WorldScape-Policy-2

WorldScape-Policy-2

WorldScape-Policy-2: Empowering Steerable World Action Modeling with Reasoning-Augmented Memory and In-Context Learning

WorldScape Policy 2.0 is a controllable World Action Model (WAM) with reasoning-augmented long short-term memory for long-horizon robotic manipulation, fine-grained instruction following, visual-context reasoning, and in-context skill transfer.

Model Checkpoints

• Pretrained checkpoint: wsp_2_pretrain
• RoboTwin2.0-C2R Interactive post-trained checkpoint: wsp_2_posttrain_robotwin2_c2r
• RoboTwin2.0-C2R Auto post-trained checkpoint: wsp_2_posttrain_robotwin2_c2r_auto
• RoboTwin2.0-Full post-trained checkpoint: wsp_2_posttrain_robotwin2_full

Project Page

https://manifoldai-research.github.io/WorldScape-Policy/

Citation

@article{worldscape-policy-2,
  title={WorldScape Policy 2.0: Empowering Steerable World Action Modeling with Reasoning-Augmented Memory and In-Context Learning},
  author={Haisheng Su and Zongdai Liu and Xin Jin and Haoxuan Dou and Chengming Hu and Baorun Li and Zhanwang Liu and Ruiyan Xu and Jianjie Fang and Xin Zhang and Zhenjie Yang and Xue Yang and Chen Gao and Junchi Yan and Yong Li and Wei Wu},
  journal={arXiv preprint arXiv:2607.18840},
  year={2026}
}