SWAM is a world action model from five collaborating institutions. It connects video and action streams through masked attention to jointly predict video and robot actions conditioned on language. The release provides pretrained and RoboTwin 2.0 Clean post-trained weights and their corresponding DMuon training configurations.

Overview

Model type
Action policy
Architecture
WAM
Action generation
Flow matching
Listed versions
SWAM

Papers & reports (1)

SWAM: Scaling World Action Models with Cross-Modal Alignment

Paper available

arXiv: 2611.12870DOI: 10.48550/arxiv.2611.12870
Authors & affiliations (11)

Author order: Original order

  1. Yu Zhou

    Shanghai Jiao Tong University · Shanghai Innovation Institute · Noematrix Ltd.

  2. Xinyuan Xia

    Shanghai Jiao Tong University · Shanghai Innovation Institute · Noematrix Ltd.

  3. Xuankun Yang

    Shanghai Jiao Tong University · Noematrix Ltd.

  4. Yinghao Yang

    Westlake University · Shanghai Innovation Institute · Noematrix Ltd.

  5. Wendi Chen

    Shanghai Jiao Tong University · Shanghai Innovation Institute · Noematrix Ltd.

  6. Bowen Liu

    Shanghai Jiao Tong University · Shanghai Innovation Institute · Noematrix Ltd.

  7. Jianhan Ma

    Noematrix Ltd.

  8. Weicheng Duan

    Tongji University · Noematrix Ltd.

  9. Jun Lv

    Noematrix Ltd.

First published
2026-11-10
Latest paper revision
v2 · 2026-11-11
Venue
Not provided

Covered model versions

Paper ↗PDF ↗

Official resources

Submit a source or correction