Acting Through Registers
Yiwei Zhang, Yantai Yang, Jin Gao, Jin Lou, Andong Chen, Yufei Liu, Boyang Xing, Yuchen Zhu and 2 others
Shanghai Jiao Tong University · Institute of Automation, Chinese Academy of Sciences · Humanoid Robot (Shanghai) Co., Ltd.
Published October 13, 2026 · on the feed October 15, 2026 · ▲ 9 · arXiv 2610.17638
The abstract
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.
The Brief
2:00A short podcast summary of the paper. Written and voiced by AI.
Transcript
Acting Through Registers. By Yiwei Zhang and nine others, from Shanghai Jiao Tong University, the Institute of Automation, Chinese Academy of Sciences, and one other institution. Published on the thirteenth of October. Its code repository is worked out: about seventy source files, with a readme and a licence.
World Action Models, or WAMs, condition robot actions on generative visual representations. How those representations organize the information an action needs is poorly understood. The authors look inside and find persistent high-norm spatial outliers. More than that, those outliers stay relevant to action even after normalization and the K and V projections have attenuated their extreme magnitude.
That observation motivates RegisterWAM. It adds learnable registers to the world model, tokens that are not tied to any spatial position, and uses those registers directly for action conditioning.
Registers on their own do not fully eliminate the spatial outliers. What does is pairing them with supervision from an observation flow-matching loss. With both in place, the spatial outliers are suppressed and the high-norm responses concentrate in the registers instead. The flow-matching supervision turned out to help on a second front: it improves action learning even without registers. That motivated one more objective, a noised-observation flow-matching loss in which the model learns to recover the current scene from corrupted inputs.
The resulting model achieves eighty-one point six four percent success on LIBERO-Plus, without additional embodied pretraining data and without foundation models. On RoboTwin two point oh it reaches ninety-three percent average success, and it improves manipulation performance in the real world as well.
Read as a sequence, the paper moves from a diagnosis to a remedy. Something action-relevant was hiding in a few loud spatial tokens. Giving that information a dedicated place to live, and supervising the observations so it moves there, is the change behind the results.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free