High-resolution or super-resolution parameter estimation, i.e. going beyond the classical diffraction limit for target separation in range, speed and angle, is of recent interest for automotive radar systems. Especially for parameter estimation in angular domain (azimuth, elevation), where the number of sampling points is physically limited by the number of antenna elements and the observation space is limited by the aperture of the antenna arrays, super resolution techniques are necessary to achieve state-of-the art performance requirements in terms of angular resolution. Challenges for automotive radar in particular are that typically only a single observation snapshot is available and the targets to be separated are coherent. This has a negative effect on the properties of the estimate of the spatial covariance matrix, the main bearer of information for angular processing. In consequence, only a certain subset of super resolution signal processing methods remains practically relevant for automotive radar high-resolution processing. In this presentation, we will introduce a novel method for high-resolution parameter estimation in automotive radar: the application of neural networks. In contrast to other practical approaches such as maximum likelihood estimation or compressive sensing, the machine-learning based approaches do not require the number of targets to be separated to be known in advance. In addition, model errors such as element coupling and distortions due to radome and vehicle bumpers are inherently trained into the networks. Last but not least, once trained offline, inference of the parameters using a neural network approach corresponds to a data-streaming processing, which could deliver a much higher data throughput in contrast to iterative methods such as compressive sensing.