This article studies the problem of RFID-based gesture recognition, which is practically important in various human-computer interaction scenarios, for example, smart homes, intelligent logistics, and smart cities. However, the existing solutions normally suffer from two major limitations: the model-driven methods are sensitive to specific environmental factors, and usually do not adapt well to a complex scenario that is full of multipath; the data-driven methods normally need the collection of massive RFID training data, and deploying the model in the remote cloud leads to long response delay. To overcome the above limitations, this article proposes a cross-domain augmentation-based AI learning (CAL) framework in the context of cloud-edge computing. In the CAL framework, we can simulate massive RFID phase profiles by converting the computer vision data that contains the gesture movement information, instead of costing lots of manpower to actually collect RFID training data. The simulated RFID phase profiles are used to train an AI model in the high-performance cloud. Note that since many sources of this kind of computer vision data are available online, we actually do not even need any manpower to collect training data. To achieve time-efficient recognition, knowledge distillation is applied to get a light and accurate model, which is deployed at the edge side. Thus, recognition response delay can be significantly reduced because the edge server where the AI model is actually deployed is much closer to users than the cloud server. We use commercial off-the-shelf RFID, Kinect, a high-performance server, and a laptop to implement the CAL framework. Extensive experiments are conducted to evaluate the performance of CAL. The results reveal that gesture recognition accuracy of CAL can reach nearly 90 percent without collection of any RFID training data. [ABSTRACT FROM AUTHOR]