The decisional capability of robotic systems has expended significantly in recent years. Their validation typically implies expensive and laborious test campaigns in the field. Part of the validation could use simulation means, but there is currently no systematic method to test robots in virtual worlds. This work aims to provide such a method. We discuss the challenges regarding the definition of virtual worlds and missions, their generation guided by test criteria and the automated analysis of test results. We also present preliminary work for an agricultural robot.