OSWAM Generates Actions in One Diffusion Step

Researchers from The Hong Kong Polytechnic University published OSWAM, a world action model for autonomous driving and embodied control, on Oct 26. The model generates actions in one diffusion step, and the authors report comparable or better results than direct-policy WAM and Joint-WAM systems that require tens of steps.

OSWAM starts diffusion with “imagination tokens,” which provide structured physical priors and action-conditioned representations. The authors also use privileged information distillation with future conditioning to improve future-aware action generation. They say this approach addresses latency from iterative denoising while maintaining action quality.

The authors report approximately 3× speedup over direct-policy WAM baselines. In the paper abstract, compiled inference reaches as low as 40ms on an NVIDIA RTX 5090 GPU. A project-page post reports 90.3 PDMS on NAVSIM and 90.00% average success on RoboTwin 2.0, with approximately 3× faster inference than corresponding multi-step baselines on NVIDIA H20; compiled RoboTwin inference reaches 40.81 ms on RTX 5090. The post also says the team tested OSWAM on real-world PiPER manipulation tasks.

Researchers from The Hong Kong Polytechnic University published OSWAM, a world action model for autonomous driving and embodied control, on Oct 26. The model generates actions in one diffusion step, and the authors report comparable or better results than direct-policy WAM and Joint-WAM systems that require tens of steps.

OSWAM starts diffusion with “imagination tokens,” which provide structured physical priors and action-conditioned representations. The authors also use privileged information distillation with future conditioning to improve future-aware action generation. They say this approach addresses latency from iterative denoising while maintaining action quality.

The authors report approximately 3× speedup over direct-policy WAM baselines. In the paper abstract, compiled inference reaches as low as 40ms on an NVIDIA RTX 5090 GPU. A project-page post reports 90.3 PDMS on NAVSIM and 90.00% average success on RoboTwin 2.0, with approximately 3× faster inference than corresponding multi-step baselines on NVIDIA H20; compiled RoboTwin inference reaches 40.81 ms on RTX 5090. The post also says the team tested OSWAM on real-world PiPER manipulation tasks.