
+ "size": 3456109,
60
+ "sha256": "4740eafc38447d916727975ae8c9a17de2330e4f9e5e85e82e266fa5e06ca9fd"
61
+ },
62
+ {
63
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200012_succ0.mp4",
64
+ "size": 3614493,
65
+ "sha256": "2a2989d3d70a929dca37c035784ef28834676e7c27cdcf2f95e2f20fab80d818"
66
+ },
67
+ {
68
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200013_succ1.mp4",
69
+ "size": 2258291,
70
+ "sha256": "15b7c06e0ffcc718ddfffe4097c66c279fac605deb0b215583750feae1a770d6"
71
+ },
72
+ {
73
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200014_succ0.mp4",
74
+ "size": 3299803,
75
+ "sha256": "d6cfddf9e228f7bd136322deb99ca5b52ae91a499b870e27cf4c38d2ef7b7328"
76
+ },
77
+ {
78
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200015_succ0.mp4",
79
+ "size": 4309718,
80
+ "sha256": "0fde98f2d9172ee1918823afb070fd9d2282eceb08cdc0c3e73bebfd2408ea87"
81
+ },
82
+ {
83
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200016_succ0.mp4",
84
+ "size": 3641302,
85
+ "sha256": "e0d2fbc5be96b6e3474bc1d31d50215003537780f25005134abd351e82eb7ccd"
86
+ },
87
+ {
88
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200017_succ0.mp4",
89
+ "size": 2983719,
90
+ "sha256": "c929ceb60173a2315049bcd184955553abe9a99befd6d37ab9be9438893c824f"
91
+ },
92
+ {
93
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200018_succ0.mp4",
94
+ "size": 3848147,
95
+ "sha256": "9a8a656062f53e07b28be976a2edb44274dfaca9137e01048565d057b3b4efb2"
96
+ },
97
+ {
98
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200019_succ0.mp4",
99
+ "size": 4243200,
100
+ "sha256": "bb9c68fff7f6ba5146132e1579f444ee20e793a6a6da7e68fbf70dcb098ca7c2"
101
+ },
102
+ {
103
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200020_succ0.mp4",
104
+ "size": 3631050,
105
+ "sha256": "4de1380d394fbd1d6aca218f3e1e77f1039384030e1b471fb0a3d8cf6696d190"
106
+ },
107
+ {
108
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200021_succ0.mp4",
109
+ "size": 3495015,
110
+ "sha256": "0c41d8176dfadc4293665b19e09b0d1b1988465e57b29721a6c952b5ddd6c744"
111
+ },
112
+ {
113
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200022_succ0.mp4",
114
+ "size": 3549856,
115
+ "sha256": "d28ba1327a664537c5a5aae848714b7eecfda0aa368c807c961682ca63d56064"
116
+ },
117
+ {
118
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200023_succ0.mp4",
119
+ "size": 2824772,
120
+ "sha256": "860a763eb491c0d552a9cb7c5f2917d93d42d51ac4196c2a5eac20604d8f40f3"
121
+ },
122
+ {
123
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200024_succ0.mp4",
124
+ "size": 3711752,
125
+ "sha256": "4ffa6c3bce52792b6be57b71795970ad1f30cdb71968c2ead1e9ccd79dea93a4"
126
+ },
127
+ {
128
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200025_succ0.mp4",
129
+ "size": 4298241,
130
+ "sha256": "341e9ce00ac78bc7d6c25ed88135e1e067da9199407e0cc82e8263d43360ee7b"
131
+ },
132
+ {
133
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200026_succ0.mp4",
134
+ "size": 2638294,
135
+ "sha256": "2866a6349b64f8b11422acb85d29ea1a8265e3d38671e4c711ef70fd8db8b910"
136
+ },
137
+ {
138
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200027_succ0.mp4",
139
+ "size": 3379445,
140
+ "sha256": "254bd50497173ff71fa20af57f52db138ed27ac116813f00682489b354104afc"
141
+ },
142
+ {
143
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200028_succ1.mp4",
144
+ "size": 2261392,
145
+ "sha256": "c657c63e50b4b877bdb63cf07b0f4de337bfaa9e10952e5ebaffa7e01e675f45"
146
+ },
147
+ {
148
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200029_succ0.mp4",
149
+ "size": 3047494,
150
+ "sha256": "d884d7cf131dc7be5bef4a6f5b6409f7a8cb287a9730a9354249f48a53b04427"
151
+ },
152
+ {
153
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200030_succ0.mp4",
154
+ "size": 3594885,
155
+ "sha256": "3ca0bc1851f367ffbbf1dd041529c626fe4547701b6b521f70b6aa22b66ff58d"
156
+ },
157
+ {
158
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200031_succ0.mp4",
159
+ "size": 3450788,
160
+ "sha256": "b4e72713d78e75527af2a9e79c62b12206006c09fc014297a5920cd6554a636a"
161
+ },
162
+ {
163
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200032_succ0.mp4",
164
+ "size": 2963344,
165
+ "sha256": "f06309657400350e1b85f745fa55e1ebfe9bdca5eb3b7004a5d8af3bc446c66e"
166
+ },
167
+ {
168
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200033_succ1.mp4",
169
+ "size": 2417315,
170
+ "sha256": "db0f81350a443a48093959dee909790dc8220bee2bd266c63675bd83b3a911be"
171
+ },
172
+ {
173
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200034_succ0.mp4",
174
+ "size": 4440753,
175
+ "sha256": "599d3ea8c6fb9f37016b307410c0d6755e74abe6c1a2b6ccc86ef8c708bea08c"
176
+ },
177
+ {
178
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200035_succ0.mp4",
179
+ "size": 2673588,
180
+ "sha256": "7bd482e44ee5b37d4a41ef3b6f09ebeac4990860a739bf94420407cb49b056b6"
181
+ },
182
+ {
183
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200036_succ0.mp4",
184
+ "size": 3850317,
185
+ "sha256": "f5199397a8210caef0b341d90b21c9c941aeb31d522a7257888dba6831cc7f2c"
186
+ },
187
+ {
188
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200037_succ1.mp4",
189
+ "size": 2206454,
190
+ "sha256": "aba0add24e901804a04d0e51472c6468fc76416741c66a7da23c69ade1b63fbf"
191
+ },
192
+ {
193
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200038_succ0.mp4",
194
+ "size": 4294076,
195
+ "sha256": "b22d32c390c0e16129d2e5182fe46dcbdc9f3735f0ac5f51f6685b882daf398f"
196
+ },
197
+ {
198
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200039_succ0.mp4",
199
+ "size": 3590257,
200
+ "sha256": "cd580fc96ce5a546f2671c03f25826f7c94cf44927d3cb124c348de2f84330ee"
201
+ },
202
+ {
203
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200040_succ0.mp4",
204
+ "size": 3556042,
205
+ "sha256": "545e6563c75c9db03b3c69c0123353b12d88d192df462c48fd6885b3c872edad"
206
+ },
207
+ {
208
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200041_succ0.mp4",
209
+ "size": 3519549,
210
+ "sha256": "278d901bc4b1652e14d644b192c780c76ad2511c64ae5985209351c341f180e3"
211
+ },
212
+ {
213
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200042_succ0.mp4",
214
+ "size": 4282318,
215
+ "sha256": "3520fb2c1bdbc6246e6dd1a7421a6f7a4e92a1f6aff2812f96153faea1edc947"
216
+ },
217
+ {
218
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200043_succ0.mp4",
219
+ "size": 4253006,
220
+ "sha256": "6a6c42451a7a71769e9c2dbbb29967651ee1b69ab3123f759d34b6c91fa93a48"
221
+ },
222
+ {
223
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200044_succ0.mp4",
224
+ "size": 3421091,
225
+ "sha256": "4af25101c8a8aceb0faaee77d3245bc7285572e018bd79efbd16092390ca9500"
226
+ },
227
+ {
228
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200045_succ0.mp4",
229
+ "size": 3484266,
230
+ "sha256": "13ea59a78f3d7db7bde49f9f02d7f149e734d9791db63760eb2511f99d33c5d6"
231
+ },
232
+ {
233
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200046_succ0.mp4",
234
+ "size": 4229209,
235
+ "sha256": "2ab408edb8aa1aab70de082f1970b85eff987c7c94c7b91eba4f8d1ada6c29dd"
236
+ },
237
+ {
238
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200047_succ0.mp4",
239
+ "size": 3978613,
240
+ "sha256": "9b5c62e446f0bce5ac2834819abf50eae979770b173a4f53dfe117df44ca06c4"
241
+ },
242
+ {
243
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200048_succ0.mp4",
244
+ "size": 3669884,
245
+ "sha256": "a8699fc3d9b6b5fcfa1d5118963e524cd6db6520ceef81e8a0aeb6283fba2439"
246
+ },
247
+ {
248
+ "remote": "BagDropIn-v1/evaluation_videos/seed_200049_succ0.mp4",
249
+ "size": 4482249,
250
+ "sha256": "be92e45e710dde730fa5237a4764557e7d954b23889bb1cc5ab9f9d33ea3ca65"
251
+ },
252
+ {
253
+ "remote": "BagDropIn-v1/ckpt_final.pt",
254
+ "size": 418818607,
255
+ "sha256": "00619c450c88352b29933f9f64c03d1292639faac921864b44f3e77e53abb4a4"
256
+ },
257
+ {
258
+ "remote": "BagDropIn-v1/norm_stats.json",
259
+ "size": 4442,
260
+ "sha256": "4f88d790058d13044db3dac1c7310e5b083e029d9d71fb800666a8ffc8e1e980"
261
+ },
262
+ {
263
+ "remote": "BagDropIn-v1/environment/bagdrop.py",
264
+ "size": 82072,
265
+ "sha256": "887e7ae2ea9188688335a8e2b39b5e3ea3532c47734387a2b82bfe18d500c041"
266
+ },
267
+ {
268
+ "remote": "BagDropIn-v1/README.md",
269
+ "size": 5185,
270
+ "sha256": "04e3a068d133b3e798360381ff02a2bec0a78b76e02d9dffe832a68981b08ba4"
271
+ },
272
+ {
273
+ "remote": "BagDropIn-v1/release.json",
274
+ "size": 1719,
275
+ "sha256": "0e5f06328781b5389a1ab0f8333f85a1b2925e47b4a86b440364adf77cec34a4"
276
+ },
277
+ {
278
+ "remote": "BagDropIn-v1/eval_metrics.json",
279
+ "size": 10343,
280
+ "sha256": "f1a3da72047d38cfb28ce7502f23b39e97d0e598550dc2d08930da7ae6adac83"
281
+ },
282
+ {
283
+ "remote": "README.md",
284
+ "size": 356,
285
+ "sha256": "610f917020346c13015fa70fc44f9454d45aece11e3854813d28ed59c19bdb57"
286
+ }
287
+ ]
small_flow_matching_policy/BagDropIn-v1/ckpt_final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:00619c450c88352b29933f9f64c03d1292639faac921864b44f3e77e53abb4a4
3
+ size 418818607
small_flow_matching_policy/BagDropIn-v1/environment/bagdrop.py ADDED
@@ -0,0 +1,1588 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+ from __future__ import annotations
3
+
4
+ import os
5
+ from typing import Any, Dict, Optional, Tuple
6
+
7
+ import numpy as np
8
+ import torch
9
+ import genesis as gs
10
+ from genesis.engine.scene import Scene
11
+ from scipy.spatial.transform import Rotation as Rot
12
+
13
+ from maniladder import ASSETS_DIR
14
+ from maniladder.agents.base_agent import Keyframe
15
+ from maniladder.agents.robots.panda.panda_leaphand import PandaLeaphand
16
+ from maniladder.envs.base.base_env import ManiLadderBaseEnv
17
+ from maniladder.utils import Pose
18
+ from maniladder.utils.registration import register_env
19
+
20
+
21
+ BAG_DROPIN_DOC_STRING = """**Task Description (single arm, dex hand):**
22
+ A cloth bag hangs at the wall by ONE handle looped over a hook peg, its
23
+ bottom resting on the table and its mouth gaping open; a payload cube sits on
24
+ a shelf nearby. Pick the payload, drop it into the bag's open mouth, then
25
+ hook the free handle with the thumb and hang it onto the hook as well,
26
+ closing the loaded bag onto the hook. No cloth is ever attached to the hand:
27
+ the thumb hooks the handle the way a person does (curled inward, the handle in
28
+ the crook between thumb and palm) and carries it to the peg tip; only the last
29
+ hand-over from the opened thumb onto the peg is assisted by the env.
30
+
31
+ **Randomizations:**
32
+ - payload spawn XY jitter on the shelf,
33
+ - bag pre-hang lateral jitter along the peg's x.
34
+
35
+ **Success Conditions (sustained for HOLD_STEPS):**
36
+ - BOTH handles are within HANG_THR of the hook hang point (handles are tracked
37
+ by particle indices cached at build time, not re-estimated),
38
+ - the payload is inside the bag body,
39
+ - hand clearance is diagnostic only and is not required.
40
+ """
41
+
42
+ # default rest posture: hand extends -y palm-down; its top-down cage is the
43
+ # grasp proven on HitBox's pendulum block and PickSoft's cylinders, and the
44
+ # -y hand extension matches this scene's -y wall
45
+ _QPOS_DEFAULT = PandaLeaphand.REST_POSTURES["default"].astype(np.float32).copy()
46
+
47
+
48
+ def _quat_xyzw_to_R(q: np.ndarray) -> np.ndarray:
49
+ return Rot.from_quat([q[0], q[1], q[2], q[3]]).as_matrix()
50
+
51
+
52
+ def _quat_wxyz_to_R_t(q: torch.Tensor) -> torch.Tensor:
53
+ w, x, y, z = q[0], q[1], q[2], q[3]
54
+ return torch.stack([
55
+ torch.stack([1 - 2 * (y * y + z * z), 2 * (x * y - w * z), 2 * (x * z + w * y)]),
56
+ torch.stack([2 * (x * y + w * z), 1 - 2 * (x * x + z * z), 2 * (y * z - w * x)]),
57
+ torch.stack([2 * (x * z - w * y), 2 * (y * z + w * x), 1 - 2 * (x * x + y * y)]),
58
+ ])
59
+
60
+
61
+ @register_env("BagDropIn-v1", max_episode_steps=700)
62
+ class BagDropInEnv(ManiLadderBaseEnv):
63
+ """See `BAG_DROPIN_DOC_STRING`."""
64
+
65
+ SUPPORTED_ROBOTS = ["panda_leaphand"]
66
+ _sample_video_link = "TODO: add sample video link here"
67
+
68
+ TABLE_Z = 0.0
69
+
70
+ # ---- bag ----
71
+ # scale 18 -> bag ~0.35 tall, ~0.24 wide: it can hang fully from a 0.38
72
+ # hook without pooling on the table (scale 25 was ~0.5 tall and dragged)
73
+ # euler (90,0,90): bag plane y-z, so the handle loops (which arch across
74
+ # the mouth from front wall to back wall) have their hole axis along y --
75
+ # matching the y-axis peg. Spawn z must keep the WHOLE bag above the table
76
+ # (z extent is +-0.175 at scale 18) or it crumples at build time and the
77
+ # cached handle particle indices become garbage.
78
+ BAG_SPEC: Dict[str, Any] = {
79
+ "file": ASSETS_DIR / "meshes/deformable/bag.obj",
80
+ "pos": (0.20, 0.00, 0.26),
81
+ "euler": (90.0, 0.0, 90.0),
82
+ "scale": 18.0,
83
+ "color": (0.85, 0.62, 0.50, 0.85),
84
+ }
85
+ HANDLE_BAND = 0.04 # top band (m) used to find the two handles at build
86
+ PREHANG_SETTLE_STEPS = 200
87
+ PREHANG_JITTER_X = 0.02
88
+ PREHANG_ARCH_DZ = 0.004 # arch apex above the peg top at teleport
89
+
90
+ # ---- wall + hook ----
91
+ WALL_XZ_CENTER: Tuple[float, float] = (0.30, 0.20)
92
+ WALL_Y: float = -0.43
93
+ WALL_SIZE: Tuple[float, float, float] = (0.55, 0.02, 0.50)
94
+ WALL_COLOR: Tuple[float, float, float, float] = (0.90, 0.90, 0.92, 1.0)
95
+
96
+ # The "hook" is a horizontal peg (cylinder + end cap) sticking out of the
97
+ # wall toward the robot: its hang point (the peg midpoint) is exactly
98
+ # known, unlike the decorative hook mesh.
99
+ HOOK_X: float = 0.15
100
+ HOOK_Z: float = 0.30 # bag length below the arch is 0.356: at 0.34 the
101
+ # bag bottom only GRAZES the table, the collapsing
102
+ # body loads the loop and it tunnels off the peg
103
+ # within ~60 settle steps (measured); at 0.30 the
104
+ # bag has ~5 cm of slack, rests on the table and
105
+ # the loop is unloaded
106
+ PEG_RADIUS: float = 0.013 # > PBD particle size: a thinner peg lets a
107
+ # loaded cloth loop tunnel straight through
108
+ # (measured at 0.011); the loop hole is only
109
+ # ~3 cm tall so much thicker cannot teleport
110
+ # in without intersecting cloth
111
+ PEG_LEN: float = 0.10
112
+ PEG_TIP_RADIUS: float = 0.007 # thin tip half for threading (see below)
113
+ PEG_TILT_DEG: float = 20.0 # up-tilt toward the tip (see _load_scene)
114
+ PEG_COLOR = (0.30, 0.30, 0.35, 1.0)
115
+
116
+ # ---- payload + shelf ----
117
+ PAYLOAD_SIZE = 0.065 # cage 窗口内且挂环稳定;放大到 0.072/0.078 能加宽 BC
118
+ # 抓取窗但会挤爆挂环段(phase8 布飞散),此消彼长
119
+ PAYLOAD_COLOR = (0.88, 0.15, 0.10, 1.0) # BRIGHT red: the old near-black
120
+ # cube was invisible at RADIO's
121
+ # 192x256 on the dark table, so
122
+ # t=0 aliased with the post-drop
123
+ # state and the FM policy skipped
124
+ # straight to the hang phase
125
+ # (same lesson as Threading v5's
126
+ # thin dark rope)
127
+ SHELF_SIZE = (0.18, 0.18, 0.02) # wide enough that the outermost curled
128
+ # fingertip clears the tray lip
129
+ SHELF_CENTER = (0.31, -0.14, 0.16) # near the bag: a long caged carry drops the cube
130
+ PAYLOAD_SPAWN_JITTER_XY = 0.03
131
+ PAYLOAD_SPAWN_YAW_DEG = 20.0 # random cube yaw; the finger cage
132
+ # tolerates a tilted cube
133
+
134
+ # ---- success ----
135
+ HANG_THR = 0.055
136
+ INSIDE_TOP_DROP_Q = 0.985
137
+ INSIDE_R_Q = 0.85
138
+ INSIDE_R_MULT = 1.15
139
+ INSIDE_Z_MARGIN = 0.01
140
+ HAND_CLEAR_DIST = 0.25 # tcp distance from the hook hang point
141
+ HOLD_STEPS = 8
142
+
143
+ # 2000 Hz: the loaded handle rests in the V of the thumb hook, where the
144
+ # Leap thumb's collision boxes leave a thin slot at the middle joint. At
145
+ # the hand-over pose (the bag lifted the most) it tunnelled out through
146
+ # that slot in 5 of 5 episodes at 1000 Hz and 0 of 5 at 2000 Hz; at
147
+ # 500 Hz it falls through the thumb during the lift.
148
+ SIM_FREQ = 2000
149
+
150
+ # ---- reward ----
151
+ W_PAYLOAD = 0.8
152
+ W_HANDLE = 0.8
153
+ INSIDE_BONUS = 0.5
154
+ SUCCESS_BONUS = 1.6
155
+ TIME_PENALTY = 0.002
156
+ NORMALIZE_REWARD = True
157
+ REWARD_CLIP = (-1.0, 3.5)
158
+
159
+ # ---- scripted policy ----
160
+ HAND_OFFSET_LINK7 = np.array([0.0191, 0.0154, 0.243], dtype=np.float32)
161
+ # gentle cage (PickSoft-derived): a hard finger curl (1.3/1.8) rams the
162
+ # rigid cube sideways out of the cage instead of enclosing it, while a
163
+ # thumb ROTATION target of 1.8 (vs 1.25) keeps the thumb opposing during
164
+ # the close -- swept 8 seeds: 7/7 full carries at (x -0.040, y +0.028,
165
+ # close 0.18) vs 4/8 at the PickSoft values
166
+ GRASP_HAND_JOINTS = np.array([
167
+ 0.0, 1.2, 1.2, 1.2,
168
+ 0.0, 1.2, 1.2, 1.2,
169
+ 0.0, 1.2, 1.2, 1.2,
170
+ 1.8, 0.0, 1.0, 1.0,
171
+ ])
172
+ SP_CLOSE_SPEED = 0.18
173
+ # pd_ee_delta_pose actions are NORMALIZED: 1.0 -> 0.1 m commanded, ~40%
174
+ # tracked, so clip 0.30 is ~1.2 cm/step gross transit and 0.08 is ~3 mm
175
+ SP_GAIN = 5.0
176
+ SP_MAX_DELTA = 0.30
177
+ SP_FINE_MAX_DELTA = 0.10
178
+ SP_POS_THRESH = 0.02
179
+ SP_HAND_SPEED = 0.3
180
+ SP_THUMB_STEPS = 12
181
+ SP_GRASP_STEPS = 40
182
+ # measured tip telemetry (palm-down): the 3 fingertips row spans hand-x
183
+ # [-0.046,+0.044] at y=-0.029 and curl DOWN-AND-BACK to y~-0.06, while
184
+ # the thumb tip sweeps in y at hand-x +0.085, z~palm-0.05. The pinch
185
+ # pocket is therefore at hand (+0.03..0.04, ~0.0): offset the palm so
186
+ # the cube sits there, between the finger wall (-y) and the thumb (+y)
187
+ SP_GRASP_X_OFF = -0.040
188
+ SP_GRASP_Y_OFF = 0.028 # the close intrinsically shoves the cube ~4 cm
189
+ # +y (fingers into the yielding thumb): start it
190
+ # deeper on the fingertip side (8/8 seeds with
191
+ # the 1.8 thumb + 0.18 close speed)
192
+ SP_PAYLOAD_APPROACH_DZ = 0.11 # palm above the payload top on approach
193
+ SP_PAYLOAD_GRASP_DZ = 0.03 # palm 3 cm ABOVE the cube top at the
194
+ # close. With the palm resting ON the cube
195
+ # (0.0) ~60% of cages pressed the cube up
196
+ # against the palm; opening the hand over
197
+ # the bag then squeezed it out sideways
198
+ # (+x, ~0.3 m/s) past the bag's side edge
199
+ # in ~1 of 4 such drops. From 3 cm up the
200
+ # cube hangs in the fingertips instead and
201
+ # falls straight: measured 0 palm-pressed
202
+ # cages, 16/16 first grasps held and 16/16
203
+ # drops in the bag at 2.5-3.5 cm, and 21/21
204
+ # full episodes at 3 cm, vs 13-20% of all
205
+ # episodes lost to the drop at 0.0.
206
+ # 1.5 cm lost 2 of 8 grasps.
207
+ SP_DESCEND_MAX_DELTA = 0.06 # final descent (~5 mm/step): slow enough
208
+ # not to disturb the cube, fast enough
209
+ # that the approach doesn't crawl
210
+ SP_CARRY_PALM_Z = 0.30
211
+ SP_DROP_ABOVE_MOUTH = 0.07 # fingertip clearance above the bag top at drop
212
+ SP_CARRY_CLEAR = 0.015 # carried cube's bottom above the free handle's top
213
+ SP_DROP_Y_BIAS = 0.0 # drop-point bias off the rim midpoint
214
+ SP_DROP_STEPS = 16
215
+ # ---- thumb hook (hanging the free handle, no cloth attachment) ----
216
+ # A human-style INWARD thumb hook, palm kept down (the reset attitude).
217
+ # The opposed thumb (q12 1.6) hangs under the palm as an L: middle
218
+ # phalanx straight down, distal phalanx level and pointing at the wall
219
+ # (-y), which is the axis of the handle's hole. The distal slides
220
+ # through the hole under the arch while the palm passes over it. Then
221
+ # the thumb curls UP toward the palm (q14 0.6, q15 1.88: the distal rises
222
+ # to ~52 deg and its tip stops ~1 cm under the palm), so the handle is
223
+ # caught in the V between the two phalanges, closed above by the palm.
224
+ # Measured: on a level distal the loaded handle slides off the tip within
225
+ # ~10 steps, because the bag pulls it toward the wall and cloth-rigid
226
+ # contact is frictionless. In the curled V the same pull presses it into
227
+ # the bottom of the V instead.
228
+ # Finger knuckles are tilted up 0.3: the index lines up with the thumb in
229
+ # x, and a straight index rams the peg tip during the insertion.
230
+ SP_HOOK_HAND = np.array([0.0, -0.3, 0.0, 0.0,
231
+ 0.0, -0.3, 0.0, 0.0,
232
+ 0.0, -0.3, 0.0, 0.0,
233
+ 1.6, 0.0, 0.0, 1.6])
234
+ SP_CURL_HAND = np.concatenate([SP_HOOK_HAND[:12], [1.6, 0.0, 0.6, 1.88]])
235
+ # The fingers stay straight to the end (only the thumb works); carried
236
+ # high enough (SP_HO_HOVER) they pass over the wall's top edge. (A tight
237
+ # fist (1.5, 1.6, 1.6) reaches into the thumb's V and knocked the handle
238
+ # out, measured.)
239
+ SP_CARRY_HAND = SP_CURL_HAND.copy()
240
+ # At the peg the thumb opens back to the level L, its tip over the peg tip.
241
+ SP_UNCURL_HAND = np.concatenate([SP_CARRY_HAND[:12], [1.6, 0.0, 0.0, 1.6]])
242
+ THUMB_TIP_OFFSET = np.array([0.0, -0.06, -0.015]) # thumb_fingertip -> tip (URDF thumb_tip)
243
+ # uncurled thumb (SP_UNCURL_HAND, palm down): tip relative to the knuckle
244
+ UNCURL_TIP_FROM_KNUCKLE = np.array([-0.015, -0.060, 0.002])
245
+ # The hand is steered by the thumb's middle knuckle (thumb_fingertip link
246
+ # origin) in phases 7-14 and by the tcp in phase 15. The "crook" (mean of
247
+ # knuckle, tip and thumb_dip origin) only detects a handle that fell out of
248
+ # the hook while lifting / carrying.
249
+ SP_HOOK_PALM_ABOVE = 0.065 # palm underside above the knuckle (L-hook, palm down)
250
+ SP_HOOK_DISTAL_TOP = 0.012 # distal phalanx top above the knuckle
251
+ SP_HOOK_FRONT = 0.08 # knuckle this far +y of the arch before inserting
252
+ SP_HOOK_INS = 0.032 # insert until the knuckle is this far +y of the
253
+ # arch band's mean y (the band then lies over
254
+ # the distal ~3.5 cm from the knuckle; +-1 cm
255
+ # measured to work)
256
+ SP_HOOK_LIFT = 0.04 # lift the hooked handle this much before carrying:
257
+ # the carry then rises diagonally to the peg.
258
+ # Lifting 9 cm first loaded the V longer and
259
+ # the handle tunnelled out of it (3 of 3 on
260
+ # one seed; 0 of 2 at 4 cm)
261
+ # The carry ends in the hand-over pose (knuckle where the OPENED thumb's
262
+ # tip sits SP_HO_HOVER above the peg tip) and the thumb opens at once:
263
+ # held still up there, the bag pulls the handle straight down into the
264
+ # bottom of the V and within ~20-30 steps it tunnels out through the
265
+ # thumb's middle joint (5 of 5 episodes that paused ~25 steps).
266
+ SP_HO_HOVER = 0.045 # opened thumb's tip this far above the peg tip's
267
+ # axis: the straight fingers then pass over the
268
+ # wall's top edge (at 0.025 their tips pressed
269
+ # on the wall and the carry stalled ~4 cm short)
270
+ SP_HOOK_RETRACT = 0.13 # pull the thumb back out of the loop along +y
271
+ SP_HOOK_X_ADJ = (0.0, 0.006, -0.006, 0.012, -0.012) # insertion retries
272
+ # The settle holds on the free rim drop once the curled thumb has closed
273
+ # around the free handle (see _hand_supports_free_handle): at least
274
+ # CAPTURE_MIN_PARTICLES of the handle inside the thumb's pocket with the
275
+ # distal at least CAPTURE_MIN_ANGLE_DEG up.
276
+ CAPTURE_MIN_PARTICLES = 4
277
+ CAPTURE_MIN_ANGLE_DEG = 45.0
278
+ # ---- final hand-over assist (the ONLY non-physical step of the hang) ----
279
+ # The thumb carries the handle physically to the peg tip and opens there.
280
+ # A frictionless thumb cannot push the loop along the peg (it slides
281
+ # through the loop, measured), so once the opened hand presents the loop
282
+ # in front of the peg tip the env slides the loop's arch off the thumb
283
+ # onto the peg and keeps it there, exactly like the pre-hung loop's
284
+ # apex (_loop_pin_idx). Let go, the bag's pull toward the robot slid the
285
+ # loop back up the tilted peg and off its tip within ~3 steps (6 of 6).
286
+ HANDOVER_X_TOL = 0.02 # arch apex within this of the peg axis (x)
287
+ HANDOVER_Y_WIN = (-0.01, 0.08) # apex y - peg tip y
288
+ HANDOVER_H_WIN = (-0.005, 0.09) # apex height above the peg's axis line
289
+ HANDOVER_AT = 0.7 # slide the apex to this fraction along the peg
290
+ HANDOVER_STEPS = 12 # control steps for the slide
291
+ HANDOVER_PATCH = (0.03, 0.035, 0.04) # moved patch: below apex z / |dy| / |dx|
292
+ SP_PHASE_TIMEOUT = 80
293
+ SP_MAX_RETRIES = 2 # payload pick retries
294
+ SP_MAX_HANG_RETRIES = 3 # pinch/hang retries (separate budget: a hard
295
+ # payload pick must not eat the hang's retries)
296
+
297
+ def __init__(self, *args: Any, robot_uids: str = "panda_leaphand", **kwargs: Any) -> None:
298
+ self.bag = None
299
+ self.wall = None
300
+ self.hook = None
301
+ self.shelf = None
302
+ self.payload = None
303
+
304
+ self._handle_idx: Tuple[Optional[torch.Tensor], Optional[torch.Tensor]] = (None, None)
305
+ self._hung_handle = 0 # which cached handle was pre-hung
306
+ self._hold_counter = 0
307
+ self._last_eval_step = -1
308
+
309
+ super().__init__(
310
+ *args,
311
+ robot_uids=robot_uids,
312
+ sim_config={"sim_freq": int(self.SIM_FREQ)},
313
+ **kwargs,
314
+ )
315
+
316
+ if self.agent.uid == "panda_leaphand":
317
+ self.agent.keyframes = dict(
318
+ rest=Keyframe(qpos=_QPOS_DEFAULT.copy(), pose=Pose())
319
+ )
320
+
321
+ # ------------------------------------------------------------------ #
322
+ # Scene construction
323
+ # ------------------------------------------------------------------ #
324
+ def _setup_scene(self) -> None:
325
+ # NOTE: keep default solver iterations -- raising
326
+ # max_stretch/bending_solver_iterations makes the solver explode when
327
+ # particles are world-pinned (fix_particles), which the pre-hung
328
+ # handle relies on.
329
+ pbd_options = gs.options.PBDOptions(
330
+ particle_size=0.010,
331
+ )
332
+ self.scene = Scene(
333
+ sim_options=gs.options.SimOptions(
334
+ dt=1.0 / self._sim_freq,
335
+ gravity=self.sim_config.gravity,
336
+ ),
337
+ rigid_options=gs.options.RigidOptions(
338
+ max_collision_pairs=self.sim_config.max_collision_pairs,
339
+ ),
340
+ pbd_options=pbd_options,
341
+ show_viewer=self.show_viewer,
342
+ viewer_options=gs.options.ViewerOptions(
343
+ camera_pos=(3.0, 0.6, 2.0),
344
+ camera_lookat=(0.2, -0.2, 0.4),
345
+ camera_fov=40,
346
+ ),
347
+ profiling_options=self.profiling_options,
348
+ )
349
+
350
+ def _cloth_material(self):
351
+ # A floppy cloth bag (bending_compliance 5e-6) cannot stand: it
352
+ # pancakes on the table within ~100 steps and the mouth closes, which
353
+ # kills both the drop-in and the handle hang. The bag is canvas-stiff
354
+ # by default; overridable for tuning sweeps.
355
+ return gs.materials.PBD.Cloth(
356
+ static_friction=1.2,
357
+ kinetic_friction=1.0,
358
+ rho=4.0,
359
+ stretch_compliance=float(os.environ.get("BAGDROP_STRETCHC", 1e-7)),
360
+ bending_compliance=float(os.environ.get("BAGDROP_BENDC", 1e-8)),
361
+ stretch_relaxation=0.5,
362
+ bending_relaxation=0.3,
363
+ air_resistance=2e-3,
364
+ )
365
+
366
+ # ------------------------------------------------------------------ #
367
+ # Accessors
368
+ # ------------------------------------------------------------------ #
369
+ def _tcp_pos(self) -> torch.Tensor:
370
+ return torch.as_tensor(self.agent.tcp_pos, dtype=torch.float32, device=self.device).view(3)
371
+
372
+ def _bag_pts(self) -> torch.Tensor:
373
+ pts = self.bag.get_particles_pos()
374
+ if isinstance(pts, torch.Tensor):
375
+ P = pts.detach().to(device=self.device, dtype=torch.float32)
376
+ if P.ndim == 3:
377
+ P = P[0]
378
+ return P.reshape(-1, 3)
379
+ pts_np = np.array(pts, dtype=np.float32, copy=True).reshape(-1, 3)
380
+ return torch.as_tensor(pts_np, dtype=torch.float32, device=self.device)
381
+
382
+ def _payload_pos(self) -> torch.Tensor:
383
+ return torch.as_tensor(self.payload.get_pos(), dtype=torch.float32, device=self.device).view(-1)[:3]
384
+
385
+ def _hook_hang_world(self) -> torch.Tensor:
386
+ # hanging loops slide down the tilted peg and rest near the wall
387
+ # root: the hang reference is 25% along the peg from the root
388
+ # (measured resting centroids sit 0.027-0.048 from here)
389
+ p = self._peg_root + 0.25 * (self._peg_tip - self._peg_root)
390
+ return torch.as_tensor(p, dtype=torch.float32, device=self.device)
391
+
392
+ def _handle_apex(self, which: int) -> torch.Tensor:
393
+ """Apex of a handle's OWN arch: the pinch and the hang servo need the
394
+ actually-pinched cloth. The cluster also contains mouth-centre cloth
395
+ that the pre-hung side hoists higher than the free arch, so restrict
396
+ to the cluster's outer half along y (away from the wall) before
397
+ taking the highest particle."""
398
+ idx = self._handle_idx[which]
399
+ P = self._bag_pts()
400
+ if idx is None or P.numel() == 0:
401
+ return torch.zeros(3, dtype=torch.float32, device=self.device)
402
+ idx = idx.to(P.device)
403
+ idx = idx[idx < P.shape[0]]
404
+ if idx.numel() == 0:
405
+ return torch.zeros(3, dtype=torch.float32, device=self.device)
406
+ H = P[idx]
407
+ if H.shape[0] >= 8:
408
+ y_cut = torch.quantile(H[:, 1].to(torch.float32), 0.6)
409
+ outer = H[H[:, 1] >= y_cut]
410
+ if outer.shape[0] > 0:
411
+ H = outer
412
+ return H[int(torch.argmax(H[:, 2]).item())]
413
+
414
+ def _loop_cov(self, which: int) -> Tuple[int, int]:
415
+ """Angular coverage (out of 12 bins) of a handle's particles around
416
+ the peg axis: >=7 means the loop encircles the peg; a loop merely
417
+ RESTING on top covers <=4 and slides off on release (frictionless)."""
418
+ idx = self._handle_idx[which]
419
+ P = self._bag_pts()
420
+ if idx is None or P.numel() == 0:
421
+ return 0, 0
422
+ H = P[idx.to(P.device)].detach().cpu().numpy()
423
+ pH = self._hook_hang_world().detach().cpu().numpy()
424
+ m = np.abs(H[:, 1] - pH[1]) < 0.5 * self.PEG_LEN + 0.02
425
+ Hp = H[m]
426
+ if len(Hp) == 0:
427
+ return 0, 0
428
+ ang = np.arctan2(Hp[:, 2] - pH[2], Hp[:, 0] - pH[0])
429
+ bins = np.unique(((ang + np.pi) / (2 * np.pi) * 12).astype(int) % 12)
430
+ return int(len(bins)), int(len(Hp))
431
+
432
+ def _handle_pos(self, which: int) -> torch.Tensor:
433
+ idx = self._handle_idx[which]
434
+ P = self._bag_pts()
435
+ if idx is None or P.numel() == 0:
436
+ return torch.zeros(3, dtype=torch.float32, device=self.device)
437
+ idx = idx.to(P.device)
438
+ idx = idx[idx < P.shape[0]]
439
+ if idx.numel() == 0:
440
+ return torch.zeros(3, dtype=torch.float32, device=self.device)
441
+ return P[idx].mean(dim=0)
442
+
443
+ # ------------------------------------------------------------------ #
444
+ # Scene loading
445
+ # ------------------------------------------------------------------ #
446
+ def _load_scene(self, options: Dict[str, Any] | None = None) -> None:
447
+ super()._load_scene(options or {})
448
+ # cached links belong to the previous scene's robot after a reconfigure
449
+ self.__dict__.pop("_link_cache", None)
450
+
451
+ # tuning-sweep overrides
452
+ self.HOOK_Z = float(os.environ.get("BAGDROP_HOOK_Z", self.HOOK_Z))
453
+ self.PEG_RADIUS = float(os.environ.get("BAGDROP_PEG_R", self.PEG_RADIUS))
454
+
455
+ wx, wz = self.WALL_XZ_CENTER
456
+ self.wall = self.scene.add_entity(
457
+ morph=gs.morphs.Box(
458
+ pos=(float(wx), float(self.WALL_Y), float(wz)),
459
+ size=self.WALL_SIZE,
460
+ fixed=True,
461
+ collision=True,
462
+ visualization=True,
463
+ ),
464
+ surface=gs.surfaces.Default(color=self.WALL_COLOR),
465
+ )
466
+
467
+ # UPWARD-TILTED peg, like a real coat hook, and NO stub at the tip.
468
+ # A horizontal peg + vertical stub is a trap for frictionless PBD
469
+ # cloth: the stub blocks threading the loop horizontally, seating
470
+ # from above cannot verify threading, and a merely-draped loop
471
+ # always slides off (many failures measured). With a 20 deg
472
+ # up-tilt the loop is threaded HORIZONTALLY past the free tip and
473
+ # gravity then slides it down-slope toward the wall root: tip
474
+ # escape is up-slope, root escape is blocked by the wall.
475
+ tilt = np.deg2rad(self.PEG_TILT_DEG)
476
+ axis = np.array([0.0, np.cos(tilt), np.sin(tilt)])
477
+ root = np.array([self.HOOK_X, self.WALL_Y + 0.5 * self.WALL_SIZE[1], self.HOOK_Z])
478
+ self._peg_root = root
479
+ self._peg_tip = root + self.PEG_LEN * axis
480
+ self._y_hook = float((root + 0.5 * self.PEG_LEN * axis)[1])
481
+ # TAPERED: thick root half (loops cannot tunnel through at rest, and
482
+ # they rest there), thin tip half (the loop hole is only ~3 cm and a
483
+ # full-thickness 2.6 cm tip leaves +-2 mm of threading tolerance --
484
+ # impossible; the 1.4 cm tip makes it +-8 mm)
485
+ euler = (-(90.0 - self.PEG_TILT_DEG), 0.0, 0.0)
486
+ seg_mids = (root + 0.25 * self.PEG_LEN * axis, root + 0.75 * self.PEG_LEN * axis)
487
+ seg_rads = (self.PEG_RADIUS, self.PEG_TIP_RADIUS)
488
+ self.hook = None
489
+ for smid, srad in zip(seg_mids, seg_rads):
490
+ ent = self.scene.add_entity(
491
+ material=gs.materials.Rigid(),
492
+ morph=gs.morphs.Cylinder(
493
+ pos=(float(smid[0]), float(smid[1]), float(smid[2])),
494
+ euler=euler,
495
+ height=0.5 * self.PEG_LEN,
496
+ radius=float(srad),
497
+ collision=True,
498
+ visualization=True,
499
+ fixed=True,
500
+ ),
501
+ surface=gs.surfaces.Default(color=self.PEG_COLOR),
502
+ )
503
+ if self.hook is None:
504
+ self.hook = ent
505
+
506
+ self.shelf = self.scene.add_entity(
507
+ morph=gs.morphs.Box(
508
+ pos=self.SHELF_CENTER,
509
+ size=self.SHELF_SIZE,
510
+ fixed=True,
511
+ collision=True,
512
+ visualization=True,
513
+ ),
514
+ surface=gs.surfaces.Default(color=(0.35, 0.35, 0.38, 1.0)),
515
+ )
516
+ # NOTE: no tray lip around the shelf. A 2 cm lip was tried as a
517
+ # backstop for the close's sideways shove, but the cube rides up onto
518
+ # the lip or wedges into its corner and the capture rate DROPS
519
+ # (0/3 vs ~2/3 per attempt on the bare shelf).
520
+
521
+ spec = self.BAG_SPEC
522
+ bag_scale = float(os.environ.get("BAGDROP_SCALE", spec.get("scale", 1.0)))
523
+ self._bag_scale = bag_scale
524
+ # spawn the bag well above the table in canonical pose (its full z
525
+ # extent must stay above the table or it crumples at build); it is
526
+ # aligned onto the hook by particle translation in _initialize_episode
527
+ init_pos = (float(spec["pos"][0]), float(self._y_hook + 0.12), float(spec["pos"][2]))
528
+ self.bag = self.scene.add_entity(
529
+ material=self._cloth_material(),
530
+ morph=gs.morphs.Mesh(
531
+ file=str(spec["file"]),
532
+ pos=init_pos,
533
+ euler=spec["euler"],
534
+ scale=bag_scale,
535
+ visualization=True,
536
+ collision=True,
537
+ fixed=False,
538
+ ),
539
+ surface=gs.surfaces.Default(color=spec.get("color", (0.85, 0.62, 0.50, 0.85))),
540
+ )
541
+
542
+ self.payload = self.scene.add_entity(
543
+ # ~190 g: too light skitters off the shelf on a graze, too
544
+ # heavy slips out of the gentle cage
545
+ material=gs.materials.Rigid(rho=700.0, friction=1.5),
546
+ morph=gs.morphs.Box(
547
+ pos=(0.6, 0.3, 0.2),
548
+ size=(self.PAYLOAD_SIZE, self.PAYLOAD_SIZE, self.PAYLOAD_SIZE),
549
+ fixed=False,
550
+ collision=True,
551
+ visualization=True,
552
+ ),
553
+ surface=gs.surfaces.Default(color=self.PAYLOAD_COLOR),
554
+ )
555
+
556
+ # ------------------------------------------------------------------ #
557
+ # Handle identification (particle indices, cached at first reset)
558
+ # ------------------------------------------------------------------ #
559
+ def _identify_handles(self) -> None:
560
+ P = self._bag_pts()
561
+ if P.numel() == 0:
562
+ self._handle_idx = (None, None)
563
+ return
564
+ z = P[:, 2]
565
+ z_hi = float(z.max().item())
566
+ band = torch.where(z >= z_hi - float(self.HANDLE_BAND))[0]
567
+ Pb = P[band]
568
+ # split the top band into two clusters along its widest horizontal axis
569
+ spans = (Pb[:, :2].max(dim=0).values - Pb[:, :2].min(dim=0).values)
570
+ ax = int(torch.argmax(spans).item())
571
+ med = float(Pb[:, ax].median().item())
572
+ left = band[Pb[:, ax] <= med]
573
+ right = band[Pb[:, ax] > med]
574
+ if left.numel() == 0 or right.numel() == 0:
575
+ half = band.shape[0] // 2
576
+ left, right = band[:half], band[half:]
577
+ self._handle_idx = (left.detach().clone(), right.detach().clone())
578
+
579
+ def _set_bag_pts(self, new_pts: torch.Tensor) -> None:
580
+ """Write a full (N,3) particle layout back, matching the entity's
581
+ storage shape, and zero the velocities."""
582
+ P = self.bag.get_particles_pos()
583
+ if isinstance(P, torch.Tensor) and P.ndim == 3:
584
+ Pn = P.detach().clone()
585
+ Pn[0] = new_pts.to(device=Pn.device, dtype=Pn.dtype)
586
+ else:
587
+ Pn = new_pts.detach().clone()
588
+ self.bag.set_particles_pos(Pn)
589
+ try:
590
+ self.bag.set_particles_vel(torch.zeros_like(Pn))
591
+ except Exception:
592
+ pass
593
+
594
+ def _translate_bag(self, delta: torch.Tensor) -> None:
595
+ P = self.bag.get_particles_pos()
596
+ if isinstance(P, torch.Tensor):
597
+ P0 = P.detach().to(device=self.device, dtype=torch.float32)
598
+ if P0.ndim == 3:
599
+ Pn = P0.clone()
600
+ Pn[0] = P0[0] + delta.view(1, 3)
601
+ else:
602
+ Pn = P0 + delta.view(1, 3)
603
+ self.bag.set_particles_pos(Pn)
604
+ try:
605
+ self.bag.set_particles_vel(torch.zeros_like(Pn))
606
+ except Exception:
607
+ pass
608
+
609
+ # ------------------------------------------------------------------ #
610
+ # Episode init
611
+ # ------------------------------------------------------------------ #
612
+ def _initialize_episode(self, env_idx: int, options: Dict[str, Any] | None) -> None:
613
+ rng = getattr(self, "np_random", np.random)
614
+ dev = self.device
615
+
616
+ self._hold_counter = 0
617
+ self._last_eval_step = -1
618
+
619
+ # restore the canonical (build-pose) particle layout on re-resets:
620
+ # without this the 2nd episode teleports a crumpled bag and the loop
621
+ # never catches the peg
622
+ P_cur = self._bag_pts()
623
+ if getattr(self, "_canonical_pts", None) is None:
624
+ self._canonical_pts = P_cur.detach().clone()
625
+ else:
626
+ self._set_bag_pts(self._canonical_pts)
627
+
628
+ # identify handles in the canonical (build) pose, then pre-hang one.
629
+ # A free-standing cloth bag is impossible: even at bending_compliance
630
+ # 1e-8 the empty bag pancakes within ~25 steps (top 0.36 -> 0.05).
631
+ # Instead the wall-side handle loop is teleported AROUND the peg and
632
+ # the bag settles hanging from it, bottom resting on the table (hook
633
+ # height chosen so the loop carries almost no load); the mouth then
634
+ # gapes open toward the robot, which is what the drop-in needs.
635
+ self._identify_handles()
636
+ pH = self._hook_hang_world()
637
+ h0 = self._handle_pos(0)
638
+ h1 = self._handle_pos(1)
639
+ # hang the handle that is already closer to the wall (the other one
640
+ # would put the bag body inside the wall)
641
+ self._hung_handle = 0 if float(h0[1]) < float(h1[1]) else 1
642
+ # the handle loop's HOLE sits just below the arch apex (the loop hole
643
+ # is only ~2-3 cm tall at this scale, so targeting the cluster
644
+ # centroid misses it); anchor on the cluster's highest particle
645
+ idx = self._handle_idx[self._hung_handle]
646
+ P = self._bag_pts()
647
+ Ph = P[idx.to(P.device)]
648
+ top_i = int(torch.argmax(Ph[:, 2]).item())
649
+ arch_top = Ph[top_i]
650
+ fixed_init = os.environ.get("ML_FIXED_INIT", "0") == "1" # fixed layout, no randomization
651
+ jx = 0.0 if fixed_init else float(rng.uniform(-self.PREHANG_JITTER_X, self.PREHANG_JITTER_X))
652
+ # target: arch apex just above the peg surface ~35% along its length,
653
+ # loop hole (~3 cm tall vs peg diameter 2.6 cm) around the axis so
654
+ # the teleport does not intersect peg or wall; during the settle the
655
+ # threaded loop slides down-slope and rests against the wall root
656
+ thread_pt = self._peg_root + 0.35 * (self._peg_tip - self._peg_root)
657
+ tgt = torch.tensor(
658
+ [self.HOOK_X + jx, float(thread_pt[1]),
659
+ float(thread_pt[2]) + float(self.PEG_RADIUS) + self.PREHANG_ARCH_DZ],
660
+ dtype=torch.float32, device=dev)
661
+ self._translate_bag(tgt - arch_top)
662
+ self._base_pin_idx = None
663
+ self._rim_pin_idx = None
664
+ self._rim_pinned = False
665
+ self._impact_freed = False
666
+ self._ho_idx = None
667
+ self._handed_over = False
668
+ self._support_streak = 0
669
+ # anchor the hung loop's arch at teleport time: on the tilted peg
670
+ # the threaded loop is STATICALLY stable, but hoisting the free rim
671
+ # later yanks it up-slope over the thin tip half and off (measured
672
+ # d0 0.02 -> 0.34). The pre-hang is the task premise; the hand
673
+ # works near the TIP, 5+ cm from these clamped root particles, so
674
+ # the press-on-clamped-cloth explosion cannot reach them.
675
+ P1 = self._bag_pts()
676
+ idx_h = idx.to(P1.device)
677
+ H1 = P1[idx_h]
678
+ near_pin = (
679
+ (H1[:, 2] > float(tgt[2]) - 0.012)
680
+ & (torch.abs(H1[:, 1] - tgt[1]) < 0.03)
681
+ & (torch.abs(H1[:, 0] - tgt[0]) < 0.035)
682
+ )
683
+ self._loop_pin_idx = idx_h[near_pin].detach().clone()
684
+ self._loop_pin_pos = P1[self._loop_pin_idx].detach().clone()
685
+ if os.environ.get("BAGDROP_DEBUG_SETTLE", "") == "1":
686
+ print(f"[BagDropIn] loop apex pinned: {self._loop_pin_idx.numel()} particles")
687
+ dbg = os.environ.get("BAGDROP_DEBUG_SETTLE", "") == "1"
688
+ if dbg:
689
+ self._debug_loop_stats("teleport")
690
+ settle = int(os.environ.get("BAGDROP_SETTLE", self.PREHANG_SETTLE_STEPS))
691
+ # pin the pooled cloth once it has settled: PBD cloth-rigid contact
692
+ # has NO friction and the pooled bag otherwise creeps sideways on the
693
+ # table, dragging everything (measured ~0.5 mm/step)
694
+ settle_free = min(int(os.environ.get("BAGDROP_SETTLE_FREE", 60)), settle)
695
+ for si in range(settle):
696
+ if si == settle_free:
697
+ self._select_base_pins()
698
+ self.scene.step()
699
+ self._clamp_pins()
700
+ if dbg and (si + 1) % 20 == 0:
701
+ self._debug_loop_stats(f"settle{si+1:4d}")
702
+
703
+ # payload on the shelf
704
+ cx, cy, cz = self.SHELF_CENTER
705
+ if os.environ.get("ML_FIXED_INIT", "0") == "1": # fixed layout, no randomization
706
+ jx = jy = yaw = 0.0
707
+ else:
708
+ jx = float(rng.uniform(-self.PAYLOAD_SPAWN_JITTER_XY, self.PAYLOAD_SPAWN_JITTER_XY))
709
+ jy = float(rng.uniform(-self.PAYLOAD_SPAWN_JITTER_XY, self.PAYLOAD_SPAWN_JITTER_XY))
710
+ yaw = float(np.deg2rad(rng.uniform(-self.PAYLOAD_SPAWN_YAW_DEG, self.PAYLOAD_SPAWN_YAW_DEG)))
711
+ pz = float(cz + 0.5 * self.SHELF_SIZE[2] + 0.5 * self.PAYLOAD_SIZE + 0.003)
712
+ self.payload.set_pos(torch.tensor([cx + jx, cy + jy, pz], dtype=torch.float32, device=dev))
713
+ self.payload.set_quat(torch.tensor(
714
+ [np.cos(0.5 * yaw), 0.0, 0.0, np.sin(0.5 * yaw)], # wxyz, yaw about z
715
+ dtype=torch.float32, device=dev))
716
+ try:
717
+ self.payload.set_vel(torch.zeros(3, device=dev))
718
+ self.payload.set_ang_vel(torch.zeros(3, device=dev))
719
+ except Exception:
720
+ pass
721
+
722
+ self._reset_scripted_policy()
723
+
724
+ try:
725
+ for sensor in self._sensors.values():
726
+ cam = getattr(sensor, "camera", None)
727
+ if cam is not None:
728
+ cam.render(rgb=True, force_render=True)
729
+ except Exception as exc:
730
+ print(f"[BagDropIn] reset render flush failed: {exc}")
731
+
732
+ def _select_base_pins(self) -> None:
733
+ """Pick the particles held by the settle holds (see settle loop): the
734
+ pooled on-table cloth (minus both handle clusters) and the FREE
735
+ handle where it settled. (The pre-hung loop's apex is pinned
736
+ separately, _loop_pin_idx.) Both holds drop once the thumb has
737
+ captured the free handle; lifting and carrying it are then physical,
738
+ and only the final hand-over onto the peg is assisted (HANDOVER_*)."""
739
+ P = self._bag_pts()
740
+ z_thr = float(os.environ.get("BAGDROP_PIN_Z", 0.014))
741
+ mask = P[:, 2] < z_thr
742
+ # never pin the handle clusters: the free handle must stay liftable
743
+ for hidx in self._handle_idx:
744
+ if hidx is not None:
745
+ m2 = torch.ones(P.shape[0], dtype=torch.bool, device=P.device)
746
+ m2[hidx.to(P.device)] = False
747
+ mask = mask & m2
748
+ idx = torch.where(mask)[0]
749
+ self._base_pin_idx = idx.detach().clone()
750
+ self._base_pin_pos = P[idx].detach().clone()
751
+ # ALSO hold the free handle's rim where it settled: the frictionless
752
+ # front wall otherwise creeps down ~0.5 mm/step and is fully pooled
753
+ # (z 0.29 -> 0.04) by the time the hand comes to hook it. The hold
754
+ # is released once the curled thumb has held the free handle in its
755
+ # pocket for 2 control steps (_hand_supports_free_handle).
756
+ fidx = self._handle_idx[1 - int(self._hung_handle)]
757
+ if fidx is not None:
758
+ fidx = fidx.to(P.device)
759
+ self._rim_pin_idx = fidx.detach().clone()
760
+ self._rim_pin_pos = P[fidx].detach().clone()
761
+ self._rim_pinned = True
762
+ if os.environ.get("BAGDROP_DEBUG_SETTLE", "") == "1":
763
+ print(f"[BagDropIn] base pinned {idx.numel()} particles, "
764
+ f"rim pinned {0 if fidx is None else fidx.numel()}")
765
+
766
+ def _clamp_pins(self) -> None:
767
+ """Hold the pinned particles by SUBSET position writes every sim
768
+ substep. Native fix_particles explodes this bag within the settle
769
+ (measured); subset writes do not."""
770
+ names = ["_base_pin_idx", "_loop_pin_idx"]
771
+ if getattr(self, "_rim_pinned", False):
772
+ names.append("_rim_pin_idx")
773
+ for iname in names:
774
+ idx = getattr(self, iname, None)
775
+ if idx is None or idx.numel() == 0:
776
+ continue
777
+ pos = getattr(self, iname.replace("_idx", "_pos"))
778
+ try:
779
+ self.bag.set_particles_pos(pos, particles_idx_local=idx)
780
+ self.bag.set_particles_vel(torch.zeros_like(pos), particles_idx_local=idx)
781
+ except Exception as exc:
782
+ print(f"[BagDropIn] clamp {iname} failed: {exc}")
783
+ ho = getattr(self, "_ho_idx", None)
784
+ if ho is not None and ho.numel() > 0:
785
+ # hand-over: smoothstep slide from the start pose onto the peg,
786
+ # then held there
787
+ self._ho_sub = min(self._ho_sub + 1, self._ho_total)
788
+ u = self._ho_sub / max(1, self._ho_total)
789
+ ua = min(1.0, u / 0.7)
790
+ un = max(0.0, (u - 0.5) / 0.5)
791
+ ua, un = ua * ua * (3.0 - 2.0 * ua), un * un * (3.0 - 2.0 * un)
792
+ pos = self._ho_p0 + (ua * self._ho_delta_a + un * self._ho_delta_n)[None, :]
793
+ try:
794
+ self.bag.set_particles_pos(pos, particles_idx_local=ho)
795
+ self.bag.set_particles_vel(torch.zeros_like(pos), particles_idx_local=ho)
796
+ except Exception as exc:
797
+ print(f"[BagDropIn] hand-over clamp failed: {exc}")
798
+
799
+ def _after_simulation_step(self):
800
+ # clamp every SIM substep: at control-step cadence the pinned
801
+ # particles drift for ~10 substeps and the big position snap-back
802
+ # injects energy until the PBD solver explodes (measured)
803
+ self._clamp_pins()
804
+ return super()._after_simulation_step()
805
+
806
+ def _debug_loop_stats(self, tag: str) -> None:
807
+ """BAGDROP_DEBUG_SETTLE=1 diagnostics: is the hung loop around the peg?"""
808
+ P = self._bag_pts().detach().cpu().numpy()
809
+ idx = self._handle_idx[self._hung_handle]
810
+ H = P[idx.detach().cpu().numpy()]
811
+ pH = self._hook_hang_world().detach().cpu().numpy()
812
+ m = np.abs(H[:, 1] - pH[1]) < 0.5 * self.PEG_LEN + 0.02
813
+ Hp = H[m]
814
+ if len(Hp):
815
+ dx = Hp[:, 0] - pH[0]
816
+ dz = Hp[:, 2] - pH[2]
817
+ ang = np.arctan2(dz, dx)
818
+ bins = np.unique(((ang + np.pi) / (2 * np.pi) * 12).astype(int) % 12)
819
+ r = np.sqrt(dx ** 2 + dz ** 2)
820
+ cov, rmin = len(bins), float(r.min())
821
+ else:
822
+ cov, rmin = 0, -1.0
823
+ h = H.mean(0)
824
+ print(f"[BagDropIn:{tag}] hung_c=({h[0]:.3f},{h[1]:.3f},{h[2]:.3f}) "
825
+ f"hmax_z={H[:,2].max():.3f} near_peg={len(Hp)} cov={cov}/12 rmin={rmin:.3f} "
826
+ f"bag_top={P[:,2].max():.3f}")
827
+
828
+ def _after_control_step(self):
829
+ try:
830
+ payload_z = float(self._payload_pos()[2].item())
831
+ except Exception:
832
+ payload_z = 1.0
833
+ if payload_z < float(self.SHELF_CENTER[2]) - 0.03:
834
+ # the payload is mid-fall into the bag: free ONLY the pinned
835
+ # cloth under the impact point BEFORE it lands (a 190 g cube
836
+ # striking position-clamped cloth blows up the PBD solver).
837
+ # The rest of the holds stay: releasing everything here lets
838
+ # the whole front wall pancake to z~0.01 where the hand cannot
839
+ # reach it any more (both measured).
840
+ idx = getattr(self, "_base_pin_idx", None)
841
+ if idx is not None and idx.numel() > 0 and not getattr(self, "_impact_freed", False):
842
+ self._impact_freed = True
843
+ pay_xy = self._payload_pos()[:2]
844
+ keep = torch.linalg.norm(
845
+ self._base_pin_pos[:, :2] - pay_xy[None, :], dim=1) > 0.07
846
+ self._base_pin_idx = idx[keep]
847
+ self._base_pin_pos = self._base_pin_pos[keep]
848
+ # The settle holds (free rim + pooled base) drop once the curled
849
+ # thumb has held the free handle for 2 steps: from then on the handle
850
+ # is carried only by resting in the thumb's hook. Measured: an
851
+ # unsupported free handle falls to the table within ~12 control steps
852
+ # of the release, so the release must wait for the support.
853
+ if getattr(self, "_rim_pinned", False):
854
+ if self._hand_supports_free_handle():
855
+ self._support_streak = getattr(self, "_support_streak", 0) + 1
856
+ if self._support_streak >= 2:
857
+ self._rim_pinned = False
858
+ self._rim_pin_idx = None
859
+ self._base_pin_idx = None
860
+ else:
861
+ self._support_streak = 0
862
+ self._update_handover()
863
+
864
+ # ------------------------------------------------------------------ #
865
+ # Thumb capture of the free handle (releases the settle holds)
866
+ # ------------------------------------------------------------------ #
867
+ def _link(self, name: str):
868
+ cache = self.__dict__.setdefault("_link_cache", {})
869
+ if name not in cache:
870
+ cache[name] = self.agent.robot.get_link(name)
871
+ return cache[name]
872
+
873
+ def _link_point(self, name: str, local) -> torch.Tensor:
874
+ lk = self._link(name)
875
+ pos = torch.as_tensor(lk.get_pos(), dtype=torch.float32, device=self.device).reshape(-1)[:3]
876
+ R = _quat_wxyz_to_R_t(torch.as_tensor(lk.get_quat(), dtype=torch.float32, device=self.device).reshape(-1)[:4])
877
+ return pos + R @ torch.as_tensor(np.asarray(local, np.float32), device=self.device)
878
+
879
+ def _thumb_points(self):
880
+ """World (knuckle, tip, thumb_dip origin) of the thumb, numpy."""
881
+ kn = self._link_point("thumb_fingertip", (0.0, 0.0, 0.0))
882
+ tip = self._link_point("thumb_fingertip", self.THUMB_TIP_OFFSET)
883
+ j14 = self._link_point("thumb_dip", (0.0, 0.0, 0.0))
884
+ return tuple(v.detach().cpu().numpy().astype(np.float64) for v in (kn, tip, j14))
885
+
886
+ def _thumb_pocket(self) -> Tuple[int, float]:
887
+ """(free-handle particles inside the curled thumb's pocket, distal
888
+ angle above horizontal in deg). The pocket is the quad knuckle ->
889
+ tip -> tip + 2 cm up -> thumb_dip + 2 cm up -> thumb_dip in the
890
+ thumb's flexion plane (world y-z with the palm down), within 2 cm of
891
+ the thumb in x."""
892
+ idx = self._handle_idx[1 - int(self._hung_handle)]
893
+ if idx is None:
894
+ return 0, 0.0
895
+ A = self._bag_pts()[idx.to(self.device)].detach().cpu().numpy()
896
+ kn, tip, j14 = self._thumb_points()
897
+ d = tip - kn
898
+ ang = float(np.degrees(np.arctan2(d[2], abs(d[1]))))
899
+ up = np.array([0.0, 0.0, 0.02])
900
+ poly = np.array([kn, tip, tip + up, j14 + up, j14])[:, 1:3]
901
+ Q = A[np.abs(A[:, 0] - tip[0]) < 0.02][:, 1:3]
902
+ inside = np.zeros(len(Q), bool)
903
+ for i in range(len(poly)):
904
+ p, q = poly[i], poly[(i + 1) % len(poly)]
905
+ dz = q[1] - p[1]
906
+ cross = (p[1] > Q[:, 1]) != (q[1] > Q[:, 1])
907
+ y_at = p[0] + (Q[:, 1] - p[1]) * (q[0] - p[0]) / (dz if abs(dz) > 1e-12 else 1e-12)
908
+ inside ^= cross & (Q[:, 0] < y_at)
909
+ return int(inside.sum()), ang
910
+
911
+ def _hand_supports_free_handle(self) -> bool:
912
+ """True once the curled thumb has closed around the free handle:
913
+ handle cloth inside the thumb's pocket AND the distal curled up
914
+ steeply enough to hold it. (A check on "cloth above a digit" fired
915
+ while only the thumb tip reached the arch's edge; the holds dropped
916
+ and the handle fell, measured.)"""
917
+ n, ang = self._thumb_pocket()
918
+ return n >= self.CAPTURE_MIN_PARTICLES and ang >= self.CAPTURE_MIN_ANGLE_DEG
919
+
920
+ def _binder_is_attached(self) -> bool:
921
+ """Kept for scripts/diag_bagdrop.py: no cloth is ever attached to the
922
+ hand any more (the thumb hooks the handle)."""
923
+ return False
924
+
925
+ # ------------------------------------------------------------------ #
926
+ # Hand-over assist (see HANDOVER_*)
927
+ # ------------------------------------------------------------------ #
928
+ def _peg_frame(self):
929
+ root = torch.as_tensor(np.asarray(self._peg_root), dtype=torch.float32, device=self.device)
930
+ tip = torch.as_tensor(np.asarray(self._peg_tip), dtype=torch.float32, device=self.device)
931
+ a = tip - root
932
+ length = torch.linalg.norm(a)
933
+ a = a / length
934
+ n = torch.tensor([0.0, 0.0, 1.0], device=self.device)
935
+ n = n - (n @ a) * a
936
+ return root, tip, a, n / torch.linalg.norm(n), length
937
+
938
+ def _handover_ready(self) -> bool:
939
+ """The opened hand presents the free handle's loop in front of the
940
+ peg tip, its arch above the peg's axis line."""
941
+ apex = self._handle_apex(1 - int(self._hung_handle))
942
+ root, tip, a, n, _ = self._peg_frame()
943
+ if abs(float(apex[0] - tip[0])) > self.HANDOVER_X_TOL:
944
+ return False
945
+ dy = float(apex[1] - tip[1])
946
+ if not (self.HANDOVER_Y_WIN[0] <= dy <= self.HANDOVER_Y_WIN[1]):
947
+ return False
948
+ h = float((apex - root) @ n)
949
+ return self.HANDOVER_H_WIN[0] <= h <= self.HANDOVER_H_WIN[1]
950
+
951
+ def _start_handover(self) -> None:
952
+ """Slide the arch patch around the apex onto the peg (see _clamp_pins)."""
953
+ fidx = self._handle_idx[1 - int(self._hung_handle)].to(self.device)
954
+ P = self._bag_pts()
955
+ H = P[fidx]
956
+ apex = self._handle_apex(1 - int(self._hung_handle))
957
+ dz, dy, dx = self.HANDOVER_PATCH
958
+ m = ((H[:, 2] > apex[2] - dz) & (torch.abs(H[:, 1] - apex[1]) < dy)
959
+ & (torch.abs(H[:, 0] - apex[0]) < dx))
960
+ root, tip, a, n, length = self._peg_frame()
961
+ f = float(self.HANDOVER_AT)
962
+ r = float(self.PEG_RADIUS) if f < 0.5 else float(self.PEG_TIP_RADIUS)
963
+ goal = root + f * length * a + n * (r + 0.006)
964
+ goal[0] = apex[0]
965
+ self._ho_idx = fidx[m].detach().clone()
966
+ self._ho_p0 = P[self._ho_idx].detach().clone()
967
+ d = (goal - apex).detach()
968
+ # split the slide: along the peg first at the presented height, then
969
+ # down onto it -- a straight line cuts through the thumb tip that
970
+ # sits just above the peg tip. x is kept (the thumb presented the
971
+ # loop at the peg's x; HANDOVER_X_TOL bounds the offset).
972
+ self._ho_delta_n = ((d @ n) * n).clone()
973
+ self._ho_delta_a = (d - self._ho_delta_n).clone()
974
+ self._ho_sub = 0
975
+ self._ho_total = int(self.HANDOVER_STEPS) * int(self._sim_steps_per_control)
976
+
977
+ def _update_handover(self) -> None:
978
+ """Called every control step: start the slide when the loop is
979
+ presented by an OPENED hand; the arch then stays held on the peg."""
980
+ if getattr(self, "_rim_pinned", False) or getattr(self, "_handed_over", False):
981
+ return
982
+ if getattr(self, "_ho_idx", None) is None:
983
+ # the settle holds are gone (the thumb had captured the handle)
984
+ # and the thumb has opened again: its distal is back below the
985
+ # capture angle. (A handle slipping out of a still-curled thumb
986
+ # is not handed over.)
987
+ if self._thumb_pocket()[1] < self.CAPTURE_MIN_ANGLE_DEG and self._handover_ready():
988
+ self._start_handover()
989
+ elif self._ho_sub >= self._ho_total:
990
+ self._handed_over = True
991
+
992
+ # ------------------------------------------------------------------ #
993
+ # Inside check
994
+ # ------------------------------------------------------------------ #
995
+ def _quantile(self, x: torch.Tensor, q: float) -> torch.Tensor:
996
+ if x.numel() == 0:
997
+ return torch.tensor(0.0, device=x.device, dtype=torch.float32)
998
+ return torch.quantile(x.to(dtype=torch.float32), q)
999
+
1000
+ def _payload_inside_bag(self, P: torch.Tensor, payload: torch.Tensor) -> Tuple[bool, torch.Tensor]:
1001
+ if P.numel() == 0 or P.shape[0] < 32:
1002
+ return False, torch.zeros(3, device=self.device)
1003
+ z = P[:, 2]
1004
+ z_cut = self._quantile(z, float(self.INSIDE_TOP_DROP_Q))
1005
+ body = P[z <= z_cut]
1006
+ if body.shape[0] < 32:
1007
+ body = P
1008
+ com = body.mean(dim=0)
1009
+ r = torch.linalg.norm(body[:, :2] - com[:2][None, :], dim=1)
1010
+ r_thr = float(self._quantile(r, float(self.INSIDE_R_Q)).item()) * float(self.INSIDE_R_MULT)
1011
+ z_lo = float(self._quantile(body[:, 2], 0.05).item())
1012
+ z_hi = float(self._quantile(body[:, 2], 0.95).item())
1013
+ d_xy = float(torch.linalg.norm(payload[:2] - com[:2]).item())
1014
+ pz = float(payload[2].item())
1015
+ inside = (d_xy <= r_thr) and (z_lo + self.INSIDE_Z_MARGIN <= pz <= z_hi - self.INSIDE_Z_MARGIN)
1016
+ return inside, com
1017
+
1018
+ # ------------------------------------------------------------------ #
1019
+ # Evaluation
1020
+ # ------------------------------------------------------------------ #
1021
+ def evaluate(self) -> Dict[str, torch.Tensor]:
1022
+ dev = self.device
1023
+ P = self._bag_pts()
1024
+ pH = self._hook_hang_world()
1025
+ h0 = self._handle_pos(0)
1026
+ h1 = self._handle_pos(1)
1027
+ d0 = torch.linalg.norm(h0 - pH)
1028
+ d1 = torch.linalg.norm(h1 - pH)
1029
+ hung_both = bool((d0 <= self.HANG_THR).item()) and bool((d1 <= self.HANG_THR).item())
1030
+
1031
+ payload = self._payload_pos()
1032
+ inside, bag_com = self._payload_inside_bag(P, payload)
1033
+
1034
+ # Diagnostic only: whether the hand has RELEASED and moved its fingertips
1035
+ # clear: the tcp (wrist) alone sits ~0.29 from the hook while the
1036
+ # fingers still pinch the loop on the peg, which latched success
1037
+ # mid-hang before this used the fingertips
1038
+ hand_clear = self._fingertip_dist(pH) >= self.HAND_CLEAR_DIST
1039
+
1040
+ criteria_now = hung_both and inside
1041
+ cur_step = int(self._elapsed_steps.reshape(-1)[0].item())
1042
+ if cur_step != self._last_eval_step:
1043
+ self._last_eval_step = cur_step
1044
+ self._hold_counter = self._hold_counter + 1 if criteria_now else 0
1045
+ success = bool(self._hold_counter >= self.HOLD_STEPS)
1046
+
1047
+ return {
1048
+ "success": torch.tensor([success], dtype=torch.bool, device=dev),
1049
+ "d_handle0": d0.view(1),
1050
+ "d_handle1": d1.view(1),
1051
+ "hung_both": torch.tensor([1.0 if hung_both else 0.0], dtype=torch.float32, device=dev),
1052
+ "inside": torch.tensor([1.0 if inside else 0.0], dtype=torch.float32, device=dev),
1053
+ "hand_clear": torch.tensor([1.0 if hand_clear else 0.0], dtype=torch.float32, device=dev),
1054
+ "hold": torch.tensor([float(self._hold_counter)], dtype=torch.float32, device=dev),
1055
+ "payload_pos": payload.view(1, 3),
1056
+ "bag_com": bag_com.view(1, 3),
1057
+ }
1058
+
1059
+ def _fingertip_dist(self, pH: torch.Tensor) -> float:
1060
+ """Distance from the hook hang point to the nearest fingertip."""
1061
+ d_tip = 10.0
1062
+ try:
1063
+ for lk in self.agent.tip_links:
1064
+ p = torch.as_tensor(lk.get_pos(), dtype=torch.float32,
1065
+ device=self.device).view(-1)[:3]
1066
+ d_tip = min(d_tip, float(torch.linalg.norm(p - pH).item()))
1067
+ except Exception:
1068
+ d_tip = float(torch.linalg.norm(self._tcp_pos() - pH).item())
1069
+ return d_tip
1070
+
1071
+ def _get_obs_extra(self, info: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
1072
+ return {
1073
+ "d_handle0": info["d_handle0"],
1074
+ "d_handle1": info["d_handle1"],
1075
+ "hung_both": info["hung_both"],
1076
+ "inside": info["inside"],
1077
+ "hand_clear": info["hand_clear"],
1078
+ "payload_pos": info["payload_pos"],
1079
+ "bag_com": info["bag_com"],
1080
+ }
1081
+
1082
+ # ------------------------------------------------------------------ #
1083
+ # Scripted policy
1084
+ # ------------------------------------------------------------------ #
1085
+ def _reset_scripted_policy(self) -> None:
1086
+ self._sp_phase = 0
1087
+ self._sp_wait = 0
1088
+ self._sp_retries = 0
1089
+ self._sp_hang_retries = 0
1090
+ self._sp_hand_offset: Optional[np.ndarray] = None
1091
+ self._sp_grasp_xyz: Optional[np.ndarray] = None
1092
+ self._sp_R_ref: Optional[np.ndarray] = None
1093
+ self._sp_phase_steps = 0
1094
+ self._sp_last_phase = 0
1095
+ self._sp_hole: Optional[np.ndarray] = None
1096
+ self._sp_ins_tries = 0
1097
+ self._sp_near = 0
1098
+ self._sp_ret_start: Optional[np.ndarray] = None
1099
+ self._sp_arch_z0 = 0.0
1100
+ self._sp_band: Optional[dict] = None
1101
+ self._sp_hold_pos: Optional[np.ndarray] = None
1102
+ self._sp_kn_hist: list = []
1103
+
1104
+ # ---- hook helpers ----
1105
+ def _thumb_tip_np(self) -> np.ndarray:
1106
+ return self._link_point("thumb_fingertip", self.THUMB_TIP_OFFSET).detach().cpu().numpy().astype(np.float32)
1107
+
1108
+ def _thumb_knuckle_np(self) -> np.ndarray:
1109
+ return self._link_point("thumb_fingertip", (0.0, 0.0, 0.0)).detach().cpu().numpy().astype(np.float32)
1110
+
1111
+ def _thumb_crook_np(self) -> np.ndarray:
1112
+ kn, tip, j14 = self._thumb_points()
1113
+ return ((kn + tip + j14) / 3.0).astype(np.float32)
1114
+
1115
+ def _retry_hook_insertion(self) -> None:
1116
+ """Back out and insert again. Once the settle holds are gone the
1117
+ handle is loose and the band measured in phase 7 no longer describes
1118
+ it: measure again (a hang retry) instead."""
1119
+ if not self._rim_pinned and self._sp_hang_retries < self.SP_MAX_HANG_RETRIES:
1120
+ self._sp_hang_retries += 1
1121
+ self._sp_band = None
1122
+ self._sp_phase, self._sp_wait, self._sp_near = 7, 0, 0
1123
+ else:
1124
+ self._sp_phase, self._sp_wait, self._sp_near = 8, 0, 0
1125
+
1126
+ def _free_arch_np(self) -> np.ndarray:
1127
+ return self._handle_pos(1 - int(self._hung_handle)).detach().cpu().numpy().astype(np.float32)
1128
+
1129
+ def _free_arch_core_np(self) -> np.ndarray:
1130
+ """Mean of the free handle's cloth within 6 cm of its median: the arch
1131
+ itself, without the stray legs."""
1132
+ A = self._bag_pts()[self._handle_idx[1 - int(self._hung_handle)].to(self.device)].detach().cpu().numpy()
1133
+ med = np.median(A, axis=0)
1134
+ core = A[np.linalg.norm(A - med, axis=1) < 0.06]
1135
+ return (core if len(core) else A).mean(axis=0).astype(np.float32)
1136
+
1137
+ def _hook_hole(self) -> Tuple[np.ndarray, float]:
1138
+ """Where to push the thumb through the free handle: the point of the
1139
+ loop plane (x-z) under the arch with the most clearance from the cloth
1140
+ around it, and that clearance. The hole is only ~4.4 cm wide for a
1141
+ ~2.4 cm thumb plus 5 mm particle radii on both sides, and the handle
1142
+ cluster's own centroid sat near one side of it (the thumb jammed)."""
1143
+ P = self._bag_pts().detach().cpu().numpy()
1144
+ A = P[self._handle_idx[1 - int(self._hung_handle)].detach().cpu().numpy()]
1145
+ xc, yl = float(np.median(A[:, 0])), float(np.median(A[:, 1]))
1146
+ z_arch = float(np.quantile(A[:, 2], 0.2))
1147
+ # every particle along the thumb's path counts: the bag's side walls
1148
+ # run from the loop plane 10 cm into the bag and jammed the thumb
1149
+ slab = P[(P[:, 1] < yl + 0.02) & (P[:, 1] > yl - 0.125) & (np.abs(P[:, 0] - xc) < 0.08)
1150
+ & (P[:, 2] > z_arch - 0.14) & (P[:, 2] < z_arch + 0.04)]
1151
+ xs = np.arange(xc - 0.035, xc + 0.0351, 0.0025)
1152
+ zs = np.arange(z_arch - 0.09, z_arch - 0.005, 0.0025)
1153
+ X, Z = np.meshgrid(xs, zs, indexing="ij")
1154
+ if len(slab) == 0:
1155
+ return np.array([xc, yl, z_arch - 0.04], np.float32), 0.0
1156
+ d = np.min(np.hypot(X[..., None] - slab[None, None, :, 0], Z[..., None] - slab[None, None, :, 2]), axis=2)
1157
+ i, j = np.unravel_index(int(np.argmax(d)), d.shape)
1158
+ return np.array([xs[i], yl, zs[j]], np.float32), float(d[i, j])
1159
+
1160
+ def _hook_band(self) -> dict:
1161
+ """Where to hook: the hole, the arch band above it (free-handle cloth
1162
+ within 1.2 cm of the hole's x, above its centre) and the knuckle
1163
+ height that threads the thumb's distal under the band while the palm
1164
+ passes over it."""
1165
+ hole, _ = self._hook_hole()
1166
+ A = self._bag_pts()[self._handle_idx[1 - int(self._hung_handle)].to(self.device)].detach().cpu().numpy()
1167
+ m = (np.abs(A[:, 0] - hole[0]) < 0.012) & (A[:, 2] > hole[2] + 0.015) & (A[:, 1] > hole[1] - 0.05)
1168
+ band = A[m] if int(m.sum()) >= 3 else A[A[:, 2] > np.quantile(A[:, 2], 0.8)]
1169
+ z_lo, z_hi = float(band[:, 2].min()) - 0.005, float(band[:, 2].max()) + 0.005
1170
+ z_kn = 0.5 * ((z_hi - self.SP_HOOK_PALM_ABOVE) + (z_lo - self.SP_HOOK_DISTAL_TOP))
1171
+ return dict(hole=hole, near_y=float(band[:, 1].max()), mean_y=float(band[:, 1].mean()), z_kn=z_kn,
1172
+ z_hi=z_hi)
1173
+
1174
+ # Include near-stationary startup histories in demonstrations: evaluation
1175
+ # repeats the reset frame, while training starts at obs_horizon - 1.
1176
+ SP_IDLE_START_STEPS = 3
1177
+
1178
+ def scripted_policy(self) -> np.ndarray:
1179
+ """Pick the payload, drop it into the hanging bag's open mouth, then
1180
+ hang the free handle onto the hook.
1181
+
1182
+ Phases (pd_ee_delta_pose, 22-dim):
1183
+ 0 ApproachPayload : above the payload cube (grasp point frozen).
1184
+ 1 Thumb : pre-rotate the thumb AT HEIGHT.
1185
+ 2 DescendPayload : slow straight descent, palm just above the cube top.
1186
+ 3 GraspPayload : cage the cube.
1187
+ 4 LiftCarry : lift straight up, then travel over the mouth.
1188
+ 5 Drop : open the hand; the cube falls into the bag.
1189
+ 6 RetreatUp : rise clear of the bag.
1190
+ 7 HookShape : shape the inward L-hook (opposed thumb under the
1191
+ palm, distal toward the wall) above the free
1192
+ handle; measure its hole and arch band.
1193
+ 8 HookFront : knuckle at the threading height, in front of
1194
+ the arch.
1195
+ 9 HookInsert : slide the thumb's distal through the hole (-y)
1196
+ under the arch, the palm passing over it.
1197
+ 10 HookCurl : curl the thumb up toward the palm around the
1198
+ arch; the env drops its settle holds.
1199
+ 11 Lift : let the handle settle into the V, lift it.
1200
+ 12 CarryToPeg : knuckle to where the OPENED thumb's tip will sit
1201
+ just above the peg tip, fingers folded away.
1202
+ 13 HandOver : open the thumb at once; the loop leaves it for
1203
+ the peg (env assist, see HANDOVER_*).
1204
+ 14 Retract : pull the thumb back (+y).
1205
+ 15 Retreat : up and away; hold still clear of the hook.
1206
+ """
1207
+ tcp_pos = self.agent.tcp_pos.detach().cpu().numpy().flatten()
1208
+ current_hand_qpos = self.agent.hand_qpos.detach().cpu().numpy().astype(np.float64)
1209
+
1210
+ if not hasattr(self, "_sp_phase"):
1211
+ self._reset_scripted_policy()
1212
+
1213
+ # live palm offset: the EE orientation drifts through large motions
1214
+ # and a stale cached offset walks the palm off its commanded position
1215
+ q = self.agent.tcp_quat.detach().cpu().numpy().flatten()
1216
+ R = Rot.from_quat([q[1], q[2], q[3], q[0]]).as_matrix()
1217
+ off = (R @ self.HAND_OFFSET_LINK7).astype(np.float32)
1218
+ self._sp_hand_offset = off
1219
+
1220
+ # orientation hold: with zero rotation actions the EE attitude drifts
1221
+ # (IK runs only 10 iterations) and the reachable workspace shrinks;
1222
+ # servo back to the reset attitude every step (delta_quat is applied
1223
+ # world-frame with a -0.5 factor, hence the -2x rotvec)
1224
+ if getattr(self, "_sp_R_ref", None) is None:
1225
+ self._sp_R_ref = R.copy()
1226
+ rot_err = Rot.from_matrix(self._sp_R_ref @ R.T).as_rotvec()
1227
+ rot_cmd = np.clip(-2.0 * rot_err, -1.0, 1.0).astype(np.float32)
1228
+
1229
+ if int(self._elapsed_steps.reshape(-1)[0].item()) < self.SP_IDLE_START_STEPS:
1230
+ action = np.zeros(22, dtype=np.float32)
1231
+ action[3:6] = rot_cmd
1232
+ return action
1233
+
1234
+ def _palm_to_tcp(palm_xyz: np.ndarray) -> np.ndarray:
1235
+ return (palm_xyz - off).astype(np.float32)
1236
+
1237
+ def _grasp_delta():
1238
+ nxt = PandaLeaphand.grasp_hand_policy(
1239
+ current_hand_qpos, speed=self.SP_CLOSE_SPEED,
1240
+ target=self.GRASP_HAND_JOINTS.astype(np.float64),
1241
+ )
1242
+ return (nxt - current_hand_qpos).astype(np.float32)
1243
+
1244
+ def _release_delta():
1245
+ nxt = PandaLeaphand.release_hand_policy(current_hand_qpos, speed=self.SP_HAND_SPEED)
1246
+ return (nxt - current_hand_qpos).astype(np.float32)
1247
+
1248
+ payload = self._payload_pos().detach().cpu().numpy().flatten()
1249
+ pH = self._hook_hang_world().detach().cpu().numpy().flatten()
1250
+ h0 = self._handle_pos(0).detach().cpu().numpy().flatten()
1251
+ h1 = self._handle_pos(1).detach().cpu().numpy().flatten()
1252
+ # only the robot-side handle is free; the wall-side one is pre-hung
1253
+ free = 1 - int(self._hung_handle)
1254
+ P = self._bag_pts()
1255
+ bag_top_z = float(P[:, 2].max().item()) if P.numel() else 0.3
1256
+ # the mouth pocket gapes between the hung rim (wall side, high) and
1257
+ # the fallen rim (robot side); aim the drop between the two rims.
1258
+ # x from the bag BODY centre: the handle-cluster mean sits ~2 cm
1259
+ # toward +x and a cube released there clips the mouth's +x edge and
1260
+ # bounces out (measured)
1261
+ bag_cx = float(P[:, 0].mean().item()) if P.numel() else self.HOOK_X
1262
+ mouth_xy = np.array(
1263
+ [bag_cx, 0.5 * (h0[1] + h1[1]) + self.SP_DROP_Y_BIAS], np.float32)
1264
+
1265
+ # A cube that slips out of the cage during the carry sometimes lands
1266
+ # IN the bag (5 of ~50 measured episodes). That is the drop done, not
1267
+ # a failed pick: the retry used to re-grasp it from inside the bag and
1268
+ # the episode was lost. Resting in the bag = in the bag body AND on
1269
+ # cloth (z > 0.04; a cube on the bare table sits at z = 0.032).
1270
+ if int(self._sp_phase) <= 4 and 0.04 < float(payload[2]) < 0.15:
1271
+ if self._payload_inside_bag(P, self._payload_pos())[0]:
1272
+ self._sp_phase = 6
1273
+ self._sp_wait = 0
1274
+
1275
+ # per-phase timeout
1276
+ if int(self._sp_phase) != int(getattr(self, "_sp_last_phase", 0)):
1277
+ self._sp_last_phase = int(self._sp_phase)
1278
+ self._sp_phase_steps = 0
1279
+ self._sp_phase_steps += 1
1280
+ if self._sp_phase_steps > int(self.SP_PHASE_TIMEOUT) and self._sp_phase < 15:
1281
+ if int(self._sp_phase) == 4 and payload[2] < 0.26 \
1282
+ and self._sp_retries < self.SP_MAX_RETRIES:
1283
+ # a phase-4 timeout with the cube not held means the grasp
1284
+ # failed: retry it instead of "dropping" nothing into the bag
1285
+ self._sp_retries += 1
1286
+ self._sp_grasp_xyz = None
1287
+ self._sp_phase = 0
1288
+ else:
1289
+ self._sp_phase = int(self._sp_phase) + 1
1290
+ self._sp_wait = 0
1291
+ self._sp_phase_steps = 0
1292
+
1293
+ phase = int(self._sp_phase)
1294
+ target = tcp_pos.copy()
1295
+ hand_delta = np.zeros(16, dtype=np.float32)
1296
+ max_delta = self.SP_MAX_DELTA
1297
+
1298
+ pay_top = payload[2] + 0.5 * self.PAYLOAD_SIZE
1299
+
1300
+ # freeze the grasp point when the approach starts: chasing the live
1301
+ # payload position turns any finger graze into a positive feedback
1302
+ # loop (hand pushes cube, cube moves, hand follows) that sweeps the
1303
+ # cube off the shelf
1304
+ if phase == 0 and self._sp_grasp_xyz is None:
1305
+ self._sp_grasp_xyz = np.array(
1306
+ [payload[0] + self.SP_GRASP_X_OFF,
1307
+ payload[1] + self.SP_GRASP_Y_OFF, pay_top], np.float32)
1308
+ g = self._sp_grasp_xyz if self._sp_grasp_xyz is not None else np.array(
1309
+ [payload[0] + self.SP_GRASP_X_OFF,
1310
+ payload[1] + self.SP_GRASP_Y_OFF, pay_top], np.float32)
1311
+
1312
+ if phase == 0:
1313
+ target = _palm_to_tcp(np.array([g[0], g[1], g[2] + self.SP_PAYLOAD_APPROACH_DZ], np.float32))
1314
+ hand_delta = _release_delta()
1315
+ if np.linalg.norm(tcp_pos - target) < self.SP_POS_THRESH:
1316
+ self._sp_phase = 1
1317
+ self._sp_wait = 0
1318
+ elif phase == 1:
1319
+ # pre-rotate the thumb at height, BEFORE the descent: the thumb
1320
+ # swing at cube level bats the cube away
1321
+ target = _palm_to_tcp(np.array([g[0], g[1], g[2] + self.SP_PAYLOAD_APPROACH_DZ], np.float32))
1322
+ hand_delta[12] = (2.8 - current_hand_qpos[12]) * self.SP_HAND_SPEED
1323
+ self._sp_wait += 1
1324
+ if self._sp_wait >= self.SP_THUMB_STEPS:
1325
+ self._sp_phase = 2
1326
+ self._sp_wait = 0
1327
+ elif phase == 2:
1328
+ target = _palm_to_tcp(np.array([g[0], g[1], g[2] + self.SP_PAYLOAD_GRASP_DZ], np.float32))
1329
+ max_delta = self.SP_DESCEND_MAX_DELTA
1330
+ if np.linalg.norm(tcp_pos - target) < self.SP_POS_THRESH:
1331
+ self._sp_phase = 3
1332
+ self._sp_wait = 0
1333
+ elif phase == 3:
1334
+ target = _palm_to_tcp(np.array([g[0], g[1], g[2] + self.SP_PAYLOAD_GRASP_DZ], np.float32))
1335
+ hand_delta = _grasp_delta()
1336
+ self._sp_wait += 1
1337
+ if self._sp_wait >= self.SP_GRASP_STEPS:
1338
+ self._sp_phase = 4
1339
+ self._sp_wait = 0
1340
+ elif phase == 4:
1341
+ drop_palm_z = max(bag_top_z + self.SP_DROP_ABOVE_MOUTH + 0.05, self.SP_CARRY_PALM_Z)
1342
+ cur_palm = tcp_pos + off
1343
+ hand_delta = _grasp_delta()
1344
+ max_delta = 0.08 # carry slowly (~3 mm/step) or the caged cube
1345
+ # is flung out
1346
+ shelf_top = float(self.SHELF_CENTER[2]) + 0.5 * float(self.SHELF_SIZE[2])
1347
+ # The free handle stands up at the robot-side rim, on the way to
1348
+ # the mouth, and the settle holds keep its cloth position-clamped:
1349
+ # a cube dragged over it hits what is in effect a rigid wall and
1350
+ # was flung out of the cage (3 of 18 episodes). Carry the cube's
1351
+ # bottom SP_CARRY_CLEAR above the handle's top.
1352
+ if payload[2] > shelf_top + 0.5 * self.PAYLOAD_SIZE + 0.01 \
1353
+ and np.linalg.norm(payload - cur_palm) < 0.16:
1354
+ free_top = float(P[self._handle_idx[free].to(P.device), 2].max().item())
1355
+ cube_z = free_top + 0.5 * self.PAYLOAD_SIZE + self.SP_CARRY_CLEAR
1356
+ drop_palm_z = max(drop_palm_z, float(cur_palm[2] + (cube_z - payload[2])))
1357
+ if cur_palm[2] < shelf_top + self.PAYLOAD_SIZE + 0.08:
1358
+ # lift STRAIGHT UP first; a diagonal pull-out flicks the cube
1359
+ # out of the cage
1360
+ target = _palm_to_tcp(np.array(
1361
+ [cur_palm[0], cur_palm[1], drop_palm_z], np.float32))
1362
+ else:
1363
+ # steer the CUBE over the mouth, not the palm: the caged cube
1364
+ # rides ~10 cm toward the knuckles, so palm-over-mouth drops
1365
+ # it a hand-width off target
1366
+ if payload[2] > 0.26:
1367
+ cube_off = payload[:2] - cur_palm[:2]
1368
+ else:
1369
+ cube_off = np.zeros(2, np.float32)
1370
+ target = _palm_to_tcp(np.array(
1371
+ [mouth_xy[0] - cube_off[0], mouth_xy[1] - cube_off[1],
1372
+ drop_palm_z], np.float32))
1373
+ cur_palm2 = tcp_pos + off
1374
+ held = (payload[2] > 0.26) and (np.linalg.norm(payload - cur_palm2) < 0.16)
1375
+ cube_on_target = held and (np.linalg.norm(payload[:2] - mouth_xy) < 0.03)
1376
+ if cube_on_target or (np.linalg.norm(tcp_pos - target) < 1.5 * self.SP_POS_THRESH):
1377
+ # gate the release on the CUBE being over the mouth, and
1378
+ # verify it actually came along (still caged near the palm,
1379
+ # well above the shelf); otherwise retry from the cube's
1380
+ # CURRENT position. Checking only "fell to the table" misses
1381
+ # the common failure of the cube left ON the shelf.
1382
+ if not held and self._sp_retries < self.SP_MAX_RETRIES:
1383
+ self._sp_retries += 1
1384
+ self._sp_grasp_xyz = None
1385
+ self._sp_phase = 0
1386
+ elif cube_on_target or not held:
1387
+ self._sp_phase = 5
1388
+ self._sp_wait = 0
1389
+ elif phase == 5:
1390
+ hand_delta = _release_delta()
1391
+ self._sp_wait += 1
1392
+ if self._sp_wait >= self.SP_DROP_STEPS:
1393
+ self._sp_phase = 6
1394
+ self._sp_wait = 0
1395
+ elif phase == 6:
1396
+ target = _palm_to_tcp(np.array([mouth_xy[0], mouth_xy[1] + 0.10, bag_top_z + 0.28], np.float32))
1397
+ hand_delta = _release_delta()
1398
+ if np.linalg.norm(tcp_pos - target) < 2.0 * self.SP_POS_THRESH:
1399
+ self._sp_phase = 7
1400
+ else:
1401
+ # ---- phases 7-15: hook the free handle with the thumb ----
1402
+ kn = self._thumb_knuckle_np()
1403
+ crook = self._thumb_crook_np()
1404
+ arch = self._free_arch_np()
1405
+ peg_tip = np.asarray(self._peg_tip, np.float32)
1406
+ ref = kn # the hand point steered to kn_goal
1407
+ kn_goal = kn.copy()
1408
+ hand_goal = self.SP_HOOK_HAND
1409
+ x_adj = float(self.SP_HOOK_X_ADJ[self._sp_ins_tries % len(self.SP_HOOK_X_ADJ)])
1410
+ if phase == 7:
1411
+ # shape the inward L-hook high above the free handle
1412
+ if self._sp_wait == 0 or self._sp_band is None:
1413
+ self._sp_band = self._hook_band()
1414
+ h = self._sp_band["hole"]
1415
+ kn_goal = np.array([h[0] + x_adj, h[1] + 0.10, self._sp_band["z_hi"] + 0.12], np.float32)
1416
+ max_delta = 0.2
1417
+ self._sp_wait += 1
1418
+ shaped = float(np.abs(current_hand_qpos - self.SP_HOOK_HAND).max()) < 0.1
1419
+ if (shaped and np.linalg.norm(kn - kn_goal) < 0.02) or self._sp_wait > 40:
1420
+ self._sp_phase, self._sp_wait, self._sp_near = 8, 0, 0
1421
+ elif phase == 8:
1422
+ # knuckle at the threading height, in front of the arch
1423
+ band = self._sp_band
1424
+ h = band["hole"]
1425
+ kn_goal = np.array([h[0] + x_adj, band["near_y"] + self.SP_HOOK_FRONT, band["z_kn"]], np.float32)
1426
+ max_delta = 0.2
1427
+ self._sp_wait += 1
1428
+ self._sp_near = self._sp_near + 1 if np.linalg.norm(kn - kn_goal) < 0.01 else 0
1429
+ if self._sp_near >= 3 or self._sp_wait > 35:
1430
+ self._sp_hole = kn_goal.copy() # insertion start
1431
+ self._sp_phase, self._sp_wait = 9, 0
1432
+ elif phase == 9:
1433
+ # slide the distal phalanx through the hole (-y), under the arch
1434
+ band = self._sp_band
1435
+ ins = np.array([self._sp_hole[0], band["mean_y"] + self.SP_HOOK_INS, band["z_kn"]], np.float32)
1436
+ s_in = min(1.0, (self._sp_wait + 1) / 25.0)
1437
+ kn_goal = self._sp_hole + (ins - self._sp_hole) * s_in
1438
+ max_delta = 0.2
1439
+ if self._sp_wait == 0:
1440
+ self._sp_kn_hist = []
1441
+ self._sp_kn_hist.append(float(kn[1]))
1442
+ self._sp_wait += 1
1443
+ stalled = (len(self._sp_kn_hist) > 12 and s_in >= 1.0
1444
+ and self._sp_kn_hist[-9] - self._sp_kn_hist[-1] < 0.003)
1445
+ if float(kn[1]) <= float(ins[1]) + 0.004:
1446
+ self._sp_hold_pos = tcp_pos.copy()
1447
+ self._sp_phase, self._sp_wait = 10, 0
1448
+ elif stalled or self._sp_wait > 45:
1449
+ # jammed on the loop, the bag wall or the peg: back out
1450
+ # and try a line shifted sideways
1451
+ self._sp_ins_tries += 1
1452
+ self._retry_hook_insertion()
1453
+ elif phase == 10:
1454
+ # curl the thumb up toward the palm, wrist held still: the
1455
+ # knuckle swings ~2.6 cm toward the wall and ends under the
1456
+ # arch. The env drops its settle holds once the curl has
1457
+ # closed around the handle.
1458
+ hand_goal = self.SP_CURL_HAND
1459
+ max_delta = 0.06
1460
+ self._sp_wait += 1
1461
+ if not self._rim_pinned and self._hand_supports_free_handle():
1462
+ self._sp_hold_pos = None
1463
+ self._sp_arch_z0 = float(arch[2])
1464
+ self._sp_phase, self._sp_wait = 11, 0
1465
+ elif self._sp_wait > 25 and getattr(self, "_support_streak", 0) == 0:
1466
+ # the curl closed on nothing: open up and insert again
1467
+ # (not while the env is one step from releasing the holds)
1468
+ self._sp_ins_tries += 1
1469
+ self._sp_hold_pos = None
1470
+ self._retry_hook_insertion()
1471
+ elif phase == 11:
1472
+ # let the handle settle into the V, then lift it
1473
+ hand_goal = self.SP_CURL_HAND
1474
+ if self._sp_wait == 0:
1475
+ self._sp_ret_start = kn.copy()
1476
+ s_up = min(1.0, max(0, self._sp_wait - 11) / 30.0)
1477
+ kn_goal = self._sp_ret_start + np.array([0.0, 0.0, self.SP_HOOK_LIFT * s_up], np.float32)
1478
+ max_delta = 0.06
1479
+ self._sp_wait += 1
1480
+ if self._sp_wait >= 46:
1481
+ self._sp_phase, self._sp_wait = 12, 0
1482
+ elif phase == 12:
1483
+ # carry the handle straight to the hand-over pose, the hook
1484
+ # still closed and the fingers folded out of the way
1485
+ hand_goal = self.SP_CARRY_HAND
1486
+ _, _, _, n_t, _ = self._peg_frame()
1487
+ tip_goal = peg_tip + n_t.detach().cpu().numpy().astype(np.float32) * self.SP_HO_HOVER
1488
+ goal = (tip_goal - self.UNCURL_TIP_FROM_KNUCKLE).astype(np.float32)
1489
+ # Use position feedback throughout the carry, so demonstrations
1490
+ # steer toward the same hand-over pose from intermediate states
1491
+ # without an unobserved interpolation clock.
1492
+ kn_goal = goal
1493
+ self._sp_hole = goal
1494
+ max_delta = 0.06
1495
+ self._sp_wait += 1
1496
+ if np.linalg.norm(kn - goal) < 0.015 or self._sp_wait > 60:
1497
+ self._sp_phase, self._sp_wait = 13, 0
1498
+ elif phase == 13:
1499
+ # open the thumb in place: the env slides the loop off it
1500
+ # onto the peg (HANDOVER_*)
1501
+ hand_goal = self.SP_UNCURL_HAND
1502
+ kn_goal = self._sp_hole.copy()
1503
+ max_delta = 0.06
1504
+ self._sp_wait += 1
1505
+ if (self._handed_over and self._sp_wait >= 8) or self._sp_wait > 40:
1506
+ self._sp_ret_start = kn.copy()
1507
+ self._sp_phase, self._sp_wait = 14, 0
1508
+ elif phase == 14:
1509
+ # pull the thumb back out of the loop along +y
1510
+ hand_goal = self.SP_UNCURL_HAND
1511
+ s_out = min(1.0, (self._sp_wait + 1) / 25.0)
1512
+ kn_goal = self._sp_ret_start + np.array([0.0, self.SP_HOOK_RETRACT * s_out, 0.0], np.float32)
1513
+ max_delta = 0.06
1514
+ self._sp_wait += 1
1515
+ if self._sp_wait >= 28:
1516
+ self._sp_phase, self._sp_wait = 15, 0
1517
+ else:
1518
+ # up and away (+y) until every fingertip is clear of the hook
1519
+ hand_goal = self.SP_UNCURL_HAND
1520
+ ref = tcp_pos
1521
+ kn_goal = tcp_pos.copy()
1522
+ if self._fingertip_dist(self._hook_hang_world()) < self.HAND_CLEAR_DIST + 0.05:
1523
+ kn_goal = kn_goal + np.array([0.0, 0.06, 0.03], np.float32)
1524
+ max_delta = 0.12
1525
+ self._sp_wait += 1
1526
+ if self._sp_wait == 40 and self._sp_hang_retries < self.SP_MAX_HANG_RETRIES \
1527
+ and self._ho_idx is None:
1528
+ if float(np.linalg.norm(arch - pH)) > 0.09:
1529
+ self._sp_hang_retries += 1
1530
+ self._sp_band = None
1531
+ self._sp_phase, self._sp_wait, self._sp_near = 7, 0, 0
1532
+ # the handle fell out of the hook while lifting / carrying: hook it again
1533
+ # (never once the env has taken the loop for the hand-over)
1534
+ if phase in (11, 12) and self._sp_wait > 3 \
1535
+ and self._sp_hang_retries < self.SP_MAX_HANG_RETRIES \
1536
+ and self._ho_idx is None \
1537
+ and float(np.linalg.norm(self._free_arch_core_np() - crook)) > 0.07:
1538
+ self._sp_hang_retries += 1
1539
+ self._sp_band = None
1540
+ self._sp_phase, self._sp_wait, self._sp_near = 7, 0, 0
1541
+ hand_delta = np.clip((hand_goal - current_hand_qpos) / 0.1 * 0.5,
1542
+ -1.0, 1.0).astype(np.float32)
1543
+ if self._sp_hold_pos is not None and int(self._sp_phase) == 10:
1544
+ target = self._sp_hold_pos.astype(np.float32)
1545
+ else:
1546
+ target = (tcp_pos + (kn_goal - ref)).astype(np.float32)
1547
+
1548
+ rot_err = Rot.from_matrix(self._sp_R_ref @ R.T).as_rotvec()
1549
+ rot_cmd = np.clip(-2.0 * rot_err, -1.0, 1.0).astype(np.float32)
1550
+ action = np.zeros(22, dtype=np.float32)
1551
+ pos_delta = (target - tcp_pos) * self.SP_GAIN
1552
+ action[:3] = np.clip(pos_delta, -max_delta, max_delta)
1553
+ action[3:6] = rot_cmd
1554
+ action[6:] = hand_delta.astype(np.float32)
1555
+ return action
1556
+
1557
+ # ------------------------------------------------------------------ #
1558
+ # Rewards
1559
+ # ------------------------------------------------------------------ #
1560
+ def compute_dense_reward(self, obs: Any, action: torch.Tensor, info: Dict[str, torch.Tensor]) -> torch.Tensor:
1561
+ dev = self.device
1562
+ d0 = info.get("d_handle0", torch.tensor([1.0], device=dev)).view(-1)[0]
1563
+ d1 = info.get("d_handle1", torch.tensor([1.0], device=dev)).view(-1)[0]
1564
+ inside = bool(info.get("inside", torch.tensor([0.0], device=dev)).item() > 0.5)
1565
+ success = bool(info.get("success", torch.tensor([False], device=dev)).item())
1566
+
1567
+ payload = self._payload_pos()
1568
+ bag_com = info.get("bag_com", torch.zeros(1, 3, device=dev)).view(3)
1569
+ d_pay = torch.linalg.norm(payload - bag_com)
1570
+ pay_prog = 1.0 - torch.tanh(3.0 * d_pay)
1571
+ handle_prog = 0.5 * ((1.0 - torch.tanh(8.0 * d0)) + (1.0 - torch.tanh(8.0 * d1)))
1572
+
1573
+ r = self.W_PAYLOAD * pay_prog + self.W_HANDLE * handle_prog
1574
+ if inside:
1575
+ r = r + self.INSIDE_BONUS
1576
+ if success:
1577
+ r = r + self.SUCCESS_BONUS
1578
+ r = r - self.TIME_PENALTY
1579
+ if self.NORMALIZE_REWARD:
1580
+ lo, hi = self.REWARD_CLIP
1581
+ r = torch.clamp(r, min=lo, max=hi)
1582
+ return r.to(dtype=torch.float32, device=dev).view(1)
1583
+
1584
+ def compute_normalized_dense_reward(self, obs: Any, action: torch.Tensor, info: Dict[str, torch.Tensor]) -> torch.Tensor:
1585
+ return self.compute_dense_reward(obs, action, info) / max(self.REWARD_CLIP[1], 1e-6)
1586
+
1587
+ def compute_sparse_reward(self, obs: Any, action: torch.Tensor, info: Dict[str, torch.Tensor]) -> torch.Tensor:
1588
+ return torch.tensor([1.0 if bool(info["success"].item()) else 0.0], dtype=torch.float32, device=self.device)
small_flow_matching_policy/BagDropIn-v1/eval_metrics.json ADDED
@@ -0,0 +1,475 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint": "ckpt_final.pt",
3
+ "episodes": 50,
4
+ "successes": 4.0,
5
+ "success_rate": 0.08,
6
+ "episode_seed_base": 200000,
7
+ "waves": [
8
+ [
9
+ 12,
10
+ 12
11
+ ],
12
+ [
13
+ 13,
14
+ 13
15
+ ]
16
+ ],
17
+ "records": [
18
+ {
19
+ "index": 0,
20
+ "seed": 200000,
21
+ "success": 0.0,
22
+ "return": 947.4615430831909,
23
+ "video": "evaluation_videos/seed_200000_succ0.mp4"
24
+ },
25
+ {
26
+ "index": 1,
27
+ "seed": 200001,
28
+ "success": 0.0,
29
+ "return": 526.417573094368,
30
+ "video": "evaluation_videos/seed_200001_succ0.mp4"
31
+ },
32
+ {
33
+ "index": 2,
34
+ "seed": 200002,
35
+ "success": 0.0,
36
+ "return": 372.8313665986061,
37
+ "video": "evaluation_videos/seed_200002_succ0.mp4"
38
+ },
39
+ {
40
+ "index": 3,
41
+ "seed": 200003,
42
+ "success": 0.0,
43
+ "return": 447.9537578821182,
44
+ "video": "evaluation_videos/seed_200003_succ0.mp4"
45
+ },
46
+ {
47
+ "index": 4,
48
+ "seed": 200004,
49
+ "success": 0.0,
50
+ "return": 956.9457547068596,
51
+ "video": "evaluation_videos/seed_200004_succ0.mp4"
52
+ },
53
+ {
54
+ "index": 5,
55
+ "seed": 200005,
56
+ "success": 0.0,
57
+ "return": 387.26814609766006,
58
+ "video": "evaluation_videos/seed_200005_succ0.mp4"
59
+ },
60
+ {
61
+ "index": 6,
62
+ "seed": 200006,
63
+ "success": 0.0,
64
+ "return": 492.5518888235092,
65
+ "video": "evaluation_videos/seed_200006_succ0.mp4"
66
+ },
67
+ {
68
+ "index": 7,
69
+ "seed": 200007,
70
+ "success": 0.0,
71
+ "return": 901.3089085221291,
72
+ "video": "evaluation_videos/seed_200007_succ0.mp4"
73
+ },
74
+ {
75
+ "index": 8,
76
+ "seed": 200008,
77
+ "success": 0.0,
78
+ "return": 441.980166465044,
79
+ "video": "evaluation_videos/seed_200008_succ0.mp4"
80
+ },
81
+ {
82
+ "index": 9,
83
+ "seed": 200009,
84
+ "success": 0.0,
85
+ "return": 405.68091064691544,
86
+ "video": "evaluation_videos/seed_200009_succ0.mp4"
87
+ },
88
+ {
89
+ "index": 10,
90
+ "seed": 200010,
91
+ "success": 0.0,
92
+ "return": 352.29464316368103,
93
+ "video": "evaluation_videos/seed_200010_succ0.mp4"
94
+ },
95
+ {
96
+ "index": 11,
97
+ "seed": 200011,
98
+ "success": 0.0,
99
+ "return": 938.030930519104,
100
+ "video": "evaluation_videos/seed_200011_succ0.mp4"
101
+ },
102
+ {
103
+ "index": 12,
104
+ "seed": 200012,
105
+ "success": 0.0,
106
+ "return": 443.4338427782059,
107
+ "video": "evaluation_videos/seed_200012_succ0.mp4"
108
+ },
109
+ {
110
+ "index": 13,
111
+ "seed": 200013,
112
+ "success": 1.0,
113
+ "return": 523.3683533668518,
114
+ "video": "evaluation_videos/seed_200013_succ1.mp4"
115
+ },
116
+ {
117
+ "index": 14,
118
+ "seed": 200014,
119
+ "success": 0.0,
120
+ "return": 933.4794335961342,
121
+ "video": "evaluation_videos/seed_200014_succ0.mp4"
122
+ },
123
+ {
124
+ "index": 15,
125
+ "seed": 200015,
126
+ "success": 0.0,
127
+ "return": 411.8796911537647,
128
+ "video": "evaluation_videos/seed_200015_succ0.mp4"
129
+ },
130
+ {
131
+ "index": 16,
132
+ "seed": 200016,
133
+ "success": 0.0,
134
+ "return": 532.6111829280853,
135
+ "video": "evaluation_videos/seed_200016_succ0.mp4"
136
+ },
137
+ {
138
+ "index": 17,
139
+ "seed": 200017,
140
+ "success": 0.0,
141
+ "return": 936.5249073505402,
142
+ "video": "evaluation_videos/seed_200017_succ0.mp4"
143
+ },
144
+ {
145
+ "index": 18,
146
+ "seed": 200018,
147
+ "success": 0.0,
148
+ "return": 485.4649856686592,
149
+ "video": "evaluation_videos/seed_200018_succ0.mp4"
150
+ },
151
+ {
152
+ "index": 19,
153
+ "seed": 200019,
154
+ "success": 0.0,
155
+ "return": 343.7198933660984,
156
+ "video": "evaluation_videos/seed_200019_succ0.mp4"
157
+ },
158
+ {
159
+ "index": 20,
160
+ "seed": 200020,
161
+ "success": 0.0,
162
+ "return": 327.3629849255085,
163
+ "video": "evaluation_videos/seed_200020_succ0.mp4"
164
+ },
165
+ {
166
+ "index": 21,
167
+ "seed": 200021,
168
+ "success": 0.0,
169
+ "return": 507.4706324338913,
170
+ "video": "evaluation_videos/seed_200021_succ0.mp4"
171
+ },
172
+ {
173
+ "index": 22,
174
+ "seed": 200022,
175
+ "success": 0.0,
176
+ "return": 544.6768824458122,
177
+ "video": "evaluation_videos/seed_200022_succ0.mp4"
178
+ },
179
+ {
180
+ "index": 23,
181
+ "seed": 200023,
182
+ "success": 0.0,
183
+ "return": 931.1281852722168,
184
+ "video": "evaluation_videos/seed_200023_succ0.mp4"
185
+ },
186
+ {
187
+ "index": 24,
188
+ "seed": 200024,
189
+ "success": 0.0,
190
+ "return": 939.2310606837273,
191
+ "video": "evaluation_videos/seed_200024_succ0.mp4"
192
+ },
193
+ {
194
+ "index": 25,
195
+ "seed": 200025,
196
+ "success": 0.0,
197
+ "return": 353.88437882065773,
198
+ "video": "evaluation_videos/seed_200025_succ0.mp4"
199
+ },
200
+ {
201
+ "index": 26,
202
+ "seed": 200026,
203
+ "success": 0.0,
204
+ "return": 894.1958553195,
205
+ "video": "evaluation_videos/seed_200026_succ0.mp4"
206
+ },
207
+ {
208
+ "index": 27,
209
+ "seed": 200027,
210
+ "success": 0.0,
211
+ "return": 931.1859244704247,
212
+ "video": "evaluation_videos/seed_200027_succ0.mp4"
213
+ },
214
+ {
215
+ "index": 28,
216
+ "seed": 200028,
217
+ "success": 1.0,
218
+ "return": 535.7420411705971,
219
+ "video": "evaluation_videos/seed_200028_succ1.mp4"
220
+ },
221
+ {
222
+ "index": 29,
223
+ "seed": 200029,
224
+ "success": 0.0,
225
+ "return": 864.7309737205505,
226
+ "video": "evaluation_videos/seed_200029_succ0.mp4"
227
+ },
228
+ {
229
+ "index": 30,
230
+ "seed": 200030,
231
+ "success": 0.0,
232
+ "return": 508.72062516212463,
233
+ "video": "evaluation_videos/seed_200030_succ0.mp4"
234
+ },
235
+ {
236
+ "index": 31,
237
+ "seed": 200031,
238
+ "success": 0.0,
239
+ "return": 946.8823100328445,
240
+ "video": "evaluation_videos/seed_200031_succ0.mp4"
241
+ },
242
+ {
243
+ "index": 32,
244
+ "seed": 200032,
245
+ "success": 0.0,
246
+ "return": 876.6623591780663,
247
+ "video": "evaluation_videos/seed_200032_succ0.mp4"
248
+ },
249
+ {
250
+ "index": 33,
251
+ "seed": 200033,
252
+ "success": 1.0,
253
+ "return": 539.879998922348,
254
+ "video": "evaluation_videos/seed_200033_succ1.mp4"
255
+ },
256
+ {
257
+ "index": 34,
258
+ "seed": 200034,
259
+ "success": 0.0,
260
+ "return": 346.2945433855057,
261
+ "video": "evaluation_videos/seed_200034_succ0.mp4"
262
+ },
263
+ {
264
+ "index": 35,
265
+ "seed": 200035,
266
+ "success": 0.0,
267
+ "return": 512.411803483963,
268
+ "video": "evaluation_videos/seed_200035_succ0.mp4"
269
+ },
270
+ {
271
+ "index": 36,
272
+ "seed": 200036,
273
+ "success": 0.0,
274
+ "return": 852.1819575428963,
275
+ "video": "evaluation_videos/seed_200036_succ0.mp4"
276
+ },
277
+ {
278
+ "index": 37,
279
+ "seed": 200037,
280
+ "success": 1.0,
281
+ "return": 527.2338510751724,
282
+ "video": "evaluation_videos/seed_200037_succ1.mp4"
283
+ },
284
+ {
285
+ "index": 38,
286
+ "seed": 200038,
287
+ "success": 0.0,
288
+ "return": 333.4079418182373,
289
+ "video": "evaluation_videos/seed_200038_succ0.mp4"
290
+ },
291
+ {
292
+ "index": 39,
293
+ "seed": 200039,
294
+ "success": 0.0,
295
+ "return": 437.03912633657455,
296
+ "video": "evaluation_videos/seed_200039_succ0.mp4"
297
+ },
298
+ {
299
+ "index": 40,
300
+ "seed": 200040,
301
+ "success": 0.0,
302
+ "return": 460.15217238664627,
303
+ "video": "evaluation_videos/seed_200040_succ0.mp4"
304
+ },
305
+ {
306
+ "index": 41,
307
+ "seed": 200041,
308
+ "success": 0.0,
309
+ "return": 414.4361147284508,
310
+ "video": "evaluation_videos/seed_200041_succ0.mp4"
311
+ },
312
+ {
313
+ "index": 42,
314
+ "seed": 200042,
315
+ "success": 0.0,
316
+ "return": 554.9326196908951,
317
+ "video": "evaluation_videos/seed_200042_succ0.mp4"
318
+ },
319
+ {
320
+ "index": 43,
321
+ "seed": 200043,
322
+ "success": 0.0,
323
+ "return": 555.0461828708649,
324
+ "video": "evaluation_videos/seed_200043_succ0.mp4"
325
+ },
326
+ {
327
+ "index": 44,
328
+ "seed": 200044,
329
+ "success": 0.0,
330
+ "return": 946.2699078321457,
331
+ "video": "evaluation_videos/seed_200044_succ0.mp4"
332
+ },
333
+ {
334
+ "index": 45,
335
+ "seed": 200045,
336
+ "success": 0.0,
337
+ "return": 505.368553340435,
338
+ "video": "evaluation_videos/seed_200045_succ0.mp4"
339
+ },
340
+ {
341
+ "index": 46,
342
+ "seed": 200046,
343
+ "success": 0.0,
344
+ "return": 627.7252280712128,
345
+ "video": "evaluation_videos/seed_200046_succ0.mp4"
346
+ },
347
+ {
348
+ "index": 47,
349
+ "seed": 200047,
350
+ "success": 0.0,
351
+ "return": 442.4451742172241,
352
+ "video": "evaluation_videos/seed_200047_succ0.mp4"
353
+ },
354
+ {
355
+ "index": 48,
356
+ "seed": 200048,
357
+ "success": 0.0,
358
+ "return": 522.9267926216125,
359
+ "video": "evaluation_videos/seed_200048_succ0.mp4"
360
+ },
361
+ {
362
+ "index": 49,
363
+ "seed": 200049,
364
+ "success": 0.0,
365
+ "return": 382.2829709649086,
366
+ "video": "evaluation_videos/seed_200049_succ0.mp4"
367
+ }
368
+ ],
369
+ "selected_checkpoint_step": 30000,
370
+ "eval/success_mean": 0.08,
371
+ "eval/episode_success": [
372
+ 0.0,
373
+ 0.0,
374
+ 0.0,
375
+ 0.0,
376
+ 0.0,
377
+ 0.0,
378
+ 0.0,
379
+ 0.0,
380
+ 0.0,
381
+ 0.0,
382
+ 0.0,
383
+ 0.0,
384
+ 0.0,
385
+ 1.0,
386
+ 0.0,
387
+ 0.0,
388
+ 0.0,
389
+ 0.0,
390
+ 0.0,
391
+ 0.0,
392
+ 0.0,
393
+ 0.0,
394
+ 0.0,
395
+ 0.0,
396
+ 0.0,
397
+ 0.0,
398
+ 0.0,
399
+ 0.0,
400
+ 1.0,
401
+ 0.0,
402
+ 0.0,
403
+ 0.0,
404
+ 0.0,
405
+ 1.0,
406
+ 0.0,
407
+ 0.0,
408
+ 0.0,
409
+ 1.0,
410
+ 0.0,
411
+ 0.0,
412
+ 0.0,
413
+ 0.0,
414
+ 0.0,
415
+ 0.0,
416
+ 0.0,
417
+ 0.0,
418
+ 0.0,
419
+ 0.0,
420
+ 0.0,
421
+ 0.0
422
+ ],
423
+ "eval/episode_return": [
424
+ 947.4615430831909,
425
+ 526.417573094368,
426
+ 372.8313665986061,
427
+ 447.9537578821182,
428
+ 956.9457547068596,
429
+ 387.26814609766006,
430
+ 492.5518888235092,
431
+ 901.3089085221291,
432
+ 441.980166465044,
433
+ 405.68091064691544,
434
+ 352.29464316368103,
435
+ 938.030930519104,
436
+ 443.4338427782059,
437
+ 523.3683533668518,
438
+ 933.4794335961342,
439
+ 411.8796911537647,
440
+ 532.6111829280853,
441
+ 936.5249073505402,
442
+ 485.4649856686592,
443
+ 343.7198933660984,
444
+ 327.3629849255085,
445
+ 507.4706324338913,
446
+ 544.6768824458122,
447
+ 931.1281852722168,
448
+ 939.2310606837273,
449
+ 353.88437882065773,
450
+ 894.1958553195,
451
+ 931.1859244704247,
452
+ 535.7420411705971,
453
+ 864.7309737205505,
454
+ 508.72062516212463,
455
+ 946.8823100328445,
456
+ 876.6623591780663,
457
+ 539.879998922348,
458
+ 346.2945433855057,
459
+ 512.411803483963,
460
+ 852.1819575428963,
461
+ 527.2338510751724,
462
+ 333.4079418182373,
463
+ 437.03912633657455,
464
+ 460.15217238664627,
465
+ 414.4361147284508,
466
+ 554.9326196908951,
467
+ 555.0461828708649,
468
+ 946.2699078321457,
469
+ 505.368553340435,
470
+ 627.7252280712128,
471
+ 442.4451742172241,
472
+ 522.9267926216125,
473
+ 382.2829709649086
474
+ ]
475
+ }
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200000_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:190783f9b9e7529a0ec46569adf5c4794868f5154cb43b19b83fc61f1e9f7aa0
3
+ size 3264118
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200001_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a90286ed1713ac357310234065a2696bc5fd6c511883eccd8391752f3338ae9
3
+ size 3783945
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200002_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bd1863802cc69ab556e69623865d192a2cf45b79c6d5f1b2a5c17013e14a5ab2
3
+ size 4493562
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200003_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ece794627193fc7bf1a2fa281b314b383aedd2bd93381d20a9ab191e91af5fde
3
+ size 3600186
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200004_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61baf86a76e4a54fb8695baa0d7e7773646c31814b06147fe665c4d0b7272451
3
+ size 3429978
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200005_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:660f7e62ebc020113447d03eb493bab81953191a30f00c7d3899a442ed4373b8
3
+ size 3417354
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200006_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f2eb34f02fdce46b678a468523affa08b8dfb877362d8908e75351cd0c780c72
3
+ size 3666759
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200007_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:206a29f5234deaadea078ca7d12d15b797e81fca64977998f31363a69263404e
3
+ size 3238840
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200008_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f0b7c38e53f05b33e89e0ce251774dbc8d41abe86c094e4b9394158c4f7227de
3
+ size 4125735
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200009_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:663837c5c85af68467adec0e6466f1a641df85fe4eac48f70f3e588bc12ce9fc
3
+ size 3501552
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200010_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:645a49b26ee5717eb284f18ca605fb3d2a2c72164574465a75945161fbc9f389
3
+ size 4268729
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200011_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4740eafc38447d916727975ae8c9a17de2330e4f9e5e85e82e266fa5e06ca9fd
3
+ size 3456109
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200012_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a2989d3d70a929dca37c035784ef28834676e7c27cdcf2f95e2f20fab80d818
3
+ size 3614493
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200013_succ1.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:15b7c06e0ffcc718ddfffe4097c66c279fac605deb0b215583750feae1a770d6
3
+ size 2258291
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200014_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d6cfddf9e228f7bd136322deb99ca5b52ae91a499b870e27cf4c38d2ef7b7328
3
+ size 3299803
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200015_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0fde98f2d9172ee1918823afb070fd9d2282eceb08cdc0c3e73bebfd2408ea87
3
+ size 4309718
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200016_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e0d2fbc5be96b6e3474bc1d31d50215003537780f25005134abd351e82eb7ccd
3
+ size 3641302
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200017_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c929ceb60173a2315049bcd184955553abe9a99befd6d37ab9be9438893c824f
3
+ size 2983719
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200018_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a8a656062f53e07b28be976a2edb44274dfaca9137e01048565d057b3b4efb2
3
+ size 3848147
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200019_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb9c68fff7f6ba5146132e1579f444ee20e793a6a6da7e68fbf70dcb098ca7c2
3
+ size 4243200
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200020_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4de1380d394fbd1d6aca218f3e1e77f1039384030e1b471fb0a3d8cf6696d190
3
+ size 3631050
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200021_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c41d8176dfadc4293665b19e09b0d1b1988465e57b29721a6c952b5ddd6c744
3
+ size 3495015
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200022_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d28ba1327a664537c5a5aae848714b7eecfda0aa368c807c961682ca63d56064
3
+ size 3549856
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200023_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:860a763eb491c0d552a9cb7c5f2917d93d42d51ac4196c2a5eac20604d8f40f3
3
+ size 2824772
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200024_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ffa6c3bce52792b6be57b71795970ad1f30cdb71968c2ead1e9ccd79dea93a4
3
+ size 3711752
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200025_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:341e9ce00ac78bc7d6c25ed88135e1e067da9199407e0cc82e8263d43360ee7b
3
+ size 4298241
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200026_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2866a6349b64f8b11422acb85d29ea1a8265e3d38671e4c711ef70fd8db8b910
3
+ size 2638294
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200027_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:254bd50497173ff71fa20af57f52db138ed27ac116813f00682489b354104afc
3
+ size 3379445
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200028_succ1.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c657c63e50b4b877bdb63cf07b0f4de337bfaa9e10952e5ebaffa7e01e675f45
3
+ size 2261392
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200029_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d884d7cf131dc7be5bef4a6f5b6409f7a8cb287a9730a9354249f48a53b04427
3
+ size 3047494
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200030_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3ca0bc1851f367ffbbf1dd041529c626fe4547701b6b521f70b6aa22b66ff58d
3
+ size 3594885
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200031_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b4e72713d78e75527af2a9e79c62b12206006c09fc014297a5920cd6554a636a
3
+ size 3450788
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200032_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f06309657400350e1b85f745fa55e1ebfe9bdca5eb3b7004a5d8af3bc446c66e
3
+ size 2963344
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200033_succ1.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:db0f81350a443a48093959dee909790dc8220bee2bd266c63675bd83b3a911be
3
+ size 2417315
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200034_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:599d3ea8c6fb9f37016b307410c0d6755e74abe6c1a2b6ccc86ef8c708bea08c
3
+ size 4440753
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200035_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7bd482e44ee5b37d4a41ef3b6f09ebeac4990860a739bf94420407cb49b056b6
3
+ size 2673588
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200036_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f5199397a8210caef0b341d90b21c9c941aeb31d522a7257888dba6831cc7f2c
3
+ size 3850317
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200037_succ1.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aba0add24e901804a04d0e51472c6468fc76416741c66a7da23c69ade1b63fbf
3
+ size 2206454
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200038_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b22d32c390c0e16129d2e5182fe46dcbdc9f3735f0ac5f51f6685b882daf398f
3
+ size 4294076
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200039_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd580fc96ce5a546f2671c03f25826f7c94cf44927d3cb124c348de2f84330ee
3
+ size 3590257
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200040_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:545e6563c75c9db03b3c69c0123353b12d88d192df462c48fd6885b3c872edad
3
+ size 3556042
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200041_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:278d901bc4b1652e14d644b192c780c76ad2511c64ae5985209351c341f180e3
3
+ size 3519549
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200042_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3520fb2c1bdbc6246e6dd1a7421a6f7a4e92a1f6aff2812f96153faea1edc947
3
+ size 4282318
small_flow_matching_policy/BagDropIn-v1/evaluation_videos/seed_200043_succ0.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6a6c42451a7a71769e9c2dbbb29967651ee1b69ab3123f759d34b6c91fa93a48
3
+ size 4253006