1Step-WAM: Your World Action Model but Faster using 1-Step Action Generation

Key points

• We demonstrate that the action-denoising trajectories in WAMs are approximately straight, rendering the first-step velocity a strong estimate of the final action direction.1
• Building on this observation, we introduce 1Step-WAM, a framework that transforms multi-step action denoising into one single forward.1
• Compressing action generation into a single step enables it to execute in parallel with visual prefilling, completely breaking the sequential bottleneck of the traditional imagination-then-action paradigm.1
• Evaluations on RoboTwin and LIBERO show that 1Step-WAM largely preserves multi-step task success with only about 13.4M trainable parameters, while achieving inference speedups of $7.44\times$, $5.98\times$, and $24.84\times$ for FastWAM, ImageWAM and Kairos, respectively.1

Sources (1)

• [1]1Step-WAM: Your World Action Model but Faster using 1-Step Action Generation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 01:50 PM

  World Action Models (WAMs) unify visual world modeling and action generation, yet their real-world deployment is heavily bottlenecked by iterative action denoising and the sequential execution of visual prefilling and action prediction.

  We demonstrate that the action-denoising trajectories in WAMs are approximately straight, rendering the first-step velocity a strong estimate of the final action direction.

Extractive summary: sentences quoted from the sources.