Learned action policies are gaining ever more traction in AI. One natural idea to support human understanding is visualization of policy behavior, enabling users (domain experts) to visually inspect and scrutinize what has been learned. Here we contribute a case study, in an abstract grid-world driving domain that extends the Racetrack planning benchmark with traffic and with policies that generalize across maps. Our visualization takes as input a set of policy traces, and provides a compact overview of critical situations – crashes or almost-crashes – at map positions, as well as means to explore individual policy traces in depth. Our user study shows that users with access to our visualization obtain a better understanding of critical situations.