Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

Elastic-WAM: Adaptive Spatiotemporal Imagination for Efficient Robot Control

Haitian Li, Tianfu Li, Xuxi Zhao, Zizheng Liu, Yangcheng Yu, Yiwen Hou, Zeyu Xiao, Mutian Xu, Ziqian Yang, Haonan Chen, Lin Shao

November 2026

Two xArm 7 arms carry out the tea-preparation task, seen from an external camera.

TL;DR Elastic-WAM learns, on top of a frozen world-action model, when to stop denoising the predicted future video, how far to step and which video tokens to refresh. It cuts inference latency by 32.2% against Joint-WAM on RoboCasa-GR1 and raises full-task success on a real bimanual tea-preparation task from 30% to 50%.

What they do

World-action models (WAMs) predict future video alongside actions, and denoising that video dominates inference cost. Existing shortcuts spend a fixed refinement budget in every state. On 10 LIBERO-Plus tasks the authors find that the best number of denoising steps varies from one to four across tasks and states, and that actions attend unevenly to future video regions. Elastic-WAM therefore decides per state how much of the future to refine, and where.

How it works

  • Refinement demand field. The 196 future video tokens are split into 64 groups. Each group gets an amplitude, and all groups share one Beta-shaped profile over denoising progress. The field’s total mass decides when to stop refining the video (“World Commit”), its mode sets the next step size, and the amplitudes rank which groups to refresh (96 tokens per sparse update by default).
  • Three-stage training, backbone frozen. The temporal profile, then the spatial priorities, then the overall demand are each supervised by the error of the resulting action sequence, on top of a pretrained FastWAM-Joint model.
  • Inference. One dense joint denoising step, then sparse steps that recompute only the selected video tokens and reuse cached states for the rest. After World Commit, action denoising continues on the cached video representation.

Results

  • RoboCasa-GR1: highest success rate among the compared methods, with inference latency 32.2% lower than Joint-WAM (measured on an NVIDIA H800).
  • The paper reports the highest average success on LIBERO, LIBERO-Pro and LIBERO-Plus, the last two without adapting the LIBERO-trained policy.
  • Real robot (two xArm 7 arms, four-stage tea preparation, 200 demonstrations, 20 trials per condition): full-task success 50% nominal and 50% under neon lighting, against 30% for Joint-WAM and 0% for Fast-WAM. With a replaced cup it ties Joint-WAM at 30% but has a higher progress score (47.5% against 35.0%).

Limitations

The paper has no limitations section. The real-world evaluation covers a single task, where half of the nominal trials still fail. The token budget per sparse update is set by hand, and a larger budget lowered success on LIBERO-Plus. The explanation offered for the long-horizon gains is stated as a hypothesis.

Videos

Wrist-camera view of the tea-preparation task.

Figures

From Figure 2: Pipeline of Elastic-WAM. Top: inference starts with dense joint refinement, followed by adaptive spars...
From Figure 2: Pipeline of Elastic-WAM. Top: inference starts with dense joint refinement, followed by adaptive sparse world refinement with state-dependent denoising progression and spatial token selection. Once the world is committed, action refinement continues using the cached world representations. Bottom: the refinement field consists of (a) a temporal branch that predicts a shared Beta distribution over denoising progression, (b) a spatial branch that estimates group-wise refinement amplitudes, and (c) a unified field that determines when to commit the world, how far to advance denoising, and where to refresh the future.
From Figure 4: Real-world evaluation settings: Nominal, Lighting OOD, and Object OOD. Lighting OOD introduces strong ...
From Figure 4: Real-world evaluation settings: Nominal, Lighting OOD, and Object OOD. Lighting OOD introduces strong illumination changes, while Object OOD replaces the white cup with a silver-colored cup of a slightly different shape.
From Figure 6: Keyframes of a successful four-stage tea-preparation rollout using Elastic-WAM, from the head camera a...
From Figure 6: Keyframes of a successful four-stage tea-preparation rollout using Elastic-WAM, from the head camera and the wrist camera. Stage 1: placing the white cup on the uncovered teapot. Stage 2: pouring tea from the covered teapot into the white cup. Stage 3: returning the white cup to its original position. Stage 4: pouring tea into the black cup and returning the uncovered teapot.
Abstract

World-Action Models improve robot control by predicting future visual dynamics, but iterative video denoising incurs substantial inference overhead. Recent approaches reduce this cost through partial future denoising, yet typically rely on fixed refinement budgets. Our empirical observations reveal that preferred refinement depths vary across tasks and states, while different future regions exhibit varying relevance to action generation. Motivated by these findings, we propose Elastic-WAM, which adaptively allocates world refinement computation across temporal denoising time and spatial future space. A unified spatiotemporal refinement demand field governs when to stop world refinement, how far to advance each denoising step, and which regions to refresh. We parameterize this field with a shared temporal basis and spatial amplitudes, and learn it through a three-stage action-grounded supervised procedure. At inference, the learned field guides recurrent sparse world refinement to improve task success while reducing inference overhead. Evaluations across diverse simulation benchmarks and real-world long-horizon manipulation demonstrate state-of-the-art control performance with reduced inference latency.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.