OSWAM: One-Step Action Generation for World Action Models in Autonomous Driving and Embodied Control

Forum topic · 小凯 · 2026-10-27

Summary

OSWAM is a World Action Model (WAM) that generates high-quality actions in a single diffusion step, targeting real-time deployment in autonomous driving and embodied control. Direct-policy WAMs decouple action generation from explicit future visual prediction but typically require tens of denoising steps, causing latency; simply reducing steps degrades action quality. The authors attribute this to the lack of structured action priors in random noise initialization. OSWAM addresses it by constructing imagination tokens that initialize the diffusion process with physical priors and action-conditioned representations, plus privileged information distillation with future conditioning for future-aware action generation. Experiments show OSWAM matches or exceeds direct-policy WAM and Joint-WAM baselines that need tens of steps, delivering roughly 3x speedup with compiled inference latency as low as 40ms on an NVIDIA RTX 5090 GPU. The paper (arXiv 2610.35850) includes a project page with additional resources.

Paper Overview

Research area: Computer Vision (CV)
Authors: Guowen Zhang, Liyi Chen, Xu Hu, Pengfei Wang, Jingfan Chen et al.
Published: 2026-10-26
arXiv: 2610.35850
Project page: https://gwenzhang.github.io/OSWAM/

Abstract

Direct-policy World Action Models (WAM) have demonstrated the benefits of decoupling action generation from explicit future visual prediction for efficient inference in autonomous driving and embodied control. However, they still require tens of denoising steps to achieve competitive action quality, incurring substantial latency and hindering real-time deployment. Simply reducing the number of denoising steps will collapse the nonlinear multi-step denoising trajectory into coarse updates, substantially degrading action quality.

To reduce the denoising steps while enabling real-time action generation, the authors present OSWAM, an effective world action model with one-step action generation. They argue that it is the absence of structured action priors in random noise initialization that forces direct-policy WAMs to recover coherent action sequences through iterative denoising. Hence, they construct imagination tokens to initialize the diffusion process, providing physical priors and action-conditioned representations. They further introduce privileged information distillation with future conditioning to enhance future-aware action generation. As a result, OSWAM can efficiently and effectively generate high quality actions in just one diffusion step.

Key Results

OSWAM achieves comparable or better results than direct-policy WAM and Joint-WAM (which require tens of steps) on both autonomous driving and embodied control benchmarks.
Approximately 3x speedup over direct-policy WAM baselines.
Supports real-time compiled inference with latency as low as 40ms on an NVIDIA RTX 5090 GPU.

*Auto-collected on 2026-10-28.*