OSWAM: An Effective World Action Model with One-Step Action Generation

OSWAM. An Effective World Action Model with One-Step Action Generation. By Guowen Zhang and six others, from the Hong Kong Polytechnic University. Published on the twenty-sixth of October.

Direct-policy World Action Models, or WAMs, separate action generation from explicit prediction of future visuals, and that decoupling has made inference efficient in autonomous driving and embodied control. They still need tens of denoising steps to reach competitive action quality, though. Tens of steps means tens of passes before a vehicle or a robot can act, and the latency this adds stands in the way of real-time deployment.

Cutting the steps naively does not work. The denoising trajectory is nonlinear and spread over many steps, and squeezing it into a few collapses it into coarse updates that substantially degrade the actions. Speed bought that way is paid for in quality.

OSWAM, a world action model with one-step action generation for driving and embodied control, starts from a diagnosis. The authors argue that what forces direct-policy WAMs into iterative denoising is the absence of structured action priors when the process is initialized from random noise. Begin from pure noise, and coherent action sequences have to be recovered gradually, step after step. Seen this way, the many steps are not demanded by the task itself. They are the cost of starting with no structure.

OSWAM therefore changes where denoising begins. It constructs imagination tokens to initialize the diffusion process, and these tokens provide physical priors and action-conditioned representations. The starting point is no longer empty: it already carries the kind of structure the diagnosis said was missing. A second component, privileged information distillation with future conditioning, has the stated purpose of making action generation future-aware. With both in place, OSWAM generates high quality actions in just one diffusion step.

In autonomous driving and in embodied control, OSWAM achieves comparable or even better results than direct-policy WAM and Joint-WAM, which both require tens of steps. The comparison holds against both kinds of model, the one that decouples action from visual prediction and the joint one. It runs approximately three times faster than direct-policy WAM baselines. Compiled for real-time use, it reaches latency as low as forty milliseconds on an NVIDIA RTX 5090 GPU.

One step instead of tens is the headline, but the argument underneath it is about initialization. If iterative denoising exists to build structure that random noise lacks, then supplying that structure at the start, as priors carried by the imagination tokens, removes most of the reason to iterate at all. The results in driving and in embodied control are the test of that argument, in two settings where acting late is acting badly.