StreamingWAM

Streaming World Action Modeling via
Chunk-Wise Staircase Flow Matching

Shaojie Ren1,2,3,*, Jingwen Sun4,*, Zezhi Liu5, Hongjin Chen6, Xin Jin7, Wei Zou3,†, Zhibo Chen4,†
1 University of Chinese Academy of Sciences 2 Beijing Zhongguancun Academy 3 Institute of Automation, Chinese Academy of Sciences 4 University of Science and Technology of China 5 Nankai University 6 Harbin Institute of Technology 7 Eastern Institute of Technology, Ningbo
* Equal contribution. † Corresponding authors.
Code Paper Checkpoint Dataset
StreamingWAM executes continuous real-robot control on a dynamic conveyor task.

Abstract

World Action Models (WAMs), often built upon pretrained video-generation world models, have emerged as a promising framework for general-purpose and open-ended robot control. However, their real-world deployment often suffers from discontinuous and jerky execution. Existing methods typically address this issue by either improving inference efficiency or enhancing action continuity across prediction chunks, without jointly addressing both challenges. To bridge this gap, we propose StreamingWAM, a streaming-diffusion framework for WAM-based robot control. StreamingWAM combines a noise ladder for chunk-wise action prediction, a prefix rollout module for cross-chunk continuity, and dense video-history memory for temporally consistent conditioning. Experiments on LIBERO, LIBERO-Plus, RoboTwin 2.0, and real-world robotic tasks demonstrate that StreamingWAM improves action continuity across static and dynamic settings, achieving competitive performance with the lowest steady-state replanning latency among the evaluated WAMs using the same backbone. The method also produces smoother action transitions and faster steady-state replanning, while the streaming staircase reduces sensitivity to the execution horizon K.

Standard action-chunk redraw compared with StreamingWAM carry-and-refine inference.
Standard policies redraw independent action chunks; StreamingWAM carries and refines a persistent action buffer.

Method

StreamingWAM combines a staircase noise schedule, a clean action prefix, and dense visual history to turn chunk-wise diffusion into a rolling control process.

StreamingWAM architecture and comparison between standard and streaming inference.
StreamingWAM architecture and inference. The persistent action buffer is shifted and refined once whenever a new observation arrives.

Results

Simulation and real-robot results are presented with matching latency and action-continuity analyses.

Simulation results and analysis

StreamingWAM improves benchmark success while reducing steady-state latency, remaining stable across execution horizons, and smoothing action-chunk transitions.

Four-panel simulation analysis showing benchmark success, latency versus success, robustness across execution horizons, and a two-part action-continuity trace.
Simulation benchmarks and analysis. The latency and continuity panels use LIBERO-Plus; markers in (d) identify the K=4 action-chunk boundaries.

Real-robot results and analysis

Per-task success, response latency, and action continuity are summarized here; closed-loop task comparisons and qualitative evidence follow below.

Per-task real-robot comparison across six policy and execution settings.
Successful trials for five manipulation tasks and valid lit-button activations across 50 whack-a-mole sessions.

Why does streaming help on the real robot?

The first view connects response time with aggregate success, while the second separates overall movement from abrupt changes. Task-level closed-loop comparisons and motion-prediction evidence follow below.

Scatter plot of steady-state latency and overall real-robot success for four asynchronous methods.
StreamingWAM combines the lowest measured steady-state latency with the highest overall success across the five manipulation tasks. All four points use asynchronous execution.
Measured joint displacement and discrete acceleration proxy for FastWAM-Joint Async, FastWAM-Joint RTC, and StreamingWAM.
StreamingWAM stays below both FastWAM-Joint execution modes in the discrete acceleration proxy for most aligned samples. The proxy is indexed by sample order rather than physical time.

Real-robot rollouts

A continuous four-trial sequence is featured first, followed by three static and three dynamic real-robot tasks.

Original 1× playback. Every video on this page is shown at its recorded speed, without speed-up or slow-motion processing.

Featured rollout

Four consecutive successes with human interaction

The presentation view is shown at full size; the three synchronized policy views are concatenated alongside it in the same order used by the model.

Presentation viewCamera 4 · 1416×1000
Exterior view 1
Exterior view 2
Wrist view

More rollouts

Six tasks grouped by whether the target remains fixed or changes during execution.

Static tasks

Fixed targets

Dynamic tasks

Moving targets and time-critical interaction

Policy comparisons

The two tasks with the clearest gains are shown first as closed-loop policy comparisons. Select a trial and camera view to compare all three policies; brief task-level interpretations and supporting motion-prediction evidence follow the rollouts.

Trial
Camera
00:00.0 / 00:00.0
Supporting evidence for the dynamic task

Qualitative motion anticipation before contact

The closed-loop conveyor videos above show the behavioral difference. To examine one plausible source, we compare future-frame predictions at a matched task stage: StreamingWAM predicts motion along the conveyor, whereas FastWAM-Joint predicts little displacement. This supports the interpretation that dense temporal history helps the WAM infer target motion, while remaining qualitative evidence rather than a controlled history-only ablation.

Prediction view
MethodPredicted displacementObserved displacementDirection cosine
FastWAM-Joint1.38 px6.93 px0.497
StreamingWAM9.66 px5.93 px0.995

Resources

The current code release contains LIBERO-Plus training and rolling inference. Real-world demonstrations and supplementary videos are hosted separately.