PrismWAM

Checkpoints for PrismWAM: Decoupling State Representations in World Action Models.

Code · Project page · arXiv

Folder	Benchmark	Actions	Result
robotwin/	RoboTwin 2.0, Clean → Randomized	14-D absolute joints	70.57
robodojo/	RoboDojo	14-D absolute joints	25.10 / 18.55
libero/	LIBERO, 4 suites	7-D delta end effector + gripper	99.35
init/	training initializers: depth VAEs, RGB VAE, action-expert init

Success rates in %; RoboDojo is score / success rate. Each model folder holds config.yaml, policy.pt, the RGB and depth VAEs, and its action statistics.

Usage

Install the code, then from the repository root:

hf download THUSI-Lab/PrismWAM --include "robotwin/*" --local-dir weights

from prismwam import load_pretrained

policy = load_pretrained("robotwin")   # also fetches the T5 text encoder on first use
observation["prompt"] = policy.format_prompt("beat the block with the hammer")
out = policy.predict(observation, seed=0)        # out["action"]: 32 x 14

The observation format, benchmark evaluation and training recipes are documented in the code repository.

Citation

@misc{prism,
  title         = {PrismWAM: Decoupling State Representations in World Action Models},
  author        = {Xin, Tianyu and Zhong, Wanqi and Hu, Bowen and Guo, Zhangchi and Wang, Lijie and Wang, Yiming and Liao, Haoran and Yu, Heyang and Tian, Xiaoyu and Yu, Zhiding and Li, Yiming},
  year          = {2026},
  eprint        = {2610.26564},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.26564}
}