Linslab commited on
Commit
ac3be32
·
verified ·
1 Parent(s): a73ee0b

Index 23 HF-data retraining PI0.5 peaks (pi05_hf21_20261025) with 50-episode evaluations

Browse files
README.md CHANGED
@@ -6,14 +6,14 @@ license: mit
6
 
7
  ## single_task
8
 
9
- One PI0.5 LoRA checkpoint per ManiLadder task (106 tasks), each finetuned on 20 demos.
10
  Checkpoints are selected by the highest **50-episode** success rate; ties use the earliest step.
11
 
12
  - The 60 previously published tasks use fixed evaluation seeds 1000–1049.
13
  - The 22 completed retraining runs published through 2026-10-30 use seeds 800000–800049. All reached 40,000 training steps. 15 have all eight 5k–40k evaluations; 7 have seven evaluations because one intermediate checkpoint was not retained. Their selected weights are the best among the available evaluated checkpoints.
14
- - The 21 tasks from the HF-data retraining campaign `pi05_hf21_20261025` (added 2026-10-29; raw demos freshly downloaded from Linslab/ManiLadder_Dataset) use seeds 1000–1049. All reached 40,000 training steps and have all eight 5k–40k evaluations. See [its manifest](single_task/pi05_hf21_20261025_manifest.json).
15
 
16
  - The 3 fresh-data runs in `pi05_newdata_3tasks_20261028` use pinned dataset revision `d120457e637b01679fbb40fd4de1d2394f0d55c9`, 20 demos per task, 40,000 training steps, and all eight 5k–40k eval50 checkpoints with seeds 800000–800049. **All three scored 0/50 at every checkpoint**; step 5,000 is the earliest tied checkpoint. See [their manifest](single_task/pi05_newdata_3tasks_20261028_manifest.json).
 
17
 
18
  See [retraining provenance and missing steps](single_task/pi05_retrain_eval50_20261022_manifest.json). These scores use different seed sets across the two groups and are not a controlled comparison.
19
 
@@ -24,7 +24,7 @@ Each `single_task/<task>/` folder contains:
24
  - `eval_result.json` — 50-episode evaluation of this checkpoint
25
 
26
  `single_task/results.csv` lists the selected checkpoint per task; `single_task/all_checkpoints_eval.csv` has the
27
- 50-episode result of every evaluated checkpoint (838 total).
28
 
29
  Evaluation: `exec_steps=16`, `sample_steps=10`. Tasks whose checkpoints all scored 0/50 are still included;
30
  for them the step choice is arbitrary.
@@ -107,6 +107,7 @@ for them the step choice is arbitrary.
107
  | HitBox-v1 | 3 | 25000 | 0.06 |
108
  | LiquidMeasuring-v1 | 3 | 35000 | 0.32 |
109
  | MobilePickPlace-v1 | 3 | 15000 | 0.08 |
 
110
  | PushTargetPose-v1 | 3 | 30000 | 0.02 |
111
  | SeparateTwoRopesEnv-v1 | 3 | 35000 | 0.26 |
112
  | Stack6Blocks-v1 | 3 | 5000 | 0.00 |
@@ -135,6 +136,7 @@ for them the step choice is arbitrary.
135
  | RodThroughHoles-v1 | 4 | 5000 | 0.02 |
136
  | SpinPenHora-v1 | 4 | 5000 | 0.00 |
137
  | StampRoll-v1 | 4 | 5000 | 0.00 |
 
138
  | TennisRally-v1 | 4 | 15000 | 0.06 |
139
  | WrapRope-v1 | 4 | 5000 | 0.02 |
140
 
 
6
 
7
  ## single_task
8
 
9
+ One PI0.5 LoRA checkpoint per ManiLadder task (108 tasks), each finetuned on 20 demos.
10
  Checkpoints are selected by the highest **50-episode** success rate; ties use the earliest step.
11
 
12
  - The 60 previously published tasks use fixed evaluation seeds 1000–1049.
13
  - The 22 completed retraining runs published through 2026-10-30 use seeds 800000–800049. All reached 40,000 training steps. 15 have all eight 5k–40k evaluations; 7 have seven evaluations because one intermediate checkpoint was not retained. Their selected weights are the best among the available evaluated checkpoints.
 
14
 
15
  - The 3 fresh-data runs in `pi05_newdata_3tasks_20261028` use pinned dataset revision `d120457e637b01679fbb40fd4de1d2394f0d55c9`, 20 demos per task, 40,000 training steps, and all eight 5k–40k eval50 checkpoints with seeds 800000–800049. **All three scored 0/50 at every checkpoint**; step 5,000 is the earliest tied checkpoint. See [their manifest](single_task/pi05_newdata_3tasks_20261028_manifest.json).
16
+ - The 23 tasks from the HF-data retraining campaign `pi05_hf21_20261025` (added 2026-10-29; raw demos freshly downloaded from Linslab/ManiLadder_Dataset) use seeds 1000–1049. All reached 40,000 training steps and have all eight 5k–40k evaluations. See [its manifest](single_task/pi05_hf21_20261025_manifest.json).
17
 
18
  See [retraining provenance and missing steps](single_task/pi05_retrain_eval50_20261022_manifest.json). These scores use different seed sets across the two groups and are not a controlled comparison.
19
 
 
24
  - `eval_result.json` — 50-episode evaluation of this checkpoint
25
 
26
  `single_task/results.csv` lists the selected checkpoint per task; `single_task/all_checkpoints_eval.csv` has the
27
+ 50-episode result of every evaluated checkpoint (854 total).
28
 
29
  Evaluation: `exec_steps=16`, `sample_steps=10`. Tasks whose checkpoints all scored 0/50 are still included;
30
  for them the step choice is arbitrary.
 
107
  | HitBox-v1 | 3 | 25000 | 0.06 |
108
  | LiquidMeasuring-v1 | 3 | 35000 | 0.32 |
109
  | MobilePickPlace-v1 | 3 | 15000 | 0.08 |
110
+ | PressFitSoftPeg-v1 | 3 | 20000 | 0.08 |
111
  | PushTargetPose-v1 | 3 | 30000 | 0.02 |
112
  | SeparateTwoRopesEnv-v1 | 3 | 35000 | 0.26 |
113
  | Stack6Blocks-v1 | 3 | 5000 | 0.00 |
 
136
  | RodThroughHoles-v1 | 4 | 5000 | 0.02 |
137
  | SpinPenHora-v1 | 4 | 5000 | 0.00 |
138
  | StampRoll-v1 | 4 | 5000 | 0.00 |
139
+ | SwingRing-v1 | 4 | 15000 | 0.04 |
140
  | TennisRally-v1 | 4 | 15000 | 0.06 |
141
  | WrapRope-v1 | 4 | 5000 | 0.02 |
142
 
single_task/all_checkpoints_eval.csv CHANGED
@@ -601,6 +601,14 @@ MobilePickPlace-v1,level_3,25000,0.02,1,nan
601
  MobilePickPlace-v1,level_3,30000,0.02,1,nan
602
  MobilePickPlace-v1,level_3,35000,0.04,2,169.25979777826228
603
  MobilePickPlace-v1,level_3,40000,0.04,2,nan
 
 
 
 
 
 
 
 
604
  PushTargetPose-v1,level_3,5000,0.0,0,-348.17943846166133
605
  PushTargetPose-v1,level_3,10000,0.0,0,-391.1001693305373
606
  PushTargetPose-v1,level_3,15000,0.0,0,-293.1471811267734
@@ -821,6 +829,14 @@ StampRoll-v1,level_4,25000,0.0,0,0.0
821
  StampRoll-v1,level_4,30000,0.0,0,0.0
822
  StampRoll-v1,level_4,35000,0.0,0,0.0
823
  StampRoll-v1,level_4,40000,0.0,0,0.0
 
 
 
 
 
 
 
 
824
  TennisRally-v1,level_4,5000,0.02,1,357.17158117782327
825
  TennisRally-v1,level_4,10000,0.02,1,582.0495989676378
826
  TennisRally-v1,level_4,15000,0.06,3,631.176888383422
 
601
  MobilePickPlace-v1,level_3,30000,0.02,1,nan
602
  MobilePickPlace-v1,level_3,35000,0.04,2,169.25979777826228
603
  MobilePickPlace-v1,level_3,40000,0.04,2,nan
604
+ PressFitSoftPeg-v1,level_3,5000,0.0,0,693.0195
605
+ PressFitSoftPeg-v1,level_3,10000,0.0,0,1621.6199599999998
606
+ PressFitSoftPeg-v1,level_3,15000,0.0,0,1761.0606
607
+ PressFitSoftPeg-v1,level_3,20000,0.08,4,1640.69166
608
+ PressFitSoftPeg-v1,level_3,25000,0.02,1,1553.72812
609
+ PressFitSoftPeg-v1,level_3,30000,0.08,4,1815.9183600000001
610
+ PressFitSoftPeg-v1,level_3,35000,0.08,4,1902.3642799999998
611
+ PressFitSoftPeg-v1,level_3,40000,0.06,3,1725.1256
612
  PushTargetPose-v1,level_3,5000,0.0,0,-348.17943846166133
613
  PushTargetPose-v1,level_3,10000,0.0,0,-391.1001693305373
614
  PushTargetPose-v1,level_3,15000,0.0,0,-293.1471811267734
 
829
  StampRoll-v1,level_4,30000,0.0,0,0.0
830
  StampRoll-v1,level_4,35000,0.0,0,0.0
831
  StampRoll-v1,level_4,40000,0.0,0,0.0
832
+ SwingRing-v1,level_4,5000,0.02,1,73.43362
833
+ SwingRing-v1,level_4,10000,0.02,1,77.27958
834
+ SwingRing-v1,level_4,15000,0.04,2,78.49958
835
+ SwingRing-v1,level_4,20000,0.0,0,76.71624
836
+ SwingRing-v1,level_4,25000,0.0,0,71.01814
837
+ SwingRing-v1,level_4,30000,0.02,1,75.07314000000001
838
+ SwingRing-v1,level_4,35000,0.0,0,73.35582000000001
839
+ SwingRing-v1,level_4,40000,0.0,0,74.07002
840
  TennisRally-v1,level_4,5000,0.02,1,357.17158117782327
841
  TennisRally-v1,level_4,10000,0.02,1,582.0495989676378
842
  TennisRally-v1,level_4,15000,0.06,3,631.176888383422
single_task/pi05_hf21_20261025_manifest.json CHANGED
@@ -155,6 +155,21 @@
155
  "checkpoint": "single_task/DualArmThreading-v1/lora_step_0020000.npz",
156
  "checkpoint_sha256": "7e815721a302d324c55dea8c9681332c939b65d432b3918e2a00d365828ffedd"
157
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
158
  {
159
  "task": "DualArmMakeLine-v1",
160
  "level": "level_3",
@@ -290,6 +305,21 @@
290
  "checkpoint": "single_task/CoverPattern-v1/lora_step_0030000.npz",
291
  "checkpoint_sha256": "0f555c7b620fa9355b6f56ee378c7531d89e630c9b6383210cdac21cd2dae020"
292
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
293
  {
294
  "task": "StampRoll-v1",
295
  "level": "level_4",
 
155
  "checkpoint": "single_task/DualArmThreading-v1/lora_step_0020000.npz",
156
  "checkpoint_sha256": "7e815721a302d324c55dea8c9681332c939b65d432b3918e2a00d365828ffedd"
157
  },
158
+ {
159
+ "task": "PressFitSoftPeg-v1",
160
+ "level": "level_3",
161
+ "suite": "deformable/single_arm_dex",
162
+ "step": 20000,
163
+ "success_rate": 0.08,
164
+ "successes": 4,
165
+ "episodes": 50,
166
+ "episode_seed_base": 1000,
167
+ "ckpts_evaluated": 8,
168
+ "missing_eval_steps": [],
169
+ "campaign": "pi05_hf21_20261025",
170
+ "checkpoint": "single_task/PressFitSoftPeg-v1/lora_step_0020000.npz",
171
+ "checkpoint_sha256": "3f798221f49b418dee3fbf610282e8bbb29499d02705a072d0e4053cfcf051c4"
172
+ },
173
  {
174
  "task": "DualArmMakeLine-v1",
175
  "level": "level_3",
 
305
  "checkpoint": "single_task/CoverPattern-v1/lora_step_0030000.npz",
306
  "checkpoint_sha256": "0f555c7b620fa9355b6f56ee378c7531d89e630c9b6383210cdac21cd2dae020"
307
  },
308
+ {
309
+ "task": "SwingRing-v1",
310
+ "level": "level_4",
311
+ "suite": "deformable/single_arm_dex",
312
+ "step": 15000,
313
+ "success_rate": 0.04,
314
+ "successes": 2,
315
+ "episodes": 50,
316
+ "episode_seed_base": 1000,
317
+ "ckpts_evaluated": 8,
318
+ "missing_eval_steps": [],
319
+ "campaign": "pi05_hf21_20261025",
320
+ "checkpoint": "single_task/SwingRing-v1/lora_step_0015000.npz",
321
+ "checkpoint_sha256": "4ba44fb21dbbc9435859c93e0796840c75f7855ba3fc996e4b748f6d4360085e"
322
+ },
323
  {
324
  "task": "StampRoll-v1",
325
  "level": "level_4",
single_task/results.csv CHANGED
@@ -75,6 +75,7 @@ HammerNail-v1,level_3,35000,0.12,6,8,single_task/HammerNail-v1/lora_step_0035000
75
  HitBox-v1,level_3,25000,0.06,3,8,single_task/HitBox-v1/lora_step_0025000.npz
76
  LiquidMeasuring-v1,level_3,35000,0.32,16,8,single_task/LiquidMeasuring-v1/lora_step_0035000.npz
77
  MobilePickPlace-v1,level_3,15000,0.08,4,8,single_task/MobilePickPlace-v1/lora_step_0015000.npz
 
78
  PushTargetPose-v1,level_3,30000,0.02,1,8,single_task/PushTargetPose-v1/lora_step_0030000.npz
79
  SeparateTwoRopesEnv-v1,level_3,35000,0.26,13,7,single_task/SeparateTwoRopesEnv-v1/lora_step_0035000.npz
80
  Stack6Blocks-v1,level_3,5000,0.0,0,8,single_task/Stack6Blocks-v1/lora_step_0005000.npz
@@ -103,5 +104,6 @@ RingTossing-v1,level_4,5000,0.0,0,8,single_task/RingTossing-v1/lora_step_0005000
103
  RodThroughHoles-v1,level_4,5000,0.02,1,8,single_task/RodThroughHoles-v1/lora_step_0005000.npz
104
  SpinPenHora-v1,level_4,5000,0.0,0,8,single_task/SpinPenHora-v1/lora_step_0005000.npz
105
  StampRoll-v1,level_4,5000,0.0,0,8,single_task/StampRoll-v1/lora_step_0005000.npz
 
106
  TennisRally-v1,level_4,15000,0.06,3,8,single_task/TennisRally-v1/lora_step_0015000.npz
107
  WrapRope-v1,level_4,5000,0.02,1,8,single_task/WrapRope-v1/lora_step_0005000.npz
 
75
  HitBox-v1,level_3,25000,0.06,3,8,single_task/HitBox-v1/lora_step_0025000.npz
76
  LiquidMeasuring-v1,level_3,35000,0.32,16,8,single_task/LiquidMeasuring-v1/lora_step_0035000.npz
77
  MobilePickPlace-v1,level_3,15000,0.08,4,8,single_task/MobilePickPlace-v1/lora_step_0015000.npz
78
+ PressFitSoftPeg-v1,level_3,20000,0.08,4,8,single_task/PressFitSoftPeg-v1/lora_step_0020000.npz
79
  PushTargetPose-v1,level_3,30000,0.02,1,8,single_task/PushTargetPose-v1/lora_step_0030000.npz
80
  SeparateTwoRopesEnv-v1,level_3,35000,0.26,13,7,single_task/SeparateTwoRopesEnv-v1/lora_step_0035000.npz
81
  Stack6Blocks-v1,level_3,5000,0.0,0,8,single_task/Stack6Blocks-v1/lora_step_0005000.npz
 
104
  RodThroughHoles-v1,level_4,5000,0.02,1,8,single_task/RodThroughHoles-v1/lora_step_0005000.npz
105
  SpinPenHora-v1,level_4,5000,0.0,0,8,single_task/SpinPenHora-v1/lora_step_0005000.npz
106
  StampRoll-v1,level_4,5000,0.0,0,8,single_task/StampRoll-v1/lora_step_0005000.npz
107
+ SwingRing-v1,level_4,15000,0.04,2,8,single_task/SwingRing-v1/lora_step_0015000.npz
108
  TennisRally-v1,level_4,15000,0.06,3,8,single_task/TennisRally-v1/lora_step_0015000.npz
109
  WrapRope-v1,level_4,5000,0.02,1,8,single_task/WrapRope-v1/lora_step_0005000.npz