Gesture recognition has broad applications in the realm of human-computer interaction, and it is currently receiving heightened at­tention from both academic and industrial circles. In recent years, owing to the rapid advancement of wireless communication and IoT technolo­gies, Wi-Fi devices have been ubiquitously deployed, giving rise to a plethora of gesture recognition methodologies founded on channel state information. These approaches offer the foundational and technological underpinnings for cost-effective pervasive sensing. Nevertheless, the prac­tical utility of Wi-Fi-based sensing systems in Wi-Fi-equipped environ­ments can be fraught with challenges in the absence of domain adapta­tion to accommodate new data domains. The attainment of cross-domain sensing hinges significantly on the development of domain-independent features. In this paper, we introduce a Wi-Fi-based cross-domain gesture recognition system. The central challenge of this system centers on the utilization of reinforcement learning to acquire spatiotemporal features from gesture signals, resulting in distinctive, domain-agnostic dynamic features that represent various gestures. Building upon this foundation, we have developed a deep learning model that requires only a single train­ing session and possesses the capability to be effectively applied across diverse domains. A rigorous series of experiments has been conducted, and the outcomes unequivocally demonstrate the commendable perfor­mance of the system, both in intra-domain and cross-domain recognition accuracy.