ForeWAM Enhances Robot Action with Latent Future Foresight

World Action Models (WAMs) are crucial for robots to understand how the physical world evolves during interaction and to generate appropriate actions. Existing WAMs either explicitly predict future visual states, incurring high inference costs, or directly predict actions from current observations, lacking an explicit way to expose predictive dynamics to the action pathway. This research introduces ForeWAM, a novel WAM that bridges this gap. ForeWAM is a dynamics-conditioned direct-policy WAM designed to provide predictive context for action generation without the need to decode future videos. Its core innovation, Future-KV, performs a single Video DiT prefill over the current visual latent and stochastic future slots, reusing the resulting key-value states throughout action denoising. Additionally, dynamics registers, supervised by a frozen latent action teacher, encourage the implicit future states to capture interaction-induced transitions like object motion and task progress. ForeWAM achieves high success rates on complex manipulation benchmarks like LIBERO and LIBERO-Plus without embodied robot data pretraining and without generating future videos during deployment, demonstrating efficient action prediction with integrated foresight.