A real-time perception–motion codesign method for image-based visual servoing in embodied intelligence systems

The embodied intelligence system achieves adaptive decision-making through the dynamic coupling between intrinsic perception and environmental interaction, wherein the extent of motion coordination is fundamentally constrained by the closed-loop optimization of multimodal perception and actuator control. To solve this problem, this study reconfigures image-based visual servoing (IBVS) into a perception–motion synergy engine within the framework of embodied intelligence. First, a perception–motion mapping model is constructed, in which Mamdani fuzzy inference is employed to achieve dynamic decoupling and adaptive regulation of multi-axis servo gains, which effectively addresses the issues of motion trajectory redundancy and visual feature visibility degradation in embodied intelligent systems operating in unstructured environments. Furthermore, a continuous velocity observer based on a polynomial decay function (PD-CVO) is designed to ensure smooth transitions in system motion velocity, effectively reducing the risk of dynamic imbalance caused by abrupt velocity changes. Experiments show that this method improves the visual-motor response efficiency of the embodied system by 13.25%, reduces redundant motion in image features by 53.63% compared to other state-of-the art approaches, and robustly maintains the initial continuity of the camera’s spatial velocity. This paradigm of embodiment control based on the dynamic adjustment of servo gain provides a new theoretical framework for realizing the real-time coupling of tool manipulation and environment deformation in flexible manufacturing scenarios.