Event
OSWAM: A World Action Model for One-Step Diffusion Action Generation
1 reports · 1 sources
Guowen Zhang and other authors released a paper proposing OSWAM, a world action model for autonomous driving and embodied control, which compresses action generation into a single diffusion step. The authors state that they initialize the diffusion process with imagination tokens to provide physical priors and action-conditioned representations, and introduce privileged information distillation with future conditioning, making action quality comparable to or better than the direct-policy WAM and Joint-WAM, which require dozens of steps. The efficiency figures given by the authors are: roughly 3x speedup compared with the direct-policy WAM baseline, and compiled inference latency as low as 40ms on an NVIDIA RTX 5090 GPU. These quality and speed conclusions are all as stated by the paper's authors.
Report timeline
- OSWAM: A World Action Model with One-Step Action Generation
OSWAM proposes a world action model that generates actions with just one diffusion step, for autonomous driving and embodied control. The authors initialize the diffusion process with imagination tokens to provide physical priors and action-conditioned representations, and introduce privileged information distillation with future conditioning, making action quality comparable to or better than direct-policy WAM and Joint-WAM, which require tens of steps. It is about 3 times faster than the direct-policy WAM baseline, with compiled inference latency as low as 40ms on an NVIDIA RTX 5090 GPU.
arXiv cs.RO ·