We investigate a simulation to real-world transfer of data-driven models for state estimation using measurements received over a wireless network. Real-world networks, such as vehicular communication systems, face issues like channel impairments and network congestion, and hence estimates are computed using a history of measurements that are received intermittently and are aged on reception. Training an estimator model directly with real-world data is expensive and impractical due to diverse network conditions. To address this, we present an approach that doesnâ\euro™t rely on real-world network data. We train a model using only data generated by a low-cost low-fidelity simulation of networks. To bridge the gap between simulation and reality, we employ domain randomization, appropriately randomizing simulation parameters during training. An in depth investigation of network simulation is provided that has a FCFS queue with two parameters, the rate of arrival of packets into the queue and packet service rates. The model is validated by two real network deployments. In one, we process vehicular trajectory from up to 70 nodes over WiFi to estimate vehicle positions and speeds. In the other, GPS coordinates are communicated by public transit buses over city-wide cellular networks to a transport authority database. This information is obtained by a server running in the AWS cloud, by polling at a certain fixed rate. The model uses the polled data to generate estimates of bus positions. Our proposed estimator model is a novel deep neural network (DNN) architecture, which outperforms existing approaches to data-driven estimation.