DECOWAM: Decoupled Whole-Body World-Action Model for Legged Mobile Manipulation
A new model called DECOWAM was introduced for legged robots that perform mobile manipulation. It separates camera motion caused by the robot’s base movement from the independent actions of the base and the arm. Earlier world-action models, built mainly for fixed-base arms, did not make this separation explicit when predicting how actions change future camera views and control signals.
The separation matters for systems that must coordinate locomotion with manipulation. By using dedicated conditional interfaces for each component, the model can represent the combined effects of walking, turning, and arm motion on what the robot sees and how it should act next.
The approach applies to any platform where ego-motion from the base must be distinguished from end-effector motion during planning or control.