Acting Through Registers

1 Introduction

• The study addresses limited understanding of how WAM video representations behave and how that behavior relates to robot action generation.WAMs adapt generative video models to jointly model visual dynamics and robot actions, using video representations to condition an action model.Source: §1 · p. 1
• High-norm spatial outliers persist across observations; despite normalization and projection attenuating their magnitude in the visual K/V cache, they retain non-trivial action sensitivity.Their action relevance is consistent even when their locations contain little task-relevant information.Source: §1 · pp. 1–2
• RegisterWAM adds dedicated non-spatial registers for action conditioning, accommodating action-useful representations that may not naturally correspond to a single image location.Registers alone do not fully eliminate spatial outliers; pairing them with observation FM supervision suppresses spatial outliers and concentrates high-norm responses in registers.Source: §1 · p. 2
• Observation FM supervision also improves action learning without registers, motivating noised-observation FM as a current-scene recovery task from corrupted visual inputs.The noised-observation objective provides additional supervision rather than another action-conditioning stream.Source: §1 · pp. 2–3
• 81.64% overall success on the full LIBERO-Plus test set, 93.0% average success on RoboTwin 2.0, and mean real-world task score improves from 0.3483 to 0.5238 over FastWAM.These results use no additional embodied pretraining data or additional foundation models.Source: §1 · p. 3

2.1 Spatial Outliers in World Action Models

• FastWAM’s largest-to-median spatial-token norm ratio rises in middle-to-late layers, with recurring high-norm locations; the outlier’s action relevance persists after K/V projection attenuates its magnitude.The outlier has stable, non-trivial sensitivity, with rank concentrated around 15 of 98 tokens, but task-relevant object tokens can cause larger action changes.Source: §2.1 · p. 3, §2.2 · pp. 3–4, Figure 3 · p. 4

2.2 Outliers and Action Generation

• The action-sensitivity probe masks a visual position from every Action DiT query; the outlier is relevant but is neither a unique action-information carrier nor the only substantially contributing spatial token.Its high norm is an observable signature of representational behavior beyond local spatial encoding, not the behavior itself.Source: §2.2 · p. 4

2.3 Registers with Observation FM Loss Supervision

• Registers are learnable non-spatial tokens jointly updated with spatial tokens, but alone they leave output-layer outliers and yield only 54.18% success versus FastWAM’s 53.64%.The 8-register result is for conditioning on both raw tokens and registers without observation FM supervision.Source: §2.3 · pp. 4–5
• Zero-noise observation FM supervision regularizes dense clean spatial-token outputs, reducing final-layer outliers while concentrating high-norm responses in registers; combined with registers, it suppresses intermediate- and final-layer outliers.The authors do not assign fixed semantic roles to individual registers.Source: §2.3 · p. 5, Figure 5 · p. 5
• With observation FM supervision, 4-register Raw + Reg. conditioning reaches 62.45%, while Only Reg. reaches 61.11%; the raw-token reference is 58.07%.The combined interface is not uniformly superior, and increasing register count does not consistently improve performance.Source: §2.3 · p. 5, Figure 6 · p. 5

2.4 Beyond Outlier Reorganization

• Without registers, observation FM supervision reaches 58.07% success versus FastWAM’s 53.64%, showing an action-learning benefit beyond register redistribution.The noised-observation FM objective then trains current-scene recovery from noisy observations as additional world-model supervision.Source: §2.4 · p. 6

3.1 Registers as the World–Action Interface

• Registers provide the Action DiT’s visual conditioning interface while the world model retains dense visual representations.The Video DiT processes learnable registers alongside observation and future tokens; action keys and values are built from register states.Source: §3.1 · p. 7

3.2 Observation-Side FM Loss Supervision

• Observation FM loss supervises dense clean-observation outputs, reducing spatial outliers and concentrating high-norm responses in registers.Registers themselves are excluded from visual FM losses.Source: §3.2 · p. 7
• Noised-observation FM supervision trains the Video DiT to denoise corrupted current-scene inputs, without adding an action-conditioning stream.Its branch is isolated by the attention mask and its gradients update Video DiT parameters.Source: §3.2 · p. 7

3.3 Training and Inference

• At inference, the Video DiT builds the register K/V cache once, which the Action DiT reuses throughout iterative action denoising.The training objective combines action-chunk denoising with three visual-output losses, all excluding registers.Source: §3.3 · p. 7

4.1 Experimental Setup

• Experiments use Wan2.2 with a 10-step Action DiT; LIBERO-Plus training uses LIBERO data, and ablations use the same 20% test subset.Unless otherwise noted, the setup uses 8 registers and σ = 0.5 for noised-observation FM supervision.Source: §4.1 · pp. 7–8

4.2 Main Results

• 81.64% overall success on full LIBERO-Plus exceeds listed FastWAM and Faster-WAM results; RoboTwin 2.0 averages 93.0%, and real-world mean task score rises from 0.3483 to 0.5238.The LIBERO-Plus comparison uses the same Wan2.2 backbone without embodied-policy pretraining or additional geometric or semantic teacher models.Source: §4.2 · pp. 8–9, Table 1 · p. 8, Table 2 · p. 8
• The six real-world settings comprise three in-distribution tasks and three bottle-handover shifts, with 20 trials per setting.Both methods score zero under the blue-tablecloth shift.Source: §4.2 · pp. 8–9, Figure 9 · p. 9

4.3 Ablation Studies

• Adding isolated noised-observation FM supervision raises LIBERO-Plus subset success from 76.29% to 78.59%; 8 registers plus observation FM supervision reaches 82.52%.Table 5 reports the cumulative component ablation.Source: §4.3 · p. 9, Table 5 · p. 10
• Among tested settings, 8 registers perform best, while performance peaks at σ = 0.5 and drops markedly at σ = 0.25.The noise sweep changes both input corruption and effective auxiliary FM loss weight.Source: §4.3 · p. 9
• Applying the design to ImageWAM raises average LIBERO-Plus success from 83.10% to 85.42%, a gain of 2.32 percentage points.ImageWAM uses the FLUX.2 image-editing backbone rather than the video-generation backbone used here.Source: Table 3 · p. 8, §4.3 · p. 9

5 Related Work

• Prior WAMs use generated visual plans, predictive features, or joint video–action learning, while related VLA methods predict visual goals or structured world knowledge before actions.Other approaches reduce inference cost or add geometric, semantic, motion, or multi-stream supervision; this work studies token organization and observation FM effects.Source: §5 · p. 10
• Unlike work on latent-token conditioning, this paper studies how observation FM supervision redistributes high-norm responses between spatial tokens and registers.The cited register literature also includes non-spatial computation in ViTs and pixel-space DiTs; high-norm outliers are distinguished from attention sinks.Source: §5 · p. 10

6 Conclusion

• Observation FM supervision suppresses output-layer spatial outliers while high-norm responses concentrate in registers, whose states condition action generation.Action learning also improves with observation FM supervision without registers, motivating noised-observation FM supervision.Source: §6 · p. 10

A Additional Outlier Analysis

• At σfuture = 0.9–1.0, first-frame outliers peak around layers 21–23, exceeding an order of magnitude over median spatial-token norm.The varied noise is on future-frame inputs, while the measured outliers are in the first-frame representation; the main-paper default σfuture = 1 is in this strong-noise regime.Source: A Additional Outlier Analysis · p. 14, Figure 10 · p. 14
• Across additional observations, each model favors recurring high-norm locations, but active locations and peak counts vary by sample.The maps compare hidden-state norms for three additional observations in FastWAM and two retrained variants.Source: A Additional Outlier Analysis · p. 14, Figure 11 · p. 15
• Registers suppress intermediate-layer peaks, observation FM supervision suppresses final-layer outliers, and together they keep Sout low throughout ten observations.With either component alone, elevated peaks remain in the layers that component primarily does not suppress.Source: Figure 12 · p. 15, A Additional Outlier Analysis · p. 15
• RegisterWAM’s spatial norm maps show broader variation without conspicuous localized outliers, consistent with combined registers and observation FM supervision.The maps show two additional observations; colors are interpreted within each panel.Source: A Additional Outlier Analysis · p. 15, Figure 13 · p. 16

B Benchmark Details

• Real-world evaluation is detailed below.Source: B Benchmark Details · p. 16

B.1 Real-World Evaluation

• RegisterWAM improves all three ID tasks and the height and distractor variants; both methods score zero with the blue tablecloth in the tested trials.Each setting is evaluated over 20 trials.Source: §B.1 · p. 16
• The reported mean task score awards partial credit for completed steps rather than measuring the fraction of fully successful trials.The overall average is the arithmetic mean of six setting-level scores, including zero-score settings.Source: §B.1 · p. 16
• The battery-insertion instructions assign the left arm two left-side batteries first, followed by the right arm inserting two right-side batteries.Source: §B.1 · p. 16
• The bowl-and-cup task instructions describe repeated stacking of bowls followed by stacking cups.Source: §B.1 · p. 16
• The bottle-handover instruction has the left arm pass a white bottle to the right arm for placement in an orange plate.Source: §B.1 · p. 16

C.1 Visualizing Register Attention

• Four selected register–head pairs attend to different regions, including the gripper, multiple objects, and a lower-right area beside the plates.The maps are overlaid on observation 003426, with register and head indices shown below each panel.Source: §C.1 · p. 16, Figure 14 · p. 17
• The selected attention maps do not imply fixed semantic roles for individual registers.Source: §C.1 · p. 16

• The evidence does not fully establish what information outlier or register tokens encode.Source: D Limitations · p. 17
• The experiments cover only a limited family of WAM architectures.Source: D Limitations · p. 17
• Observed behavior may depend on model scale, training objective, or optimization history.Source: D Limitations · p. 17