FastWAM Checkpoints

Official checkpoints for Fast-WAM: Do World Action Models Need Test-time Future Imagination?

This repository includes the original FastWAM checkpoints and the new Optional IDM variant: one model with two inference modes for studying when test-time future imagination helps.

Files

README.md
libero_uncond_2cam224.pt
libero_uncond_2cam224_dataset_stats.json
libero_optional_idm_2cam224.pt
libero_optional_idm_2cam224_dataset_stats.json
robotwin_uncond_3cam_384.pt
robotwin_uncond_3cam_384_dataset_stats.json

Included checkpoints

• libero_uncond_2cam224.pt: FastWAM Base for task=libero_uncond_2cam224_1e-4.
• libero_optional_idm_2cam224.pt: Optional IDM for task=libero_optional_idm_2cam224_1e-4.
• robotwin_uncond_3cam_384.pt: FastWAM Base for task=robotwin_uncond_3cam_384_1e-4.
• Each checkpoint has a matching *_dataset_stats.json file for evaluation.

Optional IDM: one model, two thinking modes

The Optional IDM checkpoint supports two action-inference paths without retraining:

• idm: imagine the future video first, then predict actions.
• first_frame: predict actions directly from the current observation without test-time future imagination.

Evaluate either mode with the FastWAM repository:

python experiments/libero/run_libero_manager.py \
  task=libero_optional_idm_2cam224_1e-4 \
  ckpt=./checkpoints/fastwam_release/libero_optional_idm_2cam224.pt \
  EVALUATION.dataset_stats_path=./checkpoints/fastwam_release/libero_optional_idm_2cam224_dataset_stats.json \
  EVALUATION.sigma_shift=1.0 \
  +EVALUATION.action_infer_mode=idm \
  MULTIRUN.num_gpus=8

Replace idm with first_frame to use the Fast-WAM inference mode. The released checkpoint uses an action scheduler shift of 1.0 for training and evaluation.

On the full LIBERO benchmark (40 tasks, 50 episodes per task), this checkpoint achieves:

Inference mode	Spatial	Goal	Object	Long	Average
IDM	99.0%	98.6%	99.6%	97.0%	98.55%
First frame (Fast-WAM)	98.2%	97.8%	99.2%	95.8%	97.75%

Project

• Project page: https://yuantianyuan01.github.io/FastWAM/
• Paper: https://arxiv.org/abs/2603.16666
• Code: https://github.com/yuantianyuan01/FastWAM

Citation

@misc{yuan2026fastwam,
  title={Fast-WAM: Do World Action Models Need Test-time Future Imagination?},
  author={Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao},
  year={2026},
  note={arXiv preprint arXiv:2603.16666}
}