OSWAM preprint reports one-step action generation with approximately 3× speedup

Researchers report that OSWAM, a model that generates actions for autonomous driving and embodied control, achieves approximately 3× speedup over direct-policy World Action Model baselines. Their preprint reports compiled inference latency as low as 40ms on an NVIDIA RTX 5090 GPU. OSWAM initializes action generation with learned representations that supply physical guidance, then generates actions in one diffusion step. The researchers also transfer information about future conditions during training. They report comparable or better experimental results than models requiring tens of steps; the findings concern experimental performance rather than deployed driving systems.