Learning 4D Geometric Priors for Inference-Efficient World Action Models

Jul 16, 2026

Upload by Noah

World Action Models (WAMs) enable robotic manipulation by modeling visual future dynamics together with executable action sequences, yet common co-training largely targets appearance-focused video latents and cannot adequately capture temporally evolving geometry needed for accurate grasping and contact. MECo-WAM introduces action-relevant 4D geometric priors via multi-expert co-training, using a lightweight 4D expert supervised by relational targets from a frozen VGGT encoder and decayed guidance to avoid non-causal shortcuts. Action-aware temporal geometric distillation aligns in-frame relations and their temporal evolution. Deployment removes all auxiliary 4D modules, improving LIBERO and RoboTwin performance without increasing inference cost.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Abstract

Introduction

Problem: appearance-dominated video prediction

Approach: MECo-WAM multi-expert co-training

Decayed 4D read-mask attention and deployment strategy

Action-aware temporal geometric distillation

Contributions and experimental evaluation

Learning 4D Geometric Priors for Inference-Efficient World Action Models Jianjun Zhang1,2,* Jian Zhu2,‡ Taiyi Su2 Chong Ma1,2 Zitai Huang1,2 Yi Xu2,† Hanli Wang1,† 1Tongji University 2AIRC, Midea Group †Corresponding Author ‡Project Leader [Project Page: meco-wam.github.io](Project Page: meco-wam.github.io) arXiv :2607 .05468v 1 [ cs .RO] 6 Jul 2026 Abstract World Action Models (WAMs) have shown strong potential for robotic manipulation by jointly modeling visual future dynamics and executable action sequences. However, existing video-action co-training methods primarily optimize appearance-oriented video latents, which may insufficiently capture the temporally evolving geometry required for precise manipulation. We propose MECo-WAM, a Multi-Expert CoTraining World Action Model that injects action-relevant 4D geometric priors into video-action representations while preserving the original lightweight inference graph. During training, MECo-WAM combines video and action experts with a lightweight 4D expert supervised by relational targets from a frozen VGGT encoder. Asymmetric expert visibility prevents non-causal shortcuts from auxiliary geometry to action generation. To transfer geometric knowledge into the deployed video-action pathway, we introduce decayed 4D read-mask attention, which provides restricted current-frame geometric guidance early in training and progressively removes this dependency. We further propose action-aware temporal geometric distillation, which aligns within-frame geometric relations and their temporal evolution while emphasizing visual regions most relevant to robot actions. At deployment, all auxiliary 4D components are removed. Experiments on LIBERO (98.2%), RoboTwin 2.0 (92.6%), and challenging real-world manipulation tasks show that MECo-WAM improves manipulation performance without increasing inference cost. Introduction Robotic manipulation requires a policy to map visual observations and language instructions to precise action trajectories (Hu et al. 2025; Ma et al. 2026; Su et al. 2026; Wang et al. 2026; Su et al. 2025) . World action models offer a promising formulation by jointly learning how visual states evolve under interaction and how robot actions should be generated (Ye et al. 2026c; Kim et al. 2026; Bi et al. 2026; Li et al. 2026b) . Compared with direct action policies, video-action co-training can provide richer motion and interaction priors, enabling the policy to reason over changes that unfold beyond a single observation (Ye et al. 2026a; Yuan et al. 2026) . Despite this advantage, the visual representation learned by many WAMs remains dominated by appearance-oriented * This work was completed during an internship at Midea AIResearch Centers. 95 90 80 65 60 150 200 250 1K 2K Success Rate (%) \\\\ \\ Inference Latency (ms) Figure 1: Comparison of MECo-WAM with Baselines inaction-chunk inference latency and task success rate on RoboTwin. video prediction. A video latent can support plausible future synthesis without explicitly preserving the spatial relations that determine whether a grasp is reachable, whether an object is aligned with a target, or whether contact will cause astable transition. These relations are not static: they evolve asthe robot approaches, contacts, moves, and releases objects. Recent geometry-aware VLA andWAM studies therefore introduce 3D or 4D structure to strengthen spatial grounding for manipulation (Qu et al. 2025; Li et al. 2025a,b; Guo et al. 2026; Li et al. 2026c) . Thus, manipulation-oriented WAMs require geometry-aware temporal representations beyond visual plausibility. A direct approach is to introduce explicit 4D reconstruction or dense geometric prediction into the world action modeling pipeline (Guo et al. 2026; Li et al. 2026c) . However, making geometry an explicit deployment-time output increases inference cost and may shift optimization toward geometric reconstruction that is only weakly co

FAQs

What limitation in existing WAM co-training motivates MECo-WAM?

Existing methods mainly optimize appearance-oriented video latents, which do not sufficiently represent temporally evolving geometry required for precise robotic manipulation.

How does MECo-WAM inject 4D geometric priors during training while keeping inference efficient?

MECo-WAM uses a lightweight 4D expert only during training, supervised by relational targets from a frozen VGGT encoder, and restricts early cross-expert access so only limited current-frame geometry guidance is transferred. At deployment, all auxiliary 4D components are removed.

What is decayed 4D read-mask attention used for?

It provides restricted current-frame geometric guidance early in training and progressively removes that dependency before deployment, preventing non-causal shortcuts.

How does action-aware temporal geometric distillation improve the learned representations?

It aligns predicted 4D keyframe relations and their temporal evolution while emphasizing visual regions most relevant to robot actions.

Jul 16, 2026

Upload by Noah