THUSI-Lab/PrismWAM-Star

PrismWAM*

The PrismWAM* checkpoint for RoboTwin 2.0, from PrismWAM: Decoupling State Representations in World Action Models.

Code · Project page · arXiv · PrismWAM checkpoints

Folder	Benchmark	Actions	Result
robotwin/	RoboTwin 2.0, Clean → Randomized	14-D absolute joints	83.67

Average success rate in % over the 50 tasks in clean and randomized scenes, 100 episodes per task; trained on clean scenes only.

The model folder is laid out like those of THUSI-Lab/PrismWAM: config.yaml, policy.pt, the RGB and depth VAEs, and the action statistics. On first use the code also fetches the T5 text encoder of Wan2.2-TI2V-5B and the CLIP image encoder of Wan2.1-I2V-14B-480P.

Usage

Install the code, then from the repository root:

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

from prismwam import load_pretrained

policy = load_pretrained("prismwam_star/robotwin")   # uses weights_star/robotwin
observation["prompt"] = policy.format_prompt("lift the bottle with the right arm")
out = policy.predict(observation, seed=0)        # out["action"]: 32 x 14

The observation format and the RoboTwin evaluation 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}
}