Driving · 2026-10-14

Acting Through Registers

Yiwei Zhang, Yantai Yang, Jin Gao, Jin Lou, Andong Chen, Yufei Liu, Boyang Xing, Yuchen Zhu, Weiming Hu, Zhipeng Zhang

arXiv:2610.17638PDFCode

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

World Action Models (WAMs) can show persistent high-norm spatial outlier tokens whose action relevance persists even after K/V projection attenuates their magnitude. The authors propose RegisterWAM: add learnable non-spatial register tokens for action conditioning, and pair them with observation flow-matching (FM) supervision to suppress spatial outliers while concentrating high-norm responses in registers. They also add noised-observation FM for recovery. RegisterWAM reaches 81.6% on LIBERO-Plus and 93.0% on RoboTwin 2.0.

Why it matters

The paper links internal token-norm anomalies to action-relevant behavior in WAMs and shows that representation organization can be changed with architectural non-spatial capacity plus observation FM supervision. The resulting action-conditioned world model achieves strong manipulation performance without additional embodied pretraining data or added foundation models (per the authors’ setup).

Method

  • Identify persistent high-norm spatial outliers in WAM representations; measure action sensitivity by masking visual token positions in Action DiT attention and find relevance persists despite K/V projection attenuation.
  • Introduce RegisterWAM by appending learnable non-spatial register tokens to the Video DiT; use observation FM loss supervision on dense clean observation tokens (excluding registers) to suppress spatial outliers and concentrate high-norm responses in registers.
  • Add noised-observation FM loss supervision (recovery from corrupted current observation inputs) as auxiliary training; at inference, Video DiT runs once to build register K/V cache reused during iterative action denoising.

Limitation

The authors’ evidence supports functional specialization but does not fully establish what information outlier or register tokens encode. The experiments cover a limited family of WAM architectures, and the observed behavior may depend on model scale, training objective, or optimization history.

Abstract (from arXiv)

World Action Models (WAMs) use generative visual representations to condition robot actions, yet how these representations organize action-relevant information remains poorly understood. We identify persistent high-norm spatial outliers in WAM representations and find that their action relevance persists even after normalization and K/V projection attenuate their extreme magnitude. Motivated by this observation, we introduce RegisterWAM, which adds learnable non-spatial registers to the world model and uses them directly for action conditioning. Registers alone do not fully eliminate spatial outliers; pairing them with observation flow-matching (FM) loss supervision suppresses spatial outliers while concentrating high-norm responses in registers. Beyond this reorganization, observation FM loss supervision also improves action learning without registers, motivating an additional noised-observation FM objective that learns current-scene recovery from corrupted inputs. The resulting model achieves 81.64\% success on LIBERO-Plus without additional embodied pretraining data or foundation models, reaches 93.0\% average success on RoboTwin 2.0, and improves real-world manipulation performance.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-13 · Indexed by Pub-AI: 2026-10-14 UTC

preprint

Acting Through Registers

Yiwei Zhang, Yantai Yang, Jin Gao, Jin Lou, Andong Chen, Yufei Liu, Boyang Xing, Yuchen Zhu, Weiming Hu, Zhipeng Zhang

Preprint posted: 2026-10-13 · Original record ↗

arxiv
2610.17638
Explicit source versions (1)
  • v1 · 2026-10-13
  1. Preprint available

    Source date: 2026-10-13 · First detected: 2026-10-25 UTC · Historical baseline

    Source evidence (1)
Abstract · arxiv

World Action Models (WAMs) use generative visual representations to condition robot actions, yet how these representations organize action-relevant information remains poorly understood. We identify persistent high-norm spatial outliers in WAM representations and find that their action relevance persists even after normalization and K/V projection attenuate their extreme magnitude. Motivated by this observation, we introduce RegisterWAM, which adds learnable non-spatial registers to the world model and uses them directly for action conditioning. Registers alone do not fully eliminate spatial outliers; pairing them with observation flow-matching (FM) loss supervision suppresses spatial outliers while concentrating high-norm responses in registers. Beyond this reorganization, observation FM loss supervision also improves action learning without registers, motivating an additional noised-observation FM objective that learns current-scene recovery from corrupted inputs. The resulting model achieves 81.64\% success on LIBERO-Plus without additional embodied pretraining data or foundation models, reaches 93.0\% average success on RoboTwin 2.0, and improves real-world manipulation performance.