Video-only WAM checkpoints
Checkpoints for the paper Few Actions Suffice, Diverse Videos Matter for World Action Models. Code and evaluation instructions: https://github.com/kongwanbianjinyu/video-only-wam
Each folder is a self-contained OpenWAM checkpoint (video expert, action expert, VAE, text encoder,
normalization statistics, tokenizer) that scripts/deploy.sh loads directly.
| Folder | Training data | Table 1 row |
|---|---|---|
vlabench/few_action |
20% video-action (1,000 demos) | r video-action |
vlabench/ours |
20% video-action + 80% video-only | r video-action + (1-r) video-only |
vlabench/all_action |
100% video-action (4,800 demos) | 100% video-action |
robotwin/few_action |
4% video-action (2 clean + 20 randomized per task) | r video-action |
robotwin/ours |
4% video-action + 96% video-only | r video-action + (1-r) video-only |
robotwin/all_action |
100% video-action (27,500 demos) | 100% video-action |