LoopedWAM: Looped World Action Models for Elastic Inference

Checkpoints of LoopedWAM and LoopedWAM-E for LIBERO, RoboTwin 2.0 and RoboCasa-GR1.

Benchmark LoopedWAM LoopedWAM-E
LIBERO (4 suites) libero/LoopedWAM libero/LoopedWAM-E
RoboTwin 2.0 (50 tasks) robotwin50/LoopedWAM robotwin50/LoopedWAM-E
RoboCasa-GR1 (24 tabletop tasks) robocasa_gr1/LoopedWAM robocasa_gr1/LoopedWAM-E

LoopedWAM runs its shared action core K_a = 4 times per denoising step, with N = 10 denoising steps by default. LoopedWAM-E is a single checkpoint, distilled from the LoopedWAM of the same benchmark, that runs any K_a in {1, 2, 4} with any N in {2, 5, 10}, chosen at inference time.

Files

Each folder holds:

  • policy.pt: the FP32 weights (0.58B parameters). The architecture is stored in the checkpoint. A LoopedWAM-E checkpoint also contains its teacher's video expert and loads on its own.
  • dataset_stats.json: the normalization statistics of the actions and the proprioception.

All models also use the VAE (Wan2.1_VAE.pth) and the T5 text encoder (models_t5_umt5-xxl-enc-bf16.pth, google/umt5-xxl) of Wan-AI/Wan2.1-T2V-1.3B. SHA256SUMS lists the SHA-256 of every file.

Usage

hf download anhdao69/LoopedWAM --local-dir checkpoints/LoopedWAM
hf download Wan-AI/Wan2.1-T2V-1.3B --local-dir checkpoints/Wan2.1-T2V-1.3B

# LIBERO, LoopedWAM-E with K_a = 2 action loops and N = 2 denoising steps
python experiments/libero/eval_libero.py \
  --checkpoint checkpoints/LoopedWAM/libero/LoopedWAM-E/policy.pt \
  --action-loops 2 --steps 2 --gpus 0,1,2,3 --out evaluate_results/libero/LoopedWAM-E_k2_n2

The evaluation scripts for all three benchmarks, and the training code, are in the GitHub repository.

Citation

@article{dao2026loopedwam,
  title   = {LoopedWAM: Looped World Action Models for Elastic Inference},
  author  = {Dao, Anh and Liu, Xinxin and Yu, Hang and
             Chan Santiago, Jeffrey A. and Shang, Yuzhang},
  year    = {2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading