This paper presents a self-stabilizing distributed algorithm for the recovery of large population of robots in complex environments, That is, to gather them all to a single location. We assume the robots do not have a map of the environment, but rather use network communications to physically route themselves towards the goal location. Since robot motion can disconnect the network in complex environments, performing the recovery task using in vision based navigation and limited optical communication networks can be challenging. The proposed DMLST algorithm uses an approximation of maximum-leaf spanning trees to provide a communications network that ensures connectivity during the recovery process. After the tree is built from robot networks, corresponding leaf robots move toward their stationary parents in the broadcast tree by using a mid-angle navigation algorithm. This strategy guarantees network connectivity during the recovery until all the robots get to a base location. The DMLST recovery has been both tested in simulation, and implemented on thirty robots in different complex environments. The result shows that DMLST recovery is reliable, efficient and quite successful in 90% of trials. However, a simple clustering algorithm fails in all experiments.