Reproduce “OSWAM: An Effective World Action Model with One-Step Action Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guowen Zhang, Liyi Chen, Xu Hu, Pengfei Wang, Jingfan Chen, Chenhang He, Lei Zhang
OSWAM uses an observation-conditioned action prior to make world-action models faster at inference. Action Imagination Tokens provide a structured starting point, which the Action DiT refines in one denoising step; ablations show a PDMS gain over one-step Fast-WAM on NAVSIM. OSWAM matches ten-step SimWAM at 90.3 PDMS with 136 ms latency, while RoboTwin success is 90.00%, below Fast-WAM’s 91.83%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.