DECOWAM Enhances Legged Robot Mobile Manipulation with Decoupled Model.

Mobile manipulation for legged robots presents a significant challenge: predicting how the combined movements of locomotion and arm actions will affect future observations and control. Existing world-action models, largely developed for stationary platforms, often fail to explicitly differentiate between the robot's camera ego-motion and the independent actions of its base and arm. To address this, a new model called DECOWAM (Decoupled Whole-Body World-Action Model) has been developed. DECOWAM separates these factors through dedicated conditional interfaces, freezing a FastWAM backbone and training residual adapters for efficiency. The model incorporates an action-equivalent future bottleneck, adversarially separated base and arm latents, and base-velocity conditioning for video prediction. To facilitate this research, a new real-robot dataset, ARMDOG, was created, synchronizing video, whole-body state, action, and language. In experiments, DECOWAM improved both future-video and action prediction over FastWAM, reducing action Mean Squared Error by 21.7% with a relatively small number of trainable parameters. Crucially, it demonstrated superior whole-body coordination and base-displacement robustness in closed-loop real-robot trials, indicating a significant step forward in embodiment-aware factorization for joint visual prediction and whole-body control under moving viewpoints.