X-WAM: Unified 4D World Action Modeling with Asynchronous Denoising

Date: 2026-04-30 Source: HuggingFace | Paper Raw: raw/huggingface/2026-04-30-unified-4d-world-action-modeling-video-priors-asynchronous-denoising.md

TL;DR

X-WAM unifies real-time robot action execution with high-fidelity 4D world synthesis (multi-view RGB-D video + 3D reconstruction) in a single framework. Uses pretrained video diffusion priors plus a lightweight depth-prediction branch (replicates final blocks of the diffusion transformer for depth). The key trick is Asynchronous Noise Sampling (ANS) — fewer denoising steps for actions (real-time decoding), full steps for video (high fidelity), with joint timestep distribution training to prevent inference-time distribution shift. 5,800+ hours of robotic pretraining data; 79.2% RoboCasa, 90.7% RoboTwin 2.0.

Why Tier 4 for Amit

Robotics-heavy and 3D-spatial — outside the routing/efficiency/GPU core. Recorded for the asynchronous-denoising mechanism, which is transferable: any diffusion model that produces multiple outputs at different latency targets could use the same pattern. ANS is a sibling of speculative decoding's "different speeds for different parts of the same generation" — the difference is that ANS trains for the schedule rather than computing speedups at inference.

One Sentence for the Digest

Yet another diffusion-policy world model — interesting only for the asynchronous denoising trick, which generalizes beyond robotics.