In urban logistics, various sources of uncertainty can invalidate pre-planned routes. In this context, a routing strategy that uses available information from the environment could help improve the overall performance of the routing process by dynamically choosing the next client at the online execution time. While static and deterministic testbeds for vehicle routing exist, their stochastic and dynamic counterparts are still missing. This paper proposes an interface to the microtraffic simulation package SUMO that implement a generative model of stochastic and dynamic vehicle routing problems. We formalize the latter using a reinforcement learning framework for semi-Markov decision processes. The resulting testbeds make it possible to compare single- and multi-agent reinforcement learning algorithms in customizable routing environments. We report our preliminary tests to evaluate a hand-crafted policy on some basic scenarios.