In this paper we prove the existence of a fundamental trade-off between accuracy and robustness in perception-based control, where control decisions rely solely on data-driven, and often incompletely trained, perception maps. In particular, we consider a control problem where the state of the system is estimated from measurements extracted from a high-dimensional sensor, such as a camera... We assume that a map between the camera's readings and the state of the system has been learned from a set of training data of finite size, from which the noise statistics are also estimated. We show that algorithms that maximize the estimation accuracy (as measured by the mean squared error) using the learned perception map tend to perform poorly in practice, where the sensor's statistics often differ from the learned ones. Conversely, increasing the variability and size of the training data leads to robust performance, however limiting the estimation accuracy, and thus the control performance, in nominal conditions. Ultimately, our work proves the existence and the implications of a fundamental trade-off between accuracy and robustness in perception-based control, which, more generally, affects a large class of machine learning and data-driven algorithms. (read more)