WAM-Policy adapts LingBot-VA into a language-conditioned world action policy for quadrupedal loco-manipulation. It combines on-policy self-distillation, task-specific reinforcement learning and multi-teacher distillation. The authors evaluate on QLM-Bench in MuJoCo and demonstrate the OPSD student on a two-camera Unitree Go2 Air. Supplementary LIBERO adapters cover robot-arm experiments.

Overview

Model type
Action policy
Architecture
WAM
Action generation
Flow matching
Listed versions
WAM-Policy Release snapshot c320076b

Papers & reports (1)

WAM-Policy: World Action Models with On-Policy Distillation for Quadrupedal Loco-Manipulation

Paper available

arXiv: 2611.24011DOI: 10.48550/arxiv.2611.24011
Authors & affiliations (6)

Author order: Original order

  1. Junlin LongEqual contribution

    Australian Centre for Robotics, USYD

  2. Xu DengEqual contribution

    UNSW Sydney

  3. Yiran Wang

    Australian Centre for Robotics, USYD

  4. Yue YangCorresponding author

    Maincode

  5. Yang ZhaoCorresponding author

    La Trobe University

  6. Zeyu ZhangProject lead

    Maincode

First published
2026-11-18
Latest paper revision
v1 · 2026-11-18
Venue
Not provided

Covered model versions

Paper ↗PDF ↗

Official resources

Submit a source or correction