OSWAM achieves one-step action generation for driving
Original title: OSWAM: An Effective World Action Model with One-Step Action Generation
OSWAM generates high-quality actions for autonomous driving and control in a single diffusion step using imagination tokens.
Direct-policy world action models require many denoising steps for coherent actions. OSWAM introduces imagination tokens to initialize diffusion with physical priors and action-conditioned representations. It adds privileged information distillation with future conditioning. Experiments show comparable or better performance than multi-step baselines with 3x speedup and 40ms latency on RTX 5090.
- One-step diffusion instead of tens
- 3x speedup over baselines
- 40ms latency on RTX 5090
- Structured action priors in initialization
- Supports real-time compiled inference
From the original article
1 min read at huggingface.co
Direct-policy World Action Models (WAM) have demonstrated the benefits of decoupling action generation from explicit future visual prediction for efficient inference in autonomous driving and embodied control. However, they still require tens of denoising steps to achieve competitive action quality, incurring substantial latency and hindering real-time deployment.
By Guowen Zhang, Liyi Chen, Xu Hu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.