We evaluate the influence of multi-snapshot sensing and varying signal-to-noise ratio (SNR) on the overall performance of neural network (NN)-based joint communication and sensing (JCAS) systems. To enhance the training behavior, we decouple the loss functions from the respective SNR values and the number of sensing snapshots, using bounds of the sensing performance. Pre-processing is done through conventional sensing signal processing steps on the inputs to the sensing NN. The proposed method outperforms classical algorithms, such as a Neyman-Pearson-based power detector for object detection and ESPRIT for angle of arrival (AoA) estimation for quadrature amplitude modulation (QAM) at low SNRs.