Event
DirectWAM: A Single-Step De-Diffusion World-Action Model
1 reports · 1 sources
Shaoyi Zheng et al. propose the DirectWAM method, which, through action sufficiency distillation, converts pretrained video and action experts into a single-step predictor that requires no stochastic sampling while retaining an explicit future-to-action cascade structure. During training, it first learns a single-step action model, then decodes actions from its own predicted future and back-propagates the demonstrated action error to train the future predictor. The paper claims that DirectWAM achieves an average success rate of 97.1% on the four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM under a single evaluation budget per stream, and demonstrates this advantage on three real-world tasks designed by the authors; compared with the full iterative teacher, inference latency is reduced by a factor of 28.0.
Report timeline
- DirectWAM: Distilling Cascaded World-Action Models into Single-Step Action Prediction
The paper proposes DirectWAM, which converts pretrained video and action experts into single-step predictors that require no stochastic sampling while retaining the explicit future-to-action cascade structure. After learning a single-step action model, the method trains the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error.
arXiv cs.RO ·