File size: 3,123 Bytes
134c84a
 
b4981ec
 
 
134c84a
b4981ec
 
 
 
 
0b9208d
134c84a
 
 
 
b4981ec
 
134c84a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0ec42b3
134c84a
0ec42b3
 
 
 
 
 
 
 
 
134c84a
 
 
 
 
 
 
 
 
 
 
0b9208d
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
---
pretty_name: GlanceWAM reproduction bundle
license: other
task_categories:
- robotics
tags:
- robotics
- vision-language-action
- world-model
- libero
- robocasa
- arxiv:2608.23927
---

# GlanceWAM reproduction bundle

**Paper**: [GlanceWAM: Sparse Test-Time Imagination for World-Action Models](https://huggingface.co/papers/2608.23927)

Everything needed to reproduce the GlanceWAM results on LIBERO and RoboCasa kitchen.

```
datasets/     LeRobot v3 datasets, UMT5 text caches included (glancewam_cache/)
checkpoints/  released checkpoints, one directory per run
```

Point the code at this directory:

```bash
export DATA_ROOT=/data/glancewam_release/datasets           # training + precompute
ln -s /data/glancewam_release/checkpoints  <repo>/results/Checkpoints
```

## Checkpoints

| Directory | Benchmark | Reported |
|---|---|---|
| `glancewam_robocasa_kitchen` | RoboCasa kitchen, 24 tasks x 50 episodes | **0.721** |
| `glancewam_libero` | LIBERO 4-in-1, 4 suites x 500 episodes | **0.989** |

Kitchen run-to-run noise is ~0.02 (the environment is paired but the policy is unseeded), so
treat anything within ~+/-0.02 as a match. LIBERO is saturated; differences under ~0.005 are noise.

## Datasets

| Directory | Used by |
|---|---|
| `libero_{spatial,object,goal,10}_no_noops_1.0.0_lerobot` | LIBERO (mixture `libero_all`) |
| `robocasa_cosmos_kitchen/` | RoboCasa kitchen, 24 per-task datasets (mixture `robocasa_kitchen_all`) |

Each dataset already carries its precomputed UMT5 text cache under
`<dataset>/glancewam_cache/t5/Skywork_SkyReels-V2-DF-1.3B-540P-Diffusers_L512`, so training can run
with `RESIDENT_TEXT_TABLE=True` (the default) without running the precompute step first.

The dataset trees are hard-linked from `/data/lerobot_v3`, so they cost no extra disk on this
machine; copying the directory elsewhere produces independent full copies.

## Code

<https://github.com/linhanwang/GlanceWAM>

```bash
# everything (21 GB)
hf download LinhanWang/GlanceWAM --repo-type dataset --local-dir ./glancewam_bundle

# or just one benchmark
hf download LinhanWang/GlanceWAM --repo-type dataset --local-dir ./glancewam_bundle \
    --include "checkpoints/glancewam_libero/*" "datasets/libero_*"                    # 5.0 GB
hf download LinhanWang/GlanceWAM --repo-type dataset --local-dir ./glancewam_bundle \
    --include "checkpoints/glancewam_robocasa_kitchen/*" \
              "datasets/robocasa_cosmos_kitchen/*"                                     # 16.2 GB

# then, from the code checkout
mkdir -p results
ln -s /abs/path/to/glancewam_bundle/datasets     results/Datasets
ln -s /abs/path/to/glancewam_bundle/checkpoints  results/Checkpoints
```

## Attribution

The datasets here are derived from third-party releases and remain subject to their original
terms: **LIBERO** (Lifelong Robot Learning, MIT) and the **RoboCasa** kitchen task suite as
distributed by NVIDIA's cosmos-policy release (RoboCasa Team, MIT). The checkpoints are fine-tuned
from **SkyReels-V2-DF-1.3B-540P** (Skywork) and inherit that model's terms. The GlanceWAM code
itself is MIT-licensed; see the code repository.