AnyStep-WAM Cuts Denoising Steps by Up to 85% in World Action Models Without Losing Success Rate

Rui Wang · hf · 2026-09-30

Researchers introduce AnyStep World Action Model (AnyStep-WAM), a general framework for tunable-budget prediction and scene-dependent compute allocation in world-action models (WAMs).

Motivation: Manipulation tasks contain action chunks with varying sensitivity to generation errors, yet WAMs typically use a fixed number of denoising steps for all actions.

Method:

Results (on Motus, FastWAM, and LingBotVA using RoboTwin 2.0):

Original post →

More from Embodied

Embodied channel →