This work presents the development of a MATLAB/Simulink simulation framework for a swarm of autonomous off-road ground vehicles. The operating environment is a partially known semantic map with progressive discovery. Terrain perception is handled using a Deep Learning based semantic segmentation module by assigning a terrain class to each cell of the input image. A cost map is constructed to combine the base cost per semantic class, the gradient where the class changes and the slope penalty. Furthermore, cells not yet observed receive a uniform penalty override, which is replaced by the real cost as the terrain is discovered. The architecture is a multi-agent, decentralized decision-making framework on a shared global map: the leader creates the path for itself through an enhanced A* terrain-aware algorithm, to which it is subsequently applied a path smoothing function that ensures curvature continuity.