PICA-WAM

Checkpoints for PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models (arXiv:2610.26813, project page).

InternW0-Delta-RoboTwin (preview)

File Backbone Benchmark Training data
internw0_delta/robotwin.pt InternW0-Δ, initialized from InternW0-Delta-Base RoboTwin 2.0, aloha_agilex, 14D joint Clean demonstrations only (50 tasks × 50)
  • Format and training. The file is an inference-only export of the step-5370 training weights; the PICA-WAM alignment head and EMA teacher, which are training-only, are removed. Global batch 1024, 10 epochs.
  • Reported results. Success rate, 50 tasks × 100 episodes: 90.68% clean, 74.70% randomized.
  • Usage. Training and evaluation code: InternW0-Delta-RoboTwin/ in the project repository (bash download_checkpoint.sh, then python -m eval.robotwin.run_robotwin). The model is also packaged as the XPolicyLab PICA_WAM adapter.

This is a preview release of one PICA-WAM model; further checkpoints will be added to this repository. The released weights are under the MIT License; base models (Wan2.2, RynnBrain, InternW0-Delta-Base) and the RoboTwin data keep their own licenses.

Citation

@misc{zou2026picawamvisualperturbationinvariance,
      title={PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models},
      author={Yude Zou and Yifei Yao and Guanbo Jiao and Jianming Ma and Guanjie Zheng and Xing Gao},
      year={2026},
      eprint={2610.26813},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.26813},
}
Downloads last month

-

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

Paper for xdzouyd/PICA-WAM