Linslab commited on
Commit
2d533e2
·
verified ·
1 Parent(s): 64b82d0

pi05 LoRA: 21 new single-task checkpoints (90 total); refresh README + results.csv

Browse files
Files changed (2) hide show
  1. pi05_lora/README.md +49 -55
  2. pi05_lora/results.csv +97 -76
pi05_lora/README.md CHANGED
@@ -1,16 +1,13 @@
1
  # ManiLadder — PI0.5 LoRA checkpoints
2
 
3
- LoRA fine-tunes of `pi05_base` on [ManiLadder](https://huggingface.co/datasets/Linslab/ManiLadder_Dataset).
4
- Only the **best-scoring checkpoint of each run** is published here.
5
-
6
- Two groups, kept in separate directories:
7
 
8
  | directory | what | runs |
9
  |---|---|---|
10
- | `pi05_lora/single_task/<Task>-v1/` | one LoRA per task | 69 |
11
- | `pi05_lora/suite/<level>_<embodiment>/` | one LoRA per (level, embodiment) suite, trained on 4 tasks jointly | 6 |
12
 
13
- `results.csv` in this directory lists every published checkpoint with its score.
14
 
15
  ## Training setup (identical for both groups)
16
 
@@ -23,15 +20,13 @@ Two groups, kept in separate directories:
23
  | optimizer | AdamW, lr 5e-5 cosine (1k warmup, 30k decay to 5e-6), grad-clip 1.0 |
24
  | batch / hardware | 64 on a single A100-80GB, `fsdp_devices=1` |
25
  | steps | 40 000, online eval every 5 000 steps × 20 episodes |
26
- | observations | 3 × 224×224 RGB (`base_0_rgb`, `left_wrist_0_rgb`, `right_wrist_0_rgb`; right wrist is a zero image for single-arm tasks) |
27
- | proprioception | **none** — `observation.state` is written as zeros by design, so these policies are vision-only |
28
  | action | flattened and padded to 32 dims (44 for dual-arm dexterous-hand tasks), horizon 50 |
29
 
30
  ## File format
31
 
32
- Each `lora_step_<step>.npz` is the **trainable slice** of that checkpoint — a flat
33
- npz whose keys are nnx parameter paths joined with `/` (51 tensors, ~1.87 GB;
34
- most of it is the fully-trained SigLIP tower, which is why it is not adapter-sized).
35
 
36
  ```python
37
  import numpy as np
@@ -40,33 +35,24 @@ w["PaliGemma/img/Transformer/encoder_norm/scale"] # -> (1152,)
40
  w["action_out_proj/kernel"] # -> (1024, 32) or (1024, 44)
41
  ```
42
 
43
- These weights are **not self-contained**: merge them onto `pi05_base` to get a
44
- runnable policy. `eval_log.txt` next to each file is the run's own evaluation log
45
- (`step=<n> metrics={...}`, one line per online eval).
46
 
47
  ## Choosing a checkpoint
48
 
49
- The step in each filename is that run's **best** online eval, which is usually not
50
- the last step: of the 68 single-task runs that finished 40k steps, only 7 peaked at
51
- 40k, and the mean drop from peak to step 40 000 is 0.095 success. Every file here is
52
- already the peak one.
53
 
54
  ## Single-task results
55
 
56
- Peak = best of 8 online evals, 20 episodes each. `final` = the same run at step 40 000.
57
-
58
- Mean peak 0.388, median 0.25; 13 tasks ≥ 0.80, 17 tasks at 0.00.
59
 
60
  | level | tasks | mean peak | ≥0.50 | =0.00 |
61
  |---|---|---|---|---|
62
- | L1 | 24 | 0.733 | 21 | 0 |
63
- | L2 | 19 | 0.405 | 7 | 1 |
64
- | L3 | 13 | 0.081 | 0 | 7 |
65
- | L4 | 13 | 0.031 | 0 | 9 |
66
 
67
- Success falls monotonically with the ManiLadder level, and **no level-3 or level-4
68
- task ever reached 0.50** — 20-demo LoRA fine-tuning works at levels 1–2 and breaks
69
- down above that.
70
 
71
  | task | level | embodiment | peak | step | final |
72
  |---|---|---|---|---|---|
@@ -77,14 +63,17 @@ down above that.
77
  | `RopePalmCageHoldDex-v1` | L1 | single_arm_dex | **1.00** | 10000 | 1.00 |
78
  | `DualArmButtonPress-v1` | L1 | dual_arm_gripper | **0.95** | 15000 | 0.70 |
79
  | `DualArmFoldHandKerchief-v1` | L1 | dual_arm_gripper | **0.95** | 5000 | 0.70 |
 
80
  | `TurnOnFaucet-v1` | L1 | single_arm_dex | **0.95** | 40000 | 0.95 |
81
  | `GraspCylinder-v1` | L1 | single_arm_dex | **0.90** | 10000 | 0.85 |
82
  | `RollingPin-v1` | L2 | single_arm_dex | **0.90** | 35000 | 0.75 |
83
  | `DualArmDoubleDoorBox-v1` | L1 | dual_arm_gripper | **0.85** | 35000 | 0.70 |
 
84
  | `PushSoccer-v1` | L1 | single_arm_gripper | **0.85** | 25000 | 0.70 |
85
  | `HangMug-v1` | L2 | single_arm_gripper | **0.80** | 5000 | 0.70 |
86
  | `DualArmBookPickPlace-v1` | L2 | dual_arm_gripper | **0.75** | 25000 | 0.65 |
87
  | `DualArmCoverCloth-v1` | L1 | dual_arm_dex | **0.75** | 15000 | 0.40 |
 
88
  | `OpenDoor-v1` | L1 | single_arm_gripper | **0.75** | 20000 | 0.60 |
89
  | `DualArmHammerHandover-v1` | L1 | dual_arm_gripper | **0.70** | 20000 | 0.35 |
90
  | `DualArmRopeCylinderWarpping-v1` | L1 | dual_arm_gripper | **0.70** | 30000 | 0.60 |
@@ -93,6 +82,7 @@ down above that.
93
  | `StackBlock-v1` | L1 | single_arm_gripper | **0.65** | 25000 | 0.45 |
94
  | `DualArmLiftBall-v1` | L2 | dual_arm_gripper | **0.60** | 40000 | 0.60 |
95
  | `GraspTwoBlocks-v1` | L2 | single_arm_dex | **0.60** | 10000 | 0.40 |
 
96
  | `HammerNailOut-v1` | L2 | single_arm_dex | **0.55** | 35000 | 0.35 |
97
  | `PickPlacePen-v1` | L1 | single_arm_gripper | **0.55** | 20000 | 0.45 |
98
  | `StraightenRope-v1` | L1 | single_arm_gripper | **0.55** | 40000 | 0.55 |
@@ -101,8 +91,14 @@ down above that.
101
  | `DualArmOpenLaptop-v1` | L2 | dual_arm_dex | **0.45** | 25000 | 0.25 |
102
  | `DualArmPushBox-v1` | L1 | dual_arm_gripper | **0.45** | 5000 | 0.15 |
103
  | `ReorientPen-v1` | L2 | single_arm_dex | **0.45** | 30000 | 0.40 |
 
104
  | `DualArmClothAlign-v1` | L2 | dual_arm_gripper | **0.40** | 25000 | 0.35 |
 
 
105
  | `PivotBook-v1` | L2 | single_arm_gripper | **0.40** | 10000 | 0.10 |
 
 
 
106
  | `ClutterGrasp-v1` | L3 | single_arm_dex | **0.25** | 25000 | 0.25 |
107
  | `DualArmFoldCarton-v1` | L3 | dual_arm_gripper | **0.25** | 40000 | 0.25 |
108
  | `DualArmPegInHole-v1` | L2 | dual_arm_gripper | **0.25** | 5000 | 0.05 |
@@ -114,21 +110,30 @@ down above that.
114
  | `DualArmSetupTable-v1` | L3 | dual_arm_dex | **0.15** | 25000 | 0.00 |
115
  | `DualArmStraightenRope-v1` | L2 | dual_arm_gripper | **0.15** | 30000 | 0.05 |
116
  | `DualArmSwingCup-v1` | L1 | dual_arm_dex | **0.15** | 40000 | 0.15 |
 
117
  | `HammerNail-v1` | L3 | single_arm_dex | **0.15** | 10000 | 0.15 |
 
118
  | `DualArmOpenBottle-v1` | L2 | dual_arm_dex | **0.10** | 30000 | 0.05 |
119
  | `DualArmOpenScissor-v1` | L1 | dual_arm_dex | **0.10** | 30000 | 0.00 |
120
  | `KitAssembly-v1` | L2 | single_arm_gripper | **0.10** | 15000 | 0.00 |
121
  | `PingPongBallTransport-v1` | L4 | single_arm_gripper | **0.10** | 25000 | 0.05 |
122
  | `TennisRally-v1` | L4 | single_arm_dex | **0.10** | 15000 | 0.00 |
 
 
123
  | `FlipPan-v1` | L4 | single_arm_gripper | **0.05** | 40000 | 0.05 |
 
124
  | `MobilePickPlace-v1` | L3 | single_arm_dex | **0.05** | 10000 | 0.05 |
125
  | `DualArmBallClassify-v1` | L4 | dual_arm_gripper | **0.00** | 10000 | 0.00 |
126
  | `DualArmBilliardBridge-v1` | L4 | dual_arm_gripper | **0.00** | 5000 | 0.00 |
 
 
127
  | `DualArmKeyboardType-v1` | L4 | dual_arm_dex | **0.00** | 5000 | 0.00 |
128
  | `DualArmMahjongSetup-v1` | L4 | dual_arm_gripper | **0.00** | 10000 | 0.00 |
129
  | `DualArmMakeLine-v1` | L3 | dual_arm_dex | **0.00** | 5000 | 0.00 |
130
  | `DualArmOpenFridge-v1` | L3 | dual_arm_gripper | **0.00** | 5000 | 0.00 |
131
  | `DualArmPlugCharger-v1` | L3 | dual_arm_dex | **0.00** | 5000 | 0.00 |
 
 
132
  | `DualArmShootBasketball-v1` | L4 | dual_arm_dex | **0.00** | 5000 | 0.00 |
133
  | `DualArmThrowBlock-v1` | L2 | dual_arm_dex | **0.00** | 5000 | 0.00 |
134
  | `FitnessBall-v1` | L4 | single_arm_dex | **0.00** | 5000 | 0.00 |
@@ -137,39 +142,28 @@ down above that.
137
  | `PushTargetPose-v1` | L3 | single_arm_gripper | **0.00** | 5000 | 0.00 |
138
  | `RingTossing-v1` | L4 | single_arm_gripper | **0.00** | 5000 | 0.00 |
139
  | `RodThroughHoles-v1` | L4 | single_arm_dex | **0.00** | 5000 | 0.00 |
 
 
140
  | `Stack6Blocks-v1` | L3 | single_arm_gripper | **0.00** | 5000 | 0.00 |
141
  | `TweezersPickPlace-v1` | L3 | single_arm_dex | **0.00** | 5000 | 0.00 |
142
 
143
- ## Suite (multi-task) results
144
 
145
- Each suite trains one LoRA on the 4 tasks of one (level, embodiment) cell.
146
- Scores are the suite mean over its 4 tasks.
147
 
148
  | suite | tasks | step | score |
149
  |---|---|---|---|
150
- | `l1_dual_dex` | DualArmPressButton, DualArmOpenScissor, DualArmSwingCup, DualArmOpenPen | 35000 | 0.265 (50 ep, step 35000) |
151
- | `l2_dual_gripper` | DualArmLiftBall, DualArmPegInHole, DualArmBookPickPlace, DualArmCleanupPotEnv | 40000 | 0.230 (50 ep, step 40000) |
152
- | `l2_single_dex` | ReorientPen, HammerNailOut, GraspTwoBlocks, DexOpenDoor | 40000 | 0.420 (50 ep, step 40000) |
153
- | `l2_single_gripper` | HangMug, KitAssembly, PivotBook, OpenBoxRetrieval | 40000 | 0.285 (50 ep, step 40000) |
154
- | `l3_single_gripper` | MazeBlockPush, GroceryPrepare, Stack6Blocks, PushTargetPose | 40000 | 0.000 (50 ep, step 40000) |
155
- | `l4_single_gripper` | RingTossing, PingPongBallTransport, HitBilliard, FlipPan | 40000 | 0.000 (20 ep, all steps) |
156
-
157
- `l3_single_gripper` and `l4_single_gripper` never scored above 0.00 at any step; the
158
- step-40000 checkpoint is published for completeness, as a negative result. The
159
- `l4_single_gripper` run additionally suffers a normalisation collapse: `HitBilliard-v1`
160
- contributes 10-dim actions against 7-dim for the other three tasks, and after padding
161
- one dimension is non-zero in only 626 of 47 274 frames, so `q01 == q99` in its norm
162
- stats and those frames normalise to ±7e5 — its training loss never falls below 2e6.
163
-
164
- On the 23 tasks that appear in both groups, single-task training beats suite training
165
- by only **+0.035** mean peak success (11 better / 6 worse / 6 tied), measured the same
166
- way (peak over 20-episode online evals). At 20 demos per task, joint training of 4
167
- tasks costs essentially nothing and gains essentially nothing.
168
 
169
  ## Caveats
170
 
171
- - Online eval uses 20 episodes, so the score granularity is 0.05. Repeated 50-episode
172
- offline evals of the same checkpoint differ by up to 0.03. Do not read rank
173
- differences of one or two positions as real.
174
- - Success criteria are the ones in the ManiLadder environments at the time of training
175
- (August 2026); a task whose criterion was later changed is not comparable.
 
1
  # ManiLadder — PI0.5 LoRA checkpoints
2
 
3
+ LoRA fine-tunes of `pi05_base` on [ManiLadder](https://huggingface.co/datasets/Linslab/ManiLadder_Dataset). Only the **best-scoring checkpoint of each run** is published here.
 
 
 
4
 
5
  | directory | what | runs |
6
  |---|---|---|
7
+ | `pi05_lora/single_task/<Task>-v1/` | one LoRA per task | 90 |
8
+ | `pi05_lora/suite/<level>_<embodiment>/` | one LoRA per (level, embodiment) suite, 4 tasks jointly | 6 |
9
 
10
+ `results.csv` lists every published checkpoint with its score.
11
 
12
  ## Training setup (identical for both groups)
13
 
 
20
  | optimizer | AdamW, lr 5e-5 cosine (1k warmup, 30k decay to 5e-6), grad-clip 1.0 |
21
  | batch / hardware | 64 on a single A100-80GB, `fsdp_devices=1` |
22
  | steps | 40 000, online eval every 5 000 steps × 20 episodes |
23
+ | observations | 3 × 224×224 RGB; right wrist is a zero image for single-arm tasks |
24
+ | proprioception | **none** — `observation.state` is written as zeros by design; these policies are vision-only |
25
  | action | flattened and padded to 32 dims (44 for dual-arm dexterous-hand tasks), horizon 50 |
26
 
27
  ## File format
28
 
29
+ Each `lora_step_<step>.npz` is the **trainable slice** of that checkpoint — a flat npz whose keys are nnx parameter paths joined with `/` (51 tensors, ~1.87 GB; most of it is the fully-trained SigLIP tower, which is why it is not adapter-sized).
 
 
30
 
31
  ```python
32
  import numpy as np
 
35
  w["action_out_proj/kernel"] # -> (1024, 32) or (1024, 44)
36
  ```
37
 
38
+ These weights are **not self-contained**: merge them onto `pi05_base` to get a runnable policy. `eval_log.txt` next to each file is that run's own evaluation log.
 
 
39
 
40
  ## Choosing a checkpoint
41
 
42
+ The step in each filename is that run's **best** online eval, which is usually not the last step: of the 90 single-task runs that reached 40k steps, only 9 peaked at 40k, and the mean drop from peak to step 40 000 is 0.083 success. Every file here is already the peak one.
 
 
 
43
 
44
  ## Single-task results
45
 
46
+ Peak = best of 8 online evals, 20 episodes each. Mean peak 0.362, median 0.25; 15 tasks 0.80, 23 tasks at 0.00.
 
 
47
 
48
  | level | tasks | mean peak | ≥0.50 | =0.00 |
49
  |---|---|---|---|---|
50
+ | L1 | 26 | 0.731 | 22 | 0 |
51
+ | L2 | 26 | 0.400 | 9 | 2 |
52
+ | L3 | 20 | 0.130 | 1 | 9 |
53
+ | L4 | 18 | 0.031 | 0 | 12 |
54
 
55
+ Success falls monotonically with the ManiLadder level (0.73 0.40 → 0.13 → 0.03): 20-demo LoRA fine-tuning works at levels 1–2 and largely breaks down above that.
 
 
56
 
57
  | task | level | embodiment | peak | step | final |
58
  |---|---|---|---|---|---|
 
63
  | `RopePalmCageHoldDex-v1` | L1 | single_arm_dex | **1.00** | 10000 | 1.00 |
64
  | `DualArmButtonPress-v1` | L1 | dual_arm_gripper | **0.95** | 15000 | 0.70 |
65
  | `DualArmFoldHandKerchief-v1` | L1 | dual_arm_gripper | **0.95** | 5000 | 0.70 |
66
+ | `DualArmPouringWater-v1` | L1 | dual_arm_gripper | **0.95** | 30000 | 0.85 |
67
  | `TurnOnFaucet-v1` | L1 | single_arm_dex | **0.95** | 40000 | 0.95 |
68
  | `GraspCylinder-v1` | L1 | single_arm_dex | **0.90** | 10000 | 0.85 |
69
  | `RollingPin-v1` | L2 | single_arm_dex | **0.90** | 35000 | 0.75 |
70
  | `DualArmDoubleDoorBox-v1` | L1 | dual_arm_gripper | **0.85** | 35000 | 0.70 |
71
+ | `FoldShirt-v1` | L2 | single_arm_gripper | **0.85** | 15000 | 0.65 |
72
  | `PushSoccer-v1` | L1 | single_arm_gripper | **0.85** | 25000 | 0.70 |
73
  | `HangMug-v1` | L2 | single_arm_gripper | **0.80** | 5000 | 0.70 |
74
  | `DualArmBookPickPlace-v1` | L2 | dual_arm_gripper | **0.75** | 25000 | 0.65 |
75
  | `DualArmCoverCloth-v1` | L1 | dual_arm_dex | **0.75** | 15000 | 0.40 |
76
+ | `DualArmPlaceRopeOverRod-v1` | L2 | dual_arm_dex | **0.75** | 30000 | 0.70 |
77
  | `OpenDoor-v1` | L1 | single_arm_gripper | **0.75** | 20000 | 0.60 |
78
  | `DualArmHammerHandover-v1` | L1 | dual_arm_gripper | **0.70** | 20000 | 0.35 |
79
  | `DualArmRopeCylinderWarpping-v1` | L1 | dual_arm_gripper | **0.70** | 30000 | 0.60 |
 
82
  | `StackBlock-v1` | L1 | single_arm_gripper | **0.65** | 25000 | 0.45 |
83
  | `DualArmLiftBall-v1` | L2 | dual_arm_gripper | **0.60** | 40000 | 0.60 |
84
  | `GraspTwoBlocks-v1` | L2 | single_arm_dex | **0.60** | 10000 | 0.40 |
85
+ | `MazeBlockPush-v1` [^1] | L3 | single_arm_gripper | **0.60** | 40000 | 0.60 |
86
  | `HammerNailOut-v1` | L2 | single_arm_dex | **0.55** | 35000 | 0.35 |
87
  | `PickPlacePen-v1` | L1 | single_arm_gripper | **0.55** | 20000 | 0.45 |
88
  | `StraightenRope-v1` | L1 | single_arm_gripper | **0.55** | 40000 | 0.55 |
 
91
  | `DualArmOpenLaptop-v1` | L2 | dual_arm_dex | **0.45** | 25000 | 0.25 |
92
  | `DualArmPushBox-v1` | L1 | dual_arm_gripper | **0.45** | 5000 | 0.15 |
93
  | `ReorientPen-v1` | L2 | single_arm_dex | **0.45** | 30000 | 0.40 |
94
+ | `UnfoldShirt-v1` | L1 | single_arm_gripper | **0.45** | 15000 | 0.30 |
95
  | `DualArmClothAlign-v1` | L2 | dual_arm_gripper | **0.40** | 25000 | 0.35 |
96
+ | `DualArmStuffCloth-v1` | L2 | dual_arm_dex | **0.40** | 15000 | 0.15 |
97
+ | `LiquidMeasuring-v1` | L3 | single_arm_gripper | **0.40** | 25000 | 0.20 |
98
  | `PivotBook-v1` | L2 | single_arm_gripper | **0.40** | 10000 | 0.10 |
99
+ | `ClothWipe-v1` | L2 | single_arm_dex | **0.35** | 10000 | 0.20 |
100
+ | `DualArmWeaveCloth-v1` | L3 | dual_arm_gripper | **0.35** | 20000 | 0.25 |
101
+ | `StraightenTwoRopesSequential-v1` | L2 | single_arm_gripper | **0.30** | 10000 | 0.05 |
102
  | `ClutterGrasp-v1` | L3 | single_arm_dex | **0.25** | 25000 | 0.25 |
103
  | `DualArmFoldCarton-v1` | L3 | dual_arm_gripper | **0.25** | 40000 | 0.25 |
104
  | `DualArmPegInHole-v1` | L2 | dual_arm_gripper | **0.25** | 5000 | 0.05 |
 
110
  | `DualArmSetupTable-v1` | L3 | dual_arm_dex | **0.15** | 25000 | 0.00 |
111
  | `DualArmStraightenRope-v1` | L2 | dual_arm_gripper | **0.15** | 30000 | 0.05 |
112
  | `DualArmSwingCup-v1` | L1 | dual_arm_dex | **0.15** | 40000 | 0.15 |
113
+ | `DualArmThreadRopeThroughTube-v1` | L3 | dual_arm_gripper | **0.15** | 5000 | 0.10 |
114
  | `HammerNail-v1` | L3 | single_arm_dex | **0.15** | 10000 | 0.15 |
115
+ | `BuildTent-v1` | L4 | single_arm_gripper | **0.10** | 5000 | 0.00 |
116
  | `DualArmOpenBottle-v1` | L2 | dual_arm_dex | **0.10** | 30000 | 0.05 |
117
  | `DualArmOpenScissor-v1` | L1 | dual_arm_dex | **0.10** | 30000 | 0.00 |
118
  | `KitAssembly-v1` | L2 | single_arm_gripper | **0.10** | 15000 | 0.00 |
119
  | `PingPongBallTransport-v1` | L4 | single_arm_gripper | **0.10** | 25000 | 0.05 |
120
  | `TennisRally-v1` | L4 | single_arm_dex | **0.10** | 15000 | 0.00 |
121
+ | `DualArmJugglingBall-v1` | L2 | dual_arm_dex | **0.05** | 30000 | 0.05 |
122
+ | `DualArmJumpRope-v1` | L4 | dual_arm_dex | **0.05** | 40000 | 0.05 |
123
  | `FlipPan-v1` | L4 | single_arm_gripper | **0.05** | 40000 | 0.05 |
124
+ | `HitBox-v1` | L3 | single_arm_dex | **0.05** | 20000 | 0.05 |
125
  | `MobilePickPlace-v1` | L3 | single_arm_dex | **0.05** | 10000 | 0.05 |
126
  | `DualArmBallClassify-v1` | L4 | dual_arm_gripper | **0.00** | 10000 | 0.00 |
127
  | `DualArmBilliardBridge-v1` | L4 | dual_arm_gripper | **0.00** | 5000 | 0.00 |
128
+ | `DualArmClothHanger-v1` | L4 | dual_arm_gripper | **0.00** | 5000 | 0.00 |
129
+ | `DualArmFlipTable-v1` | L4 | dual_arm_dex | **0.00** | 5000 | 0.00 |
130
  | `DualArmKeyboardType-v1` | L4 | dual_arm_dex | **0.00** | 5000 | 0.00 |
131
  | `DualArmMahjongSetup-v1` | L4 | dual_arm_gripper | **0.00** | 10000 | 0.00 |
132
  | `DualArmMakeLine-v1` | L3 | dual_arm_dex | **0.00** | 5000 | 0.00 |
133
  | `DualArmOpenFridge-v1` | L3 | dual_arm_gripper | **0.00** | 5000 | 0.00 |
134
  | `DualArmPlugCharger-v1` | L3 | dual_arm_dex | **0.00** | 5000 | 0.00 |
135
+ | `DualArmPushObject-v1` | L2 | dual_arm_gripper | **0.00** | 5000 | 0.00 |
136
+ | `DualArmShapeDoll-v1` | L3 | dual_arm_gripper | **0.00** | 5000 | 0.00 |
137
  | `DualArmShootBasketball-v1` | L4 | dual_arm_dex | **0.00** | 5000 | 0.00 |
138
  | `DualArmThrowBlock-v1` | L2 | dual_arm_dex | **0.00** | 5000 | 0.00 |
139
  | `FitnessBall-v1` | L4 | single_arm_dex | **0.00** | 5000 | 0.00 |
 
142
  | `PushTargetPose-v1` | L3 | single_arm_gripper | **0.00** | 5000 | 0.00 |
143
  | `RingTossing-v1` | L4 | single_arm_gripper | **0.00** | 5000 | 0.00 |
144
  | `RodThroughHoles-v1` | L4 | single_arm_dex | **0.00** | 5000 | 0.00 |
145
+ | `SeparateTwoRopesEnv-v1` | L3 | single_arm_gripper | **0.00** | 5000 | 0.00 |
146
+ | `SpinPenHora-v1` | L4 | single_arm_dex | **0.00** | 5000 | 0.00 |
147
  | `Stack6Blocks-v1` | L3 | single_arm_gripper | **0.00** | 5000 | 0.00 |
148
  | `TweezersPickPlace-v1` | L3 | single_arm_dex | **0.00** | 5000 | 0.00 |
149
 
150
+ [^1]: `MazeBlockPush-v1` — this score was obtained with a **seeded** maze pool (collection and evaluation drew the same 20 mazes from `MAZE_LAYOUT_SEED=2025`), and with a fixed ball start. The environment has since been changed so the maze pool is drawn from OS entropy on every construction, making train and eval mazes independent; the checkpoint published here is not comparable to results from the current environment.
151
 
152
+ ## Suite (multi-task) results
 
153
 
154
  | suite | tasks | step | score |
155
  |---|---|---|---|
156
+ | `l1_dual_dex` | DualArmPressButton, DualArmOpenScissor, DualArmSwingCup, DualArmOpenPen | 35000 | 0.265 (50 ep) |
157
+ | `l2_dual_gripper` | DualArmLiftBall, DualArmPegInHole, DualArmBookPickPlace, DualArmCleanupPotEnv | 40000 | 0.230 (50 ep) |
158
+ | `l2_single_dex` | ReorientPen, HammerNailOut, GraspTwoBlocks, DexOpenDoor | 40000 | 0.420 (50 ep) |
159
+ | `l2_single_gripper` | HangMug, KitAssembly, PivotBook, OpenBoxRetrieval | 40000 | 0.285 (50 ep) |
160
+ | `l3_single_gripper` | MazeBlockPush, GroceryPrepare, Stack6Blocks, PushTargetPose | 40000 | 0.000 (50 ep) |
161
+ | `l4_single_gripper` | RingTossing, PingPongBallTransport, HitBilliard, FlipPan | 40000 | 0.000 (20 ep) |
162
+
163
+ On the 24 tasks present in both groups, single-task training beats suite training by **+0.053** mean peak success (12 better / 6 worse / 6 tied), measured the same way (peak over 20-episode online evals). At 20 demos per task, joint training of 4 tasks costs essentially nothing and gains essentially nothing.
 
 
 
 
 
 
 
 
 
 
164
 
165
  ## Caveats
166
 
167
+ - Online eval uses 20 episodes, so score granularity is 0.05. Repeated 50-episode offline evals of the same checkpoint differ by up to 0.03. Do not read rank differences of one or two positions as real.
168
+ - Success criteria are the ones in the ManiLadder environments at training time (August–September 2026). A task whose criterion or randomisation was later changed is not comparable — see the footnotes above.
169
+ - `l4_single_gripper` suffers a normalisation collapse: `HitBilliard-v1` contributes 10-dim actions against 7-dim for its three suite-mates, and after padding one dimension is non-zero in only 626 of 47 274 frames, so `q01 == q99` in its norm stats and those frames normalise to ±7e5. Its training loss never falls below 2e6.
 
 
pi05_lora/results.csv CHANGED
@@ -1,76 +1,97 @@
1
- group,name,level,embodiment,category,checkpoint_step,peak_success,final_success_40k,file
2
- single_task,DrapeCloth-v1,level_1,single_arm_dex,deformable,5000,1.00,0.95,pi05_lora/single_task/DrapeCloth-v1/lora_step_0005000.npz
3
- single_task,DualArmRopeStretch-v1,level_1,dual_arm_dex,deformable,10000,1.00,1.00,pi05_lora/single_task/DualArmRopeStretch-v1/lora_step_0010000.npz
4
- single_task,LiftBucket-v1,level_1,single_arm_dex,rigid_body,10000,1.00,0.85,pi05_lora/single_task/LiftBucket-v1/lora_step_0010000.npz
5
- single_task,ReorientCube-v1,level_1,single_arm_dex,rigid_body,30000,1.00,0.90,pi05_lora/single_task/ReorientCube-v1/lora_step_0030000.npz
6
- single_task,RopePalmCageHoldDex-v1,level_1,single_arm_dex,deformable,10000,1.00,1.00,pi05_lora/single_task/RopePalmCageHoldDex-v1/lora_step_0010000.npz
7
- single_task,DualArmButtonPress-v1,level_1,dual_arm_gripper,rigid_body,15000,0.95,0.70,pi05_lora/single_task/DualArmButtonPress-v1/lora_step_0015000.npz
8
- single_task,DualArmFoldHandKerchief-v1,level_1,dual_arm_gripper,deformable,5000,0.95,0.70,pi05_lora/single_task/DualArmFoldHandKerchief-v1/lora_step_0005000.npz
9
- single_task,TurnOnFaucet-v1,level_1,single_arm_dex,rigid_body,40000,0.95,0.95,pi05_lora/single_task/TurnOnFaucet-v1/lora_step_0040000.npz
10
- single_task,GraspCylinder-v1,level_1,single_arm_dex,rigid_body,10000,0.90,0.85,pi05_lora/single_task/GraspCylinder-v1/lora_step_0010000.npz
11
- single_task,RollingPin-v1,level_2,single_arm_dex,deformable,35000,0.90,0.75,pi05_lora/single_task/RollingPin-v1/lora_step_0035000.npz
12
- single_task,DualArmDoubleDoorBox-v1,level_1,dual_arm_gripper,rigid_body,35000,0.85,0.70,pi05_lora/single_task/DualArmDoubleDoorBox-v1/lora_step_0035000.npz
13
- single_task,PushSoccer-v1,level_1,single_arm_gripper,rigid_body,25000,0.85,0.70,pi05_lora/single_task/PushSoccer-v1/lora_step_0025000.npz
14
- single_task,HangMug-v1,level_2,single_arm_gripper,rigid_body,5000,0.80,0.70,pi05_lora/single_task/HangMug-v1/lora_step_0005000.npz
15
- single_task,DualArmBookPickPlace-v1,level_2,dual_arm_gripper,rigid_body,25000,0.75,0.65,pi05_lora/single_task/DualArmBookPickPlace-v1/lora_step_0025000.npz
16
- single_task,DualArmCoverCloth-v1,level_1,dual_arm_dex,deformable,15000,0.75,0.40,pi05_lora/single_task/DualArmCoverCloth-v1/lora_step_0015000.npz
17
- single_task,OpenDoor-v1,level_1,single_arm_gripper,rigid_body,20000,0.75,0.60,pi05_lora/single_task/OpenDoor-v1/lora_step_0020000.npz
18
- single_task,DualArmHammerHandover-v1,level_1,dual_arm_gripper,rigid_body,20000,0.70,0.35,pi05_lora/single_task/DualArmHammerHandover-v1/lora_step_0020000.npz
19
- single_task,DualArmRopeCylinderWarpping-v1,level_1,dual_arm_gripper,deformable,30000,0.70,0.60,pi05_lora/single_task/DualArmRopeCylinderWarpping-v1/lora_step_0030000.npz
20
- single_task,DualArmOpenPen-v1,level_1,dual_arm_dex,rigid_body,30000,0.65,0.35,pi05_lora/single_task/DualArmOpenPen-v1/lora_step_0030000.npz
21
- single_task,DualArmPressButton-v1,level_1,dual_arm_dex,rigid_body,20000,0.65,0.45,pi05_lora/single_task/DualArmPressButton-v1/lora_step_0020000.npz
22
- single_task,StackBlock-v1,level_1,single_arm_gripper,rigid_body,25000,0.65,0.45,pi05_lora/single_task/StackBlock-v1/lora_step_0025000.npz
23
- single_task,DualArmLiftBall-v1,level_2,dual_arm_gripper,rigid_body,40000,0.60,0.60,pi05_lora/single_task/DualArmLiftBall-v1/lora_step_0040000.npz
24
- single_task,GraspTwoBlocks-v1,level_2,single_arm_dex,rigid_body,10000,0.60,0.40,pi05_lora/single_task/GraspTwoBlocks-v1/lora_step_0010000.npz
25
- single_task,HammerNailOut-v1,level_2,single_arm_dex,rigid_body,35000,0.55,0.35,pi05_lora/single_task/HammerNailOut-v1/lora_step_0035000.npz
26
- single_task,PickPlacePen-v1,level_1,single_arm_gripper,rigid_body,20000,0.55,0.45,pi05_lora/single_task/PickPlacePen-v1/lora_step_0020000.npz
27
- single_task,StraightenRope-v1,level_1,single_arm_gripper,deformable,40000,0.55,0.55,pi05_lora/single_task/StraightenRope-v1/lora_step_0040000.npz
28
- single_task,DualArmCleanupPotEnv-v1,level_2,dual_arm_gripper,rigid_body,35000,0.50,0.50,pi05_lora/single_task/DualArmCleanupPotEnv-v1/lora_step_0035000.npz
29
- single_task,ShapeTToI-v1,level_1,single_arm_gripper,deformable,15000,0.50,0.35,pi05_lora/single_task/ShapeTToI-v1/lora_step_0015000.npz
30
- single_task,DualArmOpenLaptop-v1,level_2,dual_arm_dex,rigid_body,25000,0.45,0.25,pi05_lora/single_task/DualArmOpenLaptop-v1/lora_step_0025000.npz
31
- single_task,DualArmPushBox-v1,level_1,dual_arm_gripper,rigid_body,5000,0.45,0.15,pi05_lora/single_task/DualArmPushBox-v1/lora_step_0005000.npz
32
- single_task,ReorientPen-v1,level_2,single_arm_dex,rigid_body,30000,0.45,0.40,pi05_lora/single_task/ReorientPen-v1/lora_step_0030000.npz
33
- single_task,DualArmClothAlign-v1,level_2,dual_arm_gripper,deformable,25000,0.40,0.35,pi05_lora/single_task/DualArmClothAlign-v1/lora_step_0025000.npz
34
- single_task,PivotBook-v1,level_2,single_arm_gripper,rigid_body,10000,0.40,0.10,pi05_lora/single_task/PivotBook-v1/lora_step_0010000.npz
35
- single_task,ClutterGrasp-v1,level_3,single_arm_dex,rigid_body,25000,0.25,0.25,pi05_lora/single_task/ClutterGrasp-v1/lora_step_0025000.npz
36
- single_task,DualArmFoldCarton-v1,level_3,dual_arm_gripper,rigid_body,40000,0.25,0.25,pi05_lora/single_task/DualArmFoldCarton-v1/lora_step_0040000.npz
37
- single_task,DualArmPegInHole-v1,level_2,dual_arm_gripper,rigid_body,5000,0.25,0.05,pi05_lora/single_task/DualArmPegInHole-v1/lora_step_0005000.npz
38
- single_task,OpenBoxRetrieval-v1,level_2,single_arm_gripper,rigid_body,20000,0.25,0.15,pi05_lora/single_task/OpenBoxRetrieval-v1/lora_step_0020000.npz
39
- single_task,ScoopWater-v1,level_2,single_arm_gripper,deformable,20000,0.25,0.15,pi05_lora/single_task/ScoopWater-v1/lora_step_0020000.npz
40
- single_task,DexOpenDoor-v1,level_2,single_arm_dex,rigid_body,40000,0.20,0.20,pi05_lora/single_task/DexOpenDoor-v1/lora_step_0040000.npz
41
- single_task,DualArmPingpongBalance-v1,level_3,dual_arm_dex,rigid_body,20000,0.20,0.20,pi05_lora/single_task/DualArmPingpongBalance-v1/lora_step_0020000.npz
42
- single_task,DualArmSendPingpong-v1,level_4,dual_arm_dex,rigid_body,35000,0.15,0.00,pi05_lora/single_task/DualArmSendPingpong-v1/lora_step_0035000.npz
43
- single_task,DualArmSetupTable-v1,level_3,dual_arm_dex,rigid_body,25000,0.15,0.00,pi05_lora/single_task/DualArmSetupTable-v1/lora_step_0025000.npz
44
- single_task,DualArmStraightenRope-v1,level_2,dual_arm_gripper,deformable,30000,0.15,0.05,pi05_lora/single_task/DualArmStraightenRope-v1/lora_step_0030000.npz
45
- single_task,DualArmSwingCup-v1,level_1,dual_arm_dex,rigid_body,40000,0.15,0.15,pi05_lora/single_task/DualArmSwingCup-v1/lora_step_0040000.npz
46
- single_task,HammerNail-v1,level_3,single_arm_dex,rigid_body,10000,0.15,0.15,pi05_lora/single_task/HammerNail-v1/lora_step_0010000.npz
47
- single_task,DualArmOpenBottle-v1,level_2,dual_arm_dex,rigid_body,30000,0.10,0.05,pi05_lora/single_task/DualArmOpenBottle-v1/lora_step_0030000.npz
48
- single_task,DualArmOpenScissor-v1,level_1,dual_arm_dex,rigid_body,30000,0.10,0.00,pi05_lora/single_task/DualArmOpenScissor-v1/lora_step_0030000.npz
49
- single_task,KitAssembly-v1,level_2,single_arm_gripper,rigid_body,15000,0.10,0.00,pi05_lora/single_task/KitAssembly-v1/lora_step_0015000.npz
50
- single_task,PingPongBallTransport-v1,level_4,single_arm_gripper,rigid_body,25000,0.10,0.05,pi05_lora/single_task/PingPongBallTransport-v1/lora_step_0025000.npz
51
- single_task,TennisRally-v1,level_4,single_arm_dex,rigid_body,15000,0.10,0.00,pi05_lora/single_task/TennisRally-v1/lora_step_0015000.npz
52
- single_task,FlipPan-v1,level_4,single_arm_gripper,rigid_body,40000,0.05,0.05,pi05_lora/single_task/FlipPan-v1/lora_step_0040000.npz
53
- single_task,MobilePickPlace-v1,level_3,single_arm_dex,rigid_body,10000,0.05,0.05,pi05_lora/single_task/MobilePickPlace-v1/lora_step_0010000.npz
54
- single_task,DualArmBallClassify-v1,level_4,dual_arm_gripper,rigid_body,10000,0.00,0.00,pi05_lora/single_task/DualArmBallClassify-v1/lora_step_0010000.npz
55
- single_task,DualArmBilliardBridge-v1,level_4,dual_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmBilliardBridge-v1/lora_step_0005000.npz
56
- single_task,DualArmKeyboardType-v1,level_4,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmKeyboardType-v1/lora_step_0005000.npz
57
- single_task,DualArmMahjongSetup-v1,level_4,dual_arm_gripper,rigid_body,10000,0.00,0.00,pi05_lora/single_task/DualArmMahjongSetup-v1/lora_step_0010000.npz
58
- single_task,DualArmMakeLine-v1,level_3,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmMakeLine-v1/lora_step_0005000.npz
59
- single_task,DualArmOpenFridge-v1,level_3,dual_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmOpenFridge-v1/lora_step_0005000.npz
60
- single_task,DualArmPlugCharger-v1,level_3,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmPlugCharger-v1/lora_step_0005000.npz
61
- single_task,DualArmShootBasketball-v1,level_4,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmShootBasketball-v1/lora_step_0005000.npz
62
- single_task,DualArmThrowBlock-v1,level_2,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmThrowBlock-v1/lora_step_0005000.npz
63
- single_task,FitnessBall-v1,level_4,single_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/FitnessBall-v1/lora_step_0005000.npz
64
- single_task,GroceryPrepare-v1,level_3,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/GroceryPrepare-v1/lora_step_0005000.npz
65
- single_task,HitBilliard-v1,level_4,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/HitBilliard-v1/lora_step_0005000.npz
66
- single_task,PushTargetPose-v1,level_3,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/PushTargetPose-v1/lora_step_0005000.npz
67
- single_task,RingTossing-v1,level_4,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/RingTossing-v1/lora_step_0005000.npz
68
- single_task,RodThroughHoles-v1,level_4,single_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/RodThroughHoles-v1/lora_step_0005000.npz
69
- single_task,Stack6Blocks-v1,level_3,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/Stack6Blocks-v1/lora_step_0005000.npz
70
- single_task,TweezersPickPlace-v1,level_3,single_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/TweezersPickPlace-v1/lora_step_0005000.npz
71
- suite,l1_dual_dex,,,,35000,,,pi05_lora/suite/l1_dual_dex/lora_step_0035000.npz
72
- suite,l2_dual_gripper,,,,40000,,,pi05_lora/suite/l2_dual_gripper/lora_step_0040000.npz
73
- suite,l2_single_dex,,,,40000,,,pi05_lora/suite/l2_single_dex/lora_step_0040000.npz
74
- suite,l2_single_gripper,,,,40000,,,pi05_lora/suite/l2_single_gripper/lora_step_0040000.npz
75
- suite,l3_single_gripper,,,,40000,,,pi05_lora/suite/l3_single_gripper/lora_step_0040000.npz
76
- suite,l4_single_gripper,,,,40000,,,pi05_lora/suite/l4_single_gripper/lora_step_0040000.npz
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ group,name,level,embodiment,category,checkpoint_step,peak_success,final_success_40k,file,note
2
+ single_task,DrapeCloth-v1,level_1,single_arm_dex,deformable,5000,1.00,0.95,pi05_lora/single_task/DrapeCloth-v1/lora_step_0005000.npz,
3
+ single_task,DualArmRopeStretch-v1,level_1,dual_arm_dex,deformable,10000,1.00,1.00,pi05_lora/single_task/DualArmRopeStretch-v1/lora_step_0010000.npz,
4
+ single_task,LiftBucket-v1,level_1,single_arm_dex,rigid_body,10000,1.00,0.85,pi05_lora/single_task/LiftBucket-v1/lora_step_0010000.npz,
5
+ single_task,ReorientCube-v1,level_1,single_arm_dex,rigid_body,30000,1.00,0.90,pi05_lora/single_task/ReorientCube-v1/lora_step_0030000.npz,
6
+ single_task,RopePalmCageHoldDex-v1,level_1,single_arm_dex,deformable,10000,1.00,1.00,pi05_lora/single_task/RopePalmCageHoldDex-v1/lora_step_0010000.npz,
7
+ single_task,DualArmButtonPress-v1,level_1,dual_arm_gripper,rigid_body,15000,0.95,0.70,pi05_lora/single_task/DualArmButtonPress-v1/lora_step_0015000.npz,
8
+ single_task,DualArmFoldHandKerchief-v1,level_1,dual_arm_gripper,deformable,5000,0.95,0.70,pi05_lora/single_task/DualArmFoldHandKerchief-v1/lora_step_0005000.npz,
9
+ single_task,DualArmPouringWater-v1,level_1,dual_arm_gripper,deformable,30000,0.95,0.85,pi05_lora/single_task/DualArmPouringWater-v1/lora_step_0030000.npz,
10
+ single_task,TurnOnFaucet-v1,level_1,single_arm_dex,rigid_body,40000,0.95,0.95,pi05_lora/single_task/TurnOnFaucet-v1/lora_step_0040000.npz,
11
+ single_task,GraspCylinder-v1,level_1,single_arm_dex,rigid_body,10000,0.90,0.85,pi05_lora/single_task/GraspCylinder-v1/lora_step_0010000.npz,
12
+ single_task,RollingPin-v1,level_2,single_arm_dex,deformable,35000,0.90,0.75,pi05_lora/single_task/RollingPin-v1/lora_step_0035000.npz,
13
+ single_task,DualArmDoubleDoorBox-v1,level_1,dual_arm_gripper,rigid_body,35000,0.85,0.70,pi05_lora/single_task/DualArmDoubleDoorBox-v1/lora_step_0035000.npz,
14
+ single_task,FoldShirt-v1,level_2,single_arm_gripper,deformable,15000,0.85,0.65,pi05_lora/single_task/FoldShirt-v1/lora_step_0015000.npz,
15
+ single_task,PushSoccer-v1,level_1,single_arm_gripper,rigid_body,25000,0.85,0.70,pi05_lora/single_task/PushSoccer-v1/lora_step_0025000.npz,
16
+ single_task,HangMug-v1,level_2,single_arm_gripper,rigid_body,5000,0.80,0.70,pi05_lora/single_task/HangMug-v1/lora_step_0005000.npz,
17
+ single_task,DualArmBookPickPlace-v1,level_2,dual_arm_gripper,rigid_body,25000,0.75,0.65,pi05_lora/single_task/DualArmBookPickPlace-v1/lora_step_0025000.npz,
18
+ single_task,DualArmCoverCloth-v1,level_1,dual_arm_dex,deformable,15000,0.75,0.40,pi05_lora/single_task/DualArmCoverCloth-v1/lora_step_0015000.npz,
19
+ single_task,DualArmPlaceRopeOverRod-v1,level_2,dual_arm_dex,deformable,30000,0.75,0.70,pi05_lora/single_task/DualArmPlaceRopeOverRod-v1/lora_step_0030000.npz,
20
+ single_task,OpenDoor-v1,level_1,single_arm_gripper,rigid_body,20000,0.75,0.60,pi05_lora/single_task/OpenDoor-v1/lora_step_0020000.npz,
21
+ single_task,DualArmHammerHandover-v1,level_1,dual_arm_gripper,rigid_body,20000,0.70,0.35,pi05_lora/single_task/DualArmHammerHandover-v1/lora_step_0020000.npz,
22
+ single_task,DualArmRopeCylinderWarpping-v1,level_1,dual_arm_gripper,deformable,30000,0.70,0.60,pi05_lora/single_task/DualArmRopeCylinderWarpping-v1/lora_step_0030000.npz,
23
+ single_task,DualArmOpenPen-v1,level_1,dual_arm_dex,rigid_body,30000,0.65,0.35,pi05_lora/single_task/DualArmOpenPen-v1/lora_step_0030000.npz,
24
+ single_task,DualArmPressButton-v1,level_1,dual_arm_dex,rigid_body,20000,0.65,0.45,pi05_lora/single_task/DualArmPressButton-v1/lora_step_0020000.npz,
25
+ single_task,StackBlock-v1,level_1,single_arm_gripper,rigid_body,25000,0.65,0.45,pi05_lora/single_task/StackBlock-v1/lora_step_0025000.npz,
26
+ single_task,DualArmLiftBall-v1,level_2,dual_arm_gripper,rigid_body,40000,0.60,0.60,pi05_lora/single_task/DualArmLiftBall-v1/lora_step_0040000.npz,
27
+ single_task,GraspTwoBlocks-v1,level_2,single_arm_dex,rigid_body,10000,0.60,0.40,pi05_lora/single_task/GraspTwoBlocks-v1/lora_step_0010000.npz,
28
+ single_task,MazeBlockPush-v1,level_3,single_arm_gripper,rigid_body,40000,0.60,0.60,pi05_lora/single_task/MazeBlockPush-v1/lora_step_0040000.npz,"this score was obtained with a **seeded** maze pool (collection and evaluation drew the same 20 mazes from `MAZE_LAYOUT_SEED=2025`), and with a fixed ball start. The environment has since been changed so the maze pool is drawn from OS entropy on every construction, making train and eval mazes independent; the checkpoint published here is not comparable to results from the current environment."
29
+ single_task,HammerNailOut-v1,level_2,single_arm_dex,rigid_body,35000,0.55,0.35,pi05_lora/single_task/HammerNailOut-v1/lora_step_0035000.npz,
30
+ single_task,PickPlacePen-v1,level_1,single_arm_gripper,rigid_body,20000,0.55,0.45,pi05_lora/single_task/PickPlacePen-v1/lora_step_0020000.npz,
31
+ single_task,StraightenRope-v1,level_1,single_arm_gripper,deformable,40000,0.55,0.55,pi05_lora/single_task/StraightenRope-v1/lora_step_0040000.npz,
32
+ single_task,DualArmCleanupPotEnv-v1,level_2,dual_arm_gripper,rigid_body,35000,0.50,0.50,pi05_lora/single_task/DualArmCleanupPotEnv-v1/lora_step_0035000.npz,
33
+ single_task,ShapeTToI-v1,level_1,single_arm_gripper,deformable,15000,0.50,0.35,pi05_lora/single_task/ShapeTToI-v1/lora_step_0015000.npz,
34
+ single_task,DualArmOpenLaptop-v1,level_2,dual_arm_dex,rigid_body,25000,0.45,0.25,pi05_lora/single_task/DualArmOpenLaptop-v1/lora_step_0025000.npz,
35
+ single_task,DualArmPushBox-v1,level_1,dual_arm_gripper,rigid_body,5000,0.45,0.15,pi05_lora/single_task/DualArmPushBox-v1/lora_step_0005000.npz,
36
+ single_task,ReorientPen-v1,level_2,single_arm_dex,rigid_body,30000,0.45,0.40,pi05_lora/single_task/ReorientPen-v1/lora_step_0030000.npz,
37
+ single_task,UnfoldShirt-v1,level_1,single_arm_gripper,deformable,15000,0.45,0.30,pi05_lora/single_task/UnfoldShirt-v1/lora_step_0015000.npz,
38
+ single_task,DualArmClothAlign-v1,level_2,dual_arm_gripper,deformable,25000,0.40,0.35,pi05_lora/single_task/DualArmClothAlign-v1/lora_step_0025000.npz,
39
+ single_task,DualArmStuffCloth-v1,level_2,dual_arm_dex,deformable,15000,0.40,0.15,pi05_lora/single_task/DualArmStuffCloth-v1/lora_step_0015000.npz,
40
+ single_task,LiquidMeasuring-v1,level_3,single_arm_gripper,deformable,25000,0.40,0.20,pi05_lora/single_task/LiquidMeasuring-v1/lora_step_0025000.npz,
41
+ single_task,PivotBook-v1,level_2,single_arm_gripper,rigid_body,10000,0.40,0.10,pi05_lora/single_task/PivotBook-v1/lora_step_0010000.npz,
42
+ single_task,ClothWipe-v1,level_2,single_arm_dex,deformable,10000,0.35,0.20,pi05_lora/single_task/ClothWipe-v1/lora_step_0010000.npz,
43
+ single_task,DualArmWeaveCloth-v1,level_3,dual_arm_gripper,deformable,20000,0.35,0.25,pi05_lora/single_task/DualArmWeaveCloth-v1/lora_step_0020000.npz,
44
+ single_task,StraightenTwoRopesSequential-v1,level_2,single_arm_gripper,deformable,10000,0.30,0.05,pi05_lora/single_task/StraightenTwoRopesSequential-v1/lora_step_0010000.npz,
45
+ single_task,ClutterGrasp-v1,level_3,single_arm_dex,rigid_body,25000,0.25,0.25,pi05_lora/single_task/ClutterGrasp-v1/lora_step_0025000.npz,
46
+ single_task,DualArmFoldCarton-v1,level_3,dual_arm_gripper,rigid_body,40000,0.25,0.25,pi05_lora/single_task/DualArmFoldCarton-v1/lora_step_0040000.npz,
47
+ single_task,DualArmPegInHole-v1,level_2,dual_arm_gripper,rigid_body,5000,0.25,0.05,pi05_lora/single_task/DualArmPegInHole-v1/lora_step_0005000.npz,
48
+ single_task,OpenBoxRetrieval-v1,level_2,single_arm_gripper,rigid_body,20000,0.25,0.15,pi05_lora/single_task/OpenBoxRetrieval-v1/lora_step_0020000.npz,
49
+ single_task,ScoopWater-v1,level_2,single_arm_gripper,deformable,20000,0.25,0.15,pi05_lora/single_task/ScoopWater-v1/lora_step_0020000.npz,
50
+ single_task,DexOpenDoor-v1,level_2,single_arm_dex,rigid_body,40000,0.20,0.20,pi05_lora/single_task/DexOpenDoor-v1/lora_step_0040000.npz,
51
+ single_task,DualArmPingpongBalance-v1,level_3,dual_arm_dex,rigid_body,20000,0.20,0.20,pi05_lora/single_task/DualArmPingpongBalance-v1/lora_step_0020000.npz,
52
+ single_task,DualArmSendPingpong-v1,level_4,dual_arm_dex,rigid_body,35000,0.15,0.00,pi05_lora/single_task/DualArmSendPingpong-v1/lora_step_0035000.npz,
53
+ single_task,DualArmSetupTable-v1,level_3,dual_arm_dex,rigid_body,25000,0.15,0.00,pi05_lora/single_task/DualArmSetupTable-v1/lora_step_0025000.npz,
54
+ single_task,DualArmStraightenRope-v1,level_2,dual_arm_gripper,deformable,30000,0.15,0.05,pi05_lora/single_task/DualArmStraightenRope-v1/lora_step_0030000.npz,
55
+ single_task,DualArmSwingCup-v1,level_1,dual_arm_dex,rigid_body,40000,0.15,0.15,pi05_lora/single_task/DualArmSwingCup-v1/lora_step_0040000.npz,
56
+ single_task,DualArmThreadRopeThroughTube-v1,level_3,dual_arm_gripper,deformable,5000,0.15,0.10,pi05_lora/single_task/DualArmThreadRopeThroughTube-v1/lora_step_0005000.npz,
57
+ single_task,HammerNail-v1,level_3,single_arm_dex,rigid_body,10000,0.15,0.15,pi05_lora/single_task/HammerNail-v1/lora_step_0010000.npz,
58
+ single_task,BuildTent-v1,level_4,single_arm_gripper,deformable,5000,0.10,0.00,pi05_lora/single_task/BuildTent-v1/lora_step_0005000.npz,
59
+ single_task,DualArmOpenBottle-v1,level_2,dual_arm_dex,rigid_body,30000,0.10,0.05,pi05_lora/single_task/DualArmOpenBottle-v1/lora_step_0030000.npz,
60
+ single_task,DualArmOpenScissor-v1,level_1,dual_arm_dex,rigid_body,30000,0.10,0.00,pi05_lora/single_task/DualArmOpenScissor-v1/lora_step_0030000.npz,
61
+ single_task,KitAssembly-v1,level_2,single_arm_gripper,rigid_body,15000,0.10,0.00,pi05_lora/single_task/KitAssembly-v1/lora_step_0015000.npz,
62
+ single_task,PingPongBallTransport-v1,level_4,single_arm_gripper,rigid_body,25000,0.10,0.05,pi05_lora/single_task/PingPongBallTransport-v1/lora_step_0025000.npz,
63
+ single_task,TennisRally-v1,level_4,single_arm_dex,rigid_body,15000,0.10,0.00,pi05_lora/single_task/TennisRally-v1/lora_step_0015000.npz,
64
+ single_task,DualArmJugglingBall-v1,level_2,dual_arm_dex,rigid_body,30000,0.05,0.05,pi05_lora/single_task/DualArmJugglingBall-v1/lora_step_0030000.npz,
65
+ single_task,DualArmJumpRope-v1,level_4,dual_arm_dex,deformable,40000,0.05,0.05,pi05_lora/single_task/DualArmJumpRope-v1/lora_step_0040000.npz,
66
+ single_task,FlipPan-v1,level_4,single_arm_gripper,rigid_body,40000,0.05,0.05,pi05_lora/single_task/FlipPan-v1/lora_step_0040000.npz,
67
+ single_task,HitBox-v1,level_3,single_arm_dex,deformable,20000,0.05,0.05,pi05_lora/single_task/HitBox-v1/lora_step_0020000.npz,
68
+ single_task,MobilePickPlace-v1,level_3,single_arm_dex,rigid_body,10000,0.05,0.05,pi05_lora/single_task/MobilePickPlace-v1/lora_step_0010000.npz,
69
+ single_task,DualArmBallClassify-v1,level_4,dual_arm_gripper,rigid_body,10000,0.00,0.00,pi05_lora/single_task/DualArmBallClassify-v1/lora_step_0010000.npz,
70
+ single_task,DualArmBilliardBridge-v1,level_4,dual_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmBilliardBridge-v1/lora_step_0005000.npz,
71
+ single_task,DualArmClothHanger-v1,level_4,dual_arm_gripper,deformable,5000,0.00,0.00,pi05_lora/single_task/DualArmClothHanger-v1/lora_step_0005000.npz,
72
+ single_task,DualArmFlipTable-v1,level_4,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmFlipTable-v1/lora_step_0005000.npz,
73
+ single_task,DualArmKeyboardType-v1,level_4,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmKeyboardType-v1/lora_step_0005000.npz,
74
+ single_task,DualArmMahjongSetup-v1,level_4,dual_arm_gripper,rigid_body,10000,0.00,0.00,pi05_lora/single_task/DualArmMahjongSetup-v1/lora_step_0010000.npz,
75
+ single_task,DualArmMakeLine-v1,level_3,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmMakeLine-v1/lora_step_0005000.npz,
76
+ single_task,DualArmOpenFridge-v1,level_3,dual_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmOpenFridge-v1/lora_step_0005000.npz,
77
+ single_task,DualArmPlugCharger-v1,level_3,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmPlugCharger-v1/lora_step_0005000.npz,
78
+ single_task,DualArmPushObject-v1,level_2,dual_arm_gripper,deformable,5000,0.00,0.00,pi05_lora/single_task/DualArmPushObject-v1/lora_step_0005000.npz,
79
+ single_task,DualArmShapeDoll-v1,level_3,dual_arm_gripper,deformable,5000,0.00,0.00,pi05_lora/single_task/DualArmShapeDoll-v1/lora_step_0005000.npz,
80
+ single_task,DualArmShootBasketball-v1,level_4,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmShootBasketball-v1/lora_step_0005000.npz,
81
+ single_task,DualArmThrowBlock-v1,level_2,dual_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/DualArmThrowBlock-v1/lora_step_0005000.npz,
82
+ single_task,FitnessBall-v1,level_4,single_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/FitnessBall-v1/lora_step_0005000.npz,
83
+ single_task,GroceryPrepare-v1,level_3,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/GroceryPrepare-v1/lora_step_0005000.npz,
84
+ single_task,HitBilliard-v1,level_4,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/HitBilliard-v1/lora_step_0005000.npz,
85
+ single_task,PushTargetPose-v1,level_3,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/PushTargetPose-v1/lora_step_0005000.npz,
86
+ single_task,RingTossing-v1,level_4,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/RingTossing-v1/lora_step_0005000.npz,
87
+ single_task,RodThroughHoles-v1,level_4,single_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/RodThroughHoles-v1/lora_step_0005000.npz,
88
+ single_task,SeparateTwoRopesEnv-v1,level_3,single_arm_gripper,deformable,5000,0.00,0.00,pi05_lora/single_task/SeparateTwoRopesEnv-v1/lora_step_0005000.npz,
89
+ single_task,SpinPenHora-v1,level_4,single_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/SpinPenHora-v1/lora_step_0005000.npz,
90
+ single_task,Stack6Blocks-v1,level_3,single_arm_gripper,rigid_body,5000,0.00,0.00,pi05_lora/single_task/Stack6Blocks-v1/lora_step_0005000.npz,
91
+ single_task,TweezersPickPlace-v1,level_3,single_arm_dex,rigid_body,5000,0.00,0.00,pi05_lora/single_task/TweezersPickPlace-v1/lora_step_0005000.npz,
92
+ suite,l1_dual_dex,,,,35000,,,pi05_lora/suite/l1_dual_dex/lora_step_0035000.npz,
93
+ suite,l2_dual_gripper,,,,40000,,,pi05_lora/suite/l2_dual_gripper/lora_step_0040000.npz,
94
+ suite,l2_single_dex,,,,40000,,,pi05_lora/suite/l2_single_dex/lora_step_0040000.npz,
95
+ suite,l2_single_gripper,,,,40000,,,pi05_lora/suite/l2_single_gripper/lora_step_0040000.npz,
96
+ suite,l3_single_gripper,,,,40000,,,pi05_lora/suite/l3_single_gripper/lora_step_0040000.npz,
97
+ suite,l4_single_gripper,,,,40000,,,pi05_lora/suite/l4_single_gripper/lora_step_0040000.npz,