Cascade world-action models (WAMs) generate predicted visual futures and then use those futures to condition robot actions. The approach works, but it inherits slow iterative inference from diffusion-pretrained experts. A new arXiv preprint by Shaoyi Zheng and colleagues asks what a single-step predicted future must actually preserve to support accurate single-step action prediction.
The authors find that in the evaluated WAM, resampling video noise has little effect on downstream actions, yet naively truncating both the video and action streams to one step substantially reduces task success. Video-space supervision only partially closes that gap, and a local sensitivity analysis shows that vision prediction errors have highly unequal consequences for action accuracy. This motivates DirectWAM, which converts pretrained video and action experts into single-step predictors without stochastic sampling.
DirectWAM first learns a one-step action model, then trains the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error through the expert. An optional future-latent reconstruction loss is also available, but the core objective evaluates the student-generated future through its downstream control consequences rather than representation matching alone. The method achieves 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM at the same one-evaluation-per-stream budget, and it extends to three real-world tasks while reducing inference latency by 28.0x compared with the full iterative teacher.