Hugging Face's logo

Datasets:
loop-wam
/
loopwam

Formats:
json
Size:
< 1K
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
1
loop-wam commited on
Commit
c7c9f35
·
verified ·
1 Parent(s): 552ebba

Add project-page figures, videos, and comparison tables

Browse files
Files changed (1) hide show
  1. README.md +169 -17
README.md CHANGED
@@ -8,23 +8,94 @@ pretty_name: LoopWAM checkpoints
8
 
9
  [Project page](https://loopwam.github.io/) · [Code](https://github.com/loop-wam/LoopWAM)
10
 
11
- LoopWAM turns deployment into continued learning. Expert demonstrations, together with successful, suboptimal, and failed rollouts, train one world–action model. Every rollout supervises future video. Action imitation is weighted by trajectory quality, and failed actions are not imitated.
12
 
13
- Demos, task videos, and paired action–video generation are on the [project page](https://loopwam.github.io/).
14
 
15
- ## Results
 
 
 
 
 
 
 
 
 
 
 
 
16
 
17
  Best overall performance among the methods compared in the paper.
18
 
19
  | Benchmark | LoopWAM |
20
  | --- | --- |
21
- | [RoboChallenge Table30-V2](https://loopwam.github.io/) | **43.33%** success · **56.80** score |
22
- | [RoboTwin 2.0](https://loopwam.github.io/) | **94.5%** average (clean 94.6, randomized 94.4) |
23
- | [RoboCasa365](https://loopwam.github.io/) | **49.5%** overall |
24
-
25
- ### Table30-V2 self-improvement
26
-
27
- 30 real tasks, four embodiments, one multi-task model per embodiment. Three deployment rounds:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28
 
29
  | Stage | Success | Score |
30
  | --- | ---: | ---: |
@@ -33,17 +104,98 @@ Best overall performance among the methods compared in the paper.
33
  | Round 2 | 37.67% | 51.65 |
34
  | Round 3 | 43.33% | 56.80 |
35
 
36
- Per-embodiment success at Round 3: ALOHA 26.00, DOS-W1 42.00, ARX5 55.71, UR5 76.67. Overall score uses task weights 10 / 10 / 7 / 3.
37
 
38
- Shirt folding, after single-task post-training, rises from 13.20% to 32.54% and then to **55.56%** across two further deployment rounds.
39
 
40
- ## Method
41
 
42
- Video and action tokens interact through bidirectional attention while they are denoised together.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
43
 
44
- - **Source-dependent noise.** Expert data trains joint generation. Lower-noise rollout actions teach the world model the consequences that were observed.
45
- - **Quality-aware learning.** Quality conditions both streams. Action loss increases with quality. Failed rollouts keep full video supervision and receive no direct action-imitation loss.
46
- - **Expert-quality steering.** At deployment, an expert-quality target steers the policy.
47
 
48
  ## Checkpoints
49
 
 
8
 
9
  [Project page](https://loopwam.github.io/) · [Code](https://github.com/loop-wam/LoopWAM)
10
 
11
+ LoopWAM turns real-world experience into better robot policies. Expert demonstrations sit alongside successful, suboptimal, and failed rollouts. Every rollout teaches the world model. Action imitation is weighted by quality, and failed actions are not copied.
12
 
13
+ ![LoopWAM self-improvement loop: deploy, collect mixed-quality rollouts, and improve through quality-aware learning.](https://loopwam.github.io/assets/branding/teaser.webp)
14
 
15
+ <video src="https://loopwam.github.io/assets/videos/overview.mp4" poster="https://loopwam.github.io/assets/posters/overview.jpg" controls playsinline width="100%"></video>
16
+
17
+ ## Method
18
+
19
+ ![Mixed-quality expert and rollout data train video and action branches through joint attention, source-dependent noise, quality-weighted supervision, and quality steering.](https://loopwam.github.io/assets/figures/method.webp)
20
+
21
+ Video and action tokens interact through bidirectional attention while they are denoised together.
22
+
23
+ - **Joint video–action modeling.** The two streams attend to each other at every layer.
24
+ - **Source-dependent noise.** Lower-noise rollout actions guide video prediction toward the physical consequences that were observed.
25
+ - **Quality-aware learning.** Quality conditions and weights guide action learning. Every rollout teaches video prediction, without imitating failed actions. At deployment, an expert-quality target steers the policy.
26
+
27
+ ## Leaderboards
28
 
29
  Best overall performance among the methods compared in the paper.
30
 
31
  | Benchmark | LoopWAM |
32
  | --- | --- |
33
+ | RoboChallenge Table30-V2 | **43.33%** success · **56.80** score |
34
+ | RoboTwin 2.0 | **94.5%** average |
35
+ | RoboCasa365 | **49.5%** overall |
36
+
37
+ ### RoboChallenge Table30-V2
38
+
39
+ Real-world, 30 tasks, 4 embodiments. One multi-task model per embodiment. Overall score uses task weights 10 / 10 / 7 / 3.
40
+
41
+ | Method | ALOHA | DOS-W1 | ARX5 | UR5 | Score | SR (%) |
42
+ | --- | ---: | ---: | ---: | ---: | ---: | ---: |
43
+ | **LoopWAM** | **26.00** | **42.00** | 55.71 | **76.67** | **56.80** | **43.33** |
44
+ | VLA-DM0.5 | 18.00 | 42.00 | **58.57** | 70.00 | 54.42 | 40.67 |
45
+ | my16 | 13.00 | 27.00 | 54.29 | 46.67 | 41.27 | 30.67 |
46
+ | LoopWAM (expert-trained) | 8.00 | 38.00 | 24.29 | 26.67 | 36.10 | 23.67 |
47
+ | JV0 | 5.00 | 16.00 | 20.00 | 43.33 | 27.58 | 16.00 |
48
+ | mc | 7.00 | 0.00 | 47.14 | 20.00 | 23.63 | 15.33 |
49
+ | MagicBot | 0.00 | 5.00 | 32.86 | 53.33 | 20.82 | 14.67 |
50
+ | holobrain | 3.00 | 6.00 | 30.00 | 43.33 | 27.93 | 14.33 |
51
+ | Baseline Model | 6.00 | 6.00 | 32.86 | 26.67 | 31.48 | 14.33 |
52
+ | test_01 | 1.00 | 3.00 | 28.57 | 43.33 | 22.73 | 12.33 |
53
+ | W0 | 2.00 | 0.00 | 30.00 | 43.33 | 21.85 | 12.00 |
54
+ | debuggers_vla | 0.00 | 2.00 | 34.29 | 23.33 | 18.72 | 11.00 |
55
+
56
+ Columns ALOHA, DOS-W1, ARX5, and UR5 are success rates (%).
57
+
58
+ ### RoboTwin 2.0
59
+
60
+ 50 bimanual tasks. No additional embodied pretraining. Clean 94.6%, randomized 94.4%.
61
+
62
+ | Method | Embodied pretraining | Clean | Randomized | Average |
63
+ | --- | --- | ---: | ---: | ---: |
64
+ | **LoopWAM** | No | **94.6** | **94.4** | **94.5** |
65
+ | LingBot-VA 2.0 | Yes | 93.8 | 93.4 | 93.6 |
66
+ | AHA-WAM | No | 93.4 | 92.2 | 92.8 |
67
+ | LingBot-VA | Yes | 92.9 | 91.5 | 92.2 |
68
+ | Fast-WAM | No | 91.9 | 91.8 | 91.8 |
69
+ | Motus | Yes | 88.7 | 87.0 | 87.8 |
70
+ | ABot-M0 | Yes | 81.2 | 80.4 | 80.8 |
71
+ | π0.5 | Yes | 82.7 | 76.8 | 79.8 |
72
+ | Motus (from Wan2.2) | No | 77.6 | 77.0 | 77.3 |
73
+ | X-VLA | Yes | 72.9 | 72.8 | 72.9 |
74
+ | π0 | Yes | 65.9 | 58.4 | 62.2 |
75
+
76
+ ### RoboCasa365
77
+
78
+ 50 evaluation tasks. Overall success uses weights 18 / 16 / 16. All 16 Composite-Unseen tasks are held out of training.
79
+
80
+ | Method | Embodied pretraining | Atomic seen | Composite seen | Composite unseen | Overall |
81
+ | --- | --- | ---: | ---: | ---: | ---: |
82
+ | **LoopWAM** | No | **79.7** | **50.3** | 14.9 | **49.5** |
83
+ | ABot-M0.6 | Yes | 79.4 | 48.3 | 7.9 | 46.6 |
84
+ | ABot-M0.5 | Yes | 75.6 | 37.7 | 3.3 | 40.3 |
85
+ | PRTS | Yes | 66.3 | 30.3 | **18.8** | 39.6 |
86
+ | RLDX-1 | Yes | 67.6 | 27.9 | 8.5 | 36.0 |
87
+ | GR00T N1.5 | Yes | 50.7 | 14.8 | 2.7 | 23.9 |
88
+ | GR00T N1.6 | Yes | 51.1 | 9.4 | 1.7 | 21.9 |
89
+ | GigaWorld-Policy 0.1 | Yes | 44.4 | 11.8 | 2.9 | 20.7 |
90
+ | π0.5 | Yes | 39.6 | 7.1 | 1.2 | 16.9 |
91
+ | π0 | Yes | 34.6 | 6.1 | 1.1 | 14.8 |
92
+ | Diffusion Policy | No | 15.7 | 0.2 | 1.3 | 6.1 |
93
+
94
+ ## Multi-task self-improvement
95
+
96
+ 30 tasks, four embodiments, three rounds of real-world deployment. The clip below is 30 full rollouts, sped up to 30 seconds.
97
+
98
+ <video src="https://loopwam.github.io/assets/videos/rollout-30-tasks.mp4" poster="https://loopwam.github.io/assets/posters/rollout-30-tasks.jpg" controls playsinline width="100%"></video>
99
 
100
  | Stage | Success | Score |
101
  | --- | ---: | ---: |
 
104
  | Round 2 | 37.67% | 51.65 |
105
  | Round 3 | 43.33% | 56.80 |
106
 
107
+ ### Tasks
108
 
109
+ Click a still to open the rollout video.
110
 
111
+ **ALOHA**
112
 
113
+ [![Lint Roller](https://loopwam.github.io/assets/posters/tasks/lint_roller_remove_dirt.jpg)](https://loopwam.github.io/assets/videos/tasks/lint_roller_remove_dirt.mp4)
114
+ [![Spread Jam](https://loopwam.github.io/assets/posters/tasks/paint_jam.jpg)](https://loopwam.github.io/assets/videos/tasks/paint_jam.mp4)
115
+ [![Put Books Back](https://loopwam.github.io/assets/posters/tasks/put_the_books_back.jpg)](https://loopwam.github.io/assets/videos/tasks/put_the_books_back.mp4)
116
+ [![Scoop with a Spoon](https://loopwam.github.io/assets/posters/tasks/scoop_with_a_small_spoon.jpg)](https://loopwam.github.io/assets/videos/tasks/scoop_with_a_small_spoon.mp4)
117
+ [![Stamp Positioning](https://loopwam.github.io/assets/posters/tasks/stamp_positioning.jpg)](https://loopwam.github.io/assets/videos/tasks/stamp_positioning.mp4)
118
+ [![Wipe Blackboard](https://loopwam.github.io/assets/posters/tasks/wipe_the_blackboard.jpg)](https://loopwam.github.io/assets/videos/tasks/wipe_the_blackboard.mp4)
119
+ [![Wrap with Cloth](https://loopwam.github.io/assets/posters/tasks/wrap_with_a_soft_cloth.jpg)](https://loopwam.github.io/assets/videos/tasks/wrap_with_a_soft_cloth.mp4)
120
+
121
+ **ARX5**
122
+
123
+ [![Arrange Flowers](https://loopwam.github.io/assets/posters/tasks/arrange_flowers.jpg)](https://loopwam.github.io/assets/videos/tasks/arrange_flowers.mp4)
124
+ [![Hang the Cup](https://loopwam.github.io/assets/posters/tasks/hang_the_cup.jpg)](https://loopwam.github.io/assets/videos/tasks/hang_the_cup.mp4)
125
+ [![Pick Green Blocks](https://loopwam.github.io/assets/posters/tasks/pick_out_the_green_blocks.jpg)](https://loopwam.github.io/assets/videos/tasks/pick_out_the_green_blocks.mp4)
126
+ [![Press the Button](https://loopwam.github.io/assets/posters/tasks/press_the_button.jpg)](https://loopwam.github.io/assets/videos/tasks/press_the_button.mp4)
127
+ [![Light Switch](https://loopwam.github.io/assets/posters/tasks/turn_on_the_light_switch.jpg)](https://loopwam.github.io/assets/videos/tasks/turn_on_the_light_switch.mp4)
128
+ [![Water Flowers](https://loopwam.github.io/assets/posters/tasks/water_the_flowers.jpg)](https://loopwam.github.io/assets/videos/tasks/water_the_flowers.mp4)
129
+ [![Wipe the Table](https://loopwam.github.io/assets/posters/tasks/wipe_the_table.jpg)](https://loopwam.github.io/assets/videos/tasks/wipe_the_table.mp4)
130
+
131
+ **UR5**
132
+
133
+ [![Arrange Fruits](https://loopwam.github.io/assets/posters/tasks/arrange_fruits.jpg)](https://loopwam.github.io/assets/videos/tasks/arrange_fruits.mp4)
134
+ [![Item Classification](https://loopwam.github.io/assets/posters/tasks/item_classification.jpg)](https://loopwam.github.io/assets/videos/tasks/item_classification.mp4)
135
+ [![Shred Paper](https://loopwam.github.io/assets/posters/tasks/shred_paper.jpg)](https://loopwam.github.io/assets/videos/tasks/shred_paper.mp4)
136
+
137
+ **DOS-W1**
138
+
139
+ [![Hold the Tray](https://loopwam.github.io/assets/posters/tasks/hold_the_tray_with_both_hands.jpg)](https://loopwam.github.io/assets/videos/tasks/hold_the_tray_with_both_hands.mp4)
140
+ [![Desk Drawer](https://loopwam.github.io/assets/posters/tasks/place_objects_into_desk_drawer.jpg)](https://loopwam.github.io/assets/videos/tasks/place_objects_into_desk_drawer.mp4)
141
+ [![Pen Container](https://loopwam.github.io/assets/posters/tasks/put_in_pen_container.jpg)](https://loopwam.github.io/assets/videos/tasks/put_in_pen_container.mp4)
142
+ [![Put Shoes Back](https://loopwam.github.io/assets/posters/tasks/put_the_shoes_back.jpg)](https://loopwam.github.io/assets/videos/tasks/put_the_shoes_back.mp4)
143
+ [![Stack Bowls](https://loopwam.github.io/assets/posters/tasks/stack_bowls.jpg)](https://loopwam.github.io/assets/videos/tasks/stack_bowls.mp4)
144
+ [![Sweep the Trash](https://loopwam.github.io/assets/posters/tasks/sweep_the_trash.jpg)](https://loopwam.github.io/assets/videos/tasks/sweep_the_trash.mp4)
145
+ [![Makeup Table](https://loopwam.github.io/assets/posters/tasks/tidy_up_the_makeup_table.jpg)](https://loopwam.github.io/assets/videos/tasks/tidy_up_the_makeup_table.mp4)
146
+ [![Tie a Knot](https://loopwam.github.io/assets/posters/tasks/tie_a_knot.jpg)](https://loopwam.github.io/assets/videos/tasks/tie_a_knot.mp4)
147
+
148
+ ## Shirt folding
149
+
150
+ After expert post-training, two deployment rounds improve DOS-W1 shirt folding from 13.20% to 55.56%.
151
+
152
+ | Stage | Training data | Success | Selected trials |
153
+ | --- | --- | ---: | --- |
154
+ | Stage 1 | Expert demonstrations | 13.20% | 1 success, 8 failures |
155
+ | Stage 2 | Expert + round 1 rollouts | 32.54% | 3 successes, 6 failures |
156
+ | Stage 3 | Expert + rounds 1 and 2 | 55.56% | 5 successes, 4 failures |
157
+
158
+ Stage 1, the success:
159
+
160
+ [![Stage 1 success](https://loopwam.github.io/assets/posters/folding/s16.jpg)](https://loopwam.github.io/assets/videos/folding/s16.mp4)
161
+
162
+ Stage 2 successes:
163
+
164
+ [![Stage 2](https://loopwam.github.io/assets/posters/folding/s11.jpg)](https://loopwam.github.io/assets/videos/folding/s11.mp4)
165
+ [![Stage 2](https://loopwam.github.io/assets/posters/folding/s10.jpg)](https://loopwam.github.io/assets/videos/folding/s10.mp4)
166
+ [![Stage 2](https://loopwam.github.io/assets/posters/folding/s9.jpg)](https://loopwam.github.io/assets/videos/folding/s9.mp4)
167
+
168
+ Stage 3 successes:
169
+
170
+ [![Stage 3](https://loopwam.github.io/assets/posters/folding/s5.jpg)](https://loopwam.github.io/assets/videos/folding/s5.mp4)
171
+ [![Stage 3](https://loopwam.github.io/assets/posters/folding/s4.jpg)](https://loopwam.github.io/assets/videos/folding/s4.mp4)
172
+ [![Stage 3](https://loopwam.github.io/assets/posters/folding/s3.jpg)](https://loopwam.github.io/assets/videos/folding/s3.mp4)
173
+ [![Stage 3](https://loopwam.github.io/assets/posters/folding/s2.jpg)](https://loopwam.github.io/assets/videos/folding/s2.mp4)
174
+ [![Stage 3](https://loopwam.github.io/assets/posters/folding/s1.jpg)](https://loopwam.github.io/assets/videos/folding/s1.mp4)
175
+
176
+ ## Action and video generation
177
+
178
+ Each pair is the observed video on the left and LoopWAM’s prediction on the right.
179
+
180
+ **Arrange Fruits**
181
+
182
+ [![Arrange Fruits, observed](https://loopwam.github.io/assets/posters/generation/arrange_fruits_episode_000000.jpg)](https://loopwam.github.io/assets/videos/generation/arrange_fruits_episode_000000.mp4)
183
+ [![Arrange Fruits, predicted](https://loopwam.github.io/assets/posters/generation/arrange_fruits_episode_000002.jpg)](https://loopwam.github.io/assets/videos/generation/arrange_fruits_episode_000002.mp4)
184
+
185
+ **Pen Container**
186
+
187
+ [![Pen Container, observed](https://loopwam.github.io/assets/posters/generation/put_in_pen_container_episode_000001.jpg)](https://loopwam.github.io/assets/videos/generation/put_in_pen_container_episode_000001.mp4)
188
+ [![Pen Container, predicted](https://loopwam.github.io/assets/posters/generation/put_in_pen_container_episode_000003.jpg)](https://loopwam.github.io/assets/videos/generation/put_in_pen_container_episode_000003.mp4)
189
+
190
+ **Stack Bowls**
191
+
192
+ [![Stack Bowls, observed](https://loopwam.github.io/assets/posters/generation/stack_bowls_episode_000005.jpg)](https://loopwam.github.io/assets/videos/generation/stack_bowls_episode_000005.mp4)
193
+ [![Stack Bowls, predicted](https://loopwam.github.io/assets/posters/generation/stack_bowls_episode_000007.jpg)](https://loopwam.github.io/assets/videos/generation/stack_bowls_episode_000007.mp4)
194
+
195
+ **Lint Roller**
196
 
197
+ [![Lint Roller, observed](https://loopwam.github.io/assets/posters/generation/lint_roller_remove_dirt_episode_000004.jpg)](https://loopwam.github.io/assets/videos/generation/lint_roller_remove_dirt_episode_000004.mp4)
198
+ [![Lint Roller, predicted](https://loopwam.github.io/assets/posters/generation/lint_roller_remove_dirt_episode_000006.jpg)](https://loopwam.github.io/assets/videos/generation/lint_roller_remove_dirt_episode_000006.mp4)
 
199
 
200
  ## Checkpoints
201