RegisterWAM adds learnable registers to world action models for robot control
An arXiv paper posted on 13 Oct 2026, "Acting Through Registers" by Yiwei Zhang and nine co-authors, examines how World Action Models (WAMs) organize action-relevant information. The authors identify persistent high-norm spatial outliers in WAM representations whose action relevance survives normalization and K/V projection, and propose RegisterWAM, which adds learnable non-spatial registers to the world model and uses them directly for action conditioning. Registers alone do not remove the spatial outliers, but pairing them with observation flow-matching (FM) loss supervision concentrates high-norm responses in the registers; the authors report 81.64% success on LIBERO-Plus without extra embodied pretraining data or foundation models, 93.0% average success on RoboTwin 2.0, and improved real-world manipulation.
It offers a concrete mechanism for understanding and reorganizing action-relevant information inside world models, potentially helping robot policies generalize with less data.